CA2084396C - Method of decarburizing molten metal in the refining of steel using neural networks - Google Patents
Method of decarburizing molten metal in the refining of steel using neural networksInfo
- Publication number
- CA2084396C CA2084396C CA002084396A CA2084396A CA2084396C CA 2084396 C CA2084396 C CA 2084396C CA 002084396 A CA002084396 A CA 002084396A CA 2084396 A CA2084396 A CA 2084396A CA 2084396 C CA2084396 C CA 2084396C
- Authority
- CA
- Canada
- Prior art keywords
- oxygen
- bath
- neural network
- temperature
- process period
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Classifications
-
- C—CHEMISTRY; METALLURGY
- C21—METALLURGY OF IRON
- C21C—PROCESSING OF PIG-IRON, e.g. REFINING, MANUFACTURE OF WROUGHT-IRON OR STEEL; TREATMENT IN MOLTEN STATE OF FERROUS ALLOYS
- C21C7/00—Treating molten ferrous alloys, e.g. steel, not covered by groups C21C1/00 - C21C5/00
- C21C7/04—Removing impurities by adding a treating agent
- C21C7/068—Decarburising
- C21C7/0685—Decarburising of stainless steel
-
- C—CHEMISTRY; METALLURGY
- C21—METALLURGY OF IRON
- C21C—PROCESSING OF PIG-IRON, e.g. REFINING, MANUFACTURE OF WROUGHT-IRON OR STEEL; TREATMENT IN MOLTEN STATE OF FERROUS ALLOYS
- C21C5/00—Manufacture of carbon-steel, e.g. plain mild steel, medium carbon steel or cast steel or stainless steel
- C21C5/28—Manufacture of steel in the converter
- C21C5/30—Regulating or controlling the blowing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y10—TECHNICAL SUBJECTS COVERED BY FORMER USPC
- Y10S—TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y10S706/00—Data processing: artificial intelligence
- Y10S706/902—Application using ai with detail of the ai system
- Y10S706/903—Control
- Y10S706/904—Manufacturing or machine, e.g. agricultural machinery, machine tool
Abstract
A method of decarburizing molten metal in the refining of steel using neural networks with a first neural network trained to analyze data representative of many process periods of one or more decarburization operations for providing an oxygen count for a preselected gas ratio of oxygen to diluent gas to cause the temperature of the molten metal bath to be decarburized to rise to a specified aim temperature and with a second neural network trained to analyze data representative of many process periods of one or more decarburization operations for providing an output schedule of oxygen counts to be injected into the bath to reduce the carbon level to a predetermined aim level in one or more successive stages corresponding to a preselected schedule of ratios of oxygen to diluent gas.
Description
METHOD OF DECARBURIZING MOLTEN METAL IN
THE ~FINING OF STEEL USING NEURAL NETWORKS
5 ~ield of the Invention This invention relates to an AOD process for decarburizing molten metal in the refining of steel and more particularly to an AOD process for decarburizing molten metal using neural networks to 10 control the decarburization operation.
Backqround of the Invention A process which has received wide acceptance in the steel industry for refining metal is the 15 argon-02ygen decarburization process also referred to as the "AOD" process. It is the purpose of AOD
refining to first remove carbon from a bath of metal, ne~t reduce any metals that may have oxidized during decarburization, and finally adjust the temperature 20 and chemistry of the bath before casting the metal into a product. Decarburization is achieved by injecting mixtures of 02ygen and inert gases in such a way as to favor the o~idation of carbon over the oxidation of other metal components present in the 25 bath. At progressively lower carbon contents during the process of decarburization progressively greater dilution of the 02ygen by inert gases is injected to favor the oxidation or removal of carbon.
Relationships between the bath weight, 30 chemistry, and temperature, the injections of o~ygen and inert gases, and the resultant changes in metal chemistry and temperature have been theorized to achieve both control and understanding of how to optimize the economics of the process. Thermodynamic models have tracked the general relationships between these parameters, but have limited accuracy and have not obviated the need for intermediate sampling of the bath temperature and chemistry in processing any 5 given heat of metal. Some theorists have adopted the approach that the decarburization reaction may be better understood, and hence controlled, by considering the chemical kinetics of the competing o~idations of carbon and the various metal species 10 present. It follows that approaches incorporating both thermodynamic and kinetic considerations have also been constructed. Finally, statistical approaches have been used to empirically model decarburization in an AOD converter.
The traditional modeling of the decarburization cycle of the AOD operation requires not only a comprehensive understanding of how to represent the thermodynamics and/or kinetics for use in a computer program, but also requires the 20 knowledge of many properties of the species involved in the reactions. For instance, normal thermodynamic modeling requires the knowledge of at least 25 pertinent interaction coefficients. The free enthalpies and entropies associated with each 25 potential reaction must also be known as well as a representative pressure exerted on the bubbles passing through and reacting with the bath. Kinetic models that are based on assumptions that diffusion, adsorption and desorption rates significantly affect 30 the relative extents to which the competing o~idation reactions occur are similarly dependent on accurate knowledge of these rates with respect to temperature and base composition. They must also be capable of modeling the surface areas, velocities of the bubbles relative to the surrounding liquid, and the residence times of the bubbles in the metal phase. Thus, the 5 modeling of decarburization based on chemical theories is subject to many items of data being all accurately measured. They also require a correct understanding of the mechanisms of the various reactions. Since models are deficient in at least 10 one of these two requirements, it is normal for known physical "constants" to be altered to make the results of the model fit actual results better. Due to the complexity of these models, great skill is required to adjust the parameters to improve the 15 overall accuracy of an entire population of results.
Often it is found that one particular solution or combination of adjusted constants is optimal for representing the results of only one particular set of working conditions. That is, solutions tend not 20 to be general, but rather geared to specific small sets of data for which they were adjusted.
In spite of the variety of approaches, inaccuracies remain and some form of measuring the carbon content during the decarburization process 25 step is normally required. This usually necessitates halting the process, withdrawing a metal sample, analyzing the carbon content and measuring the bath temperature before resuming. Lack of process control during decarburization not only necessitates extra 30 sampling, but precludes operation at the optimal conditions for cost reduction and production maximization.
A computerized system using "neural networks" benefits from the fact that a theoretical understanding of decarburization is not required.
Xnowledge of the physical properties of the species 5 and thermodynamic and kinetic reactions involved is also not required nor are the heat transfer properties of the reactor vessel required. Given the pertinent input parameters, a neural network can evaluate the input data and provide appropriate 10 output data for controlling the decarburization operation based upon the recognition of patterns between the input and output data which it has learned through a learning or training procedure involving the evaluation of random egamples presented 15 to the neural network thousands of times.
The processing of a computer to perform parallel distributive processing logic based upon neural models which simulate the operation of the human brain is, in general, referred to as "neural 20 networks". A neural network utilizes numerous nonlinear elements referred to as "neurons" to simulate the function of neurons in a human brain with each neuron representing a processing element.
Each processing element is connected to other 25 processing elements through a connecting weight or "synapse" which is combined by summation. The connecting weights are modified by adaptive learning from multiple egamples. Once trained, the neural network is capable of recognizing a pattern between 30 the input and output data which may be utilized, as hereinafter explained in detail, to provide information for controlling a decarburization operation without concern for the thermodynamic activity of the constituents in the bath and/or the kinetics of the reactions. The bath represents the mass of molten metal which is transferred to a 5 refractory lined vessel to be refined in accordance with the present invention.
Summary of the Invention In its broadest aspects, the present 10 invention is a method for refining steel by controlling the decarburization of a predetermined molten metal bath having a known composition of elements including carbon and having a known or estimated initial temperature and weight at the 15 outset of decarburization of a molten metal bath in a refractory vessel with said process of decarburization performed through the injection of -oxygen and a diluting gas into said bath under adjustable conditions of gas flow, comprising the 20 steps of:
(a) training a first neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, 25 weight and temperature at the outset of each process period, the gas ratio of o~ygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period, and the final temperature obtained at the conclusion of 30 each process period, until said first neural network is able to provide a substantially accurate output representing the counts of o~ygen required to be injected into said predetermined bath at any preselected gas ratio to cause the temperature of the bath to rise to a specified aim temperature level as a result of such gas injection;
(b) training a second neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, weight and temperature at the outset of the process 10 period, the gas ratio of oxygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period and the final carbon content obtained at the conclusion of each process period until the second neural 15 network is able to provide a substantially accurate output schedule of o~ygen counts to be injected into said predetermined bath to reduce the carbon level to a predetermined aim level in one or more successive stages corresponding to a preselected schedule of 20 ratios of oxygen to diluent gas;
(c) employing said first neural network to compute the o~ygen counts to be injected into said predetermined bath, from its known initial chemistry, weight and temperature at a first 25 preselected ratio of oxygen to diluent gas to raise the bath temperature to a specified aim temperature level;
(d) injecting o~ygen and diluent gas into said bath at said first preselected ratio until 30 the o~ygen counts computed by said first neural network are satisfied;
(e) employing said second neural network to provide an output schedule of oxygen counts to be injected into the bath from its known initial chemistry, weight and temperature to 5 successively reduce the carbon level in said bath to a predetermined aim carbon level in one or more stages corresponding to a preselected schedule of ratios of o~ygen to diluent gas; and (f) injecting oxygen and diluent gas 10 into said bath at said preselected schedule of oxygen counts corresponding to said output schedule as computed by said second neural network.
Brief Description of the Drawings Further advantages of the present invention will become apparent from the following detailed description of the invention when read in conjunction with the accompanying drawings of which:
Figure 1 is a general schematic diagram of a 20 decarburization system which utilizes the present invention;
Figure 2 is a schematic diagram of the type of neural network used in the present invention;
Figure 3 illustrates the preferred type of 25 transfer function used in training the neural network of Figure 2 in accordance with the training technique of Figure 4;
Figure 4 is a flowchart of the training technique for training a neural network in accordance 30 with the present invention; and Figure 5 is the preferred decarburization logic for the carrying out the process of decarburization in accordance with the present inventlon.
DescriPtion of a Preferred Embodiment The decarburization system as shown in Figure 1 includes a refractory lined vessel 10 charged with a predetermined mass of molten metal 12 having a known composition including carbon and other alloying constituents such as chromium, nickel, 10 manganese, silicon, iron and molybdenum in the production of steel particularly stainless steel, or nickel or cobalt based alloys. The weights of the liquid metal charged into the vessel is measured or estimated. The weight of solid additions, if any, 15 are independently computed, using conventional methods well known to those skilled in the art, to adjust the bath chemistry and weight to desired levels. Also the initial bath temperature is either estimated or measured. Conventional apparatus is 20 available to weigh the liquid metal charged into the vessel and to measure the temperature of the bath.
The flow of oxygen from a source (not shown) is regulated by a conventional o~ygen flow controller 14. Likewise, the flow of diluting gas from a source 25 (not shown) is regulated by a conventional gas flow controller 15. The gases are combined and injected directly into the melt 12 through a conventional tuyere assembly 16 or another suitable gas injector.
Following decarburization the molten metal 30 bath is reduced, finished and tapped with all of the finishing steps, including reduction, practiced in a conventional manner. The method of decarburization is achieved in accordance with the present invention by the injection of o~ygen and diluent gas, preferably subsurfacely, alone or in combination with a supply of o~ygen and/or a diluent gas blown from 5 above the bath. Alternatively, all oxygen and diluent qas, if any, may be blown onto the bath from above its surface. The diluent gas may be selected from the group consisting of argon, nitrogen and carbon dioxide. The metal bath is heated through the 10 exothermic oxidation reactions which take place during decarburization. If e~tra heat is needed, solid additions are added to the molten bath generally through the addition of aluminum and/or silicon with oxygen subsequently supplied to the bath 15 to oxidize those additions. The control of the slag chemistry is independent of the present invention.
The heat or bath of molten metal is generally blown at the maximum gas flow rate obtainable for the refining vessel and heat size 20 which is roughly 500 to 4,000 cubic feet per hour of total gas flow per ton of metal refining capacity for an AOD vessel and keeping the ratio of oxygen flow rate to the flow rate of diluent gas relatively high, preferably between 3:1 and 10:1, until the refractory 25 is threatened by high temperature. A given amount of o~ygen injected into the vessel is defined for purposes of the present invention as a count of oxygen or o~ygen "count". Likewise, a given amount of argon or other diluent gas to be injected into the 30 vessel is defined as a "count" of diluent gas.
A set of flowmeters 19 and 19' and a set of integrators 25 and 25' are used to measure the counts of oxygen and diluent gases injected into the bath 12. The ratio of oxygen to diluent gas is controlled by adjusting the flow of each gas through their respective flow controllers which can be manually or 5 automatically adjusted under the direction of the computer 18. The computer 18 is programmed to perform the decarburization logic as outlined in Figure 5 in conjunction with the selective operation of a plurality of neural networks numbered 1-5, 10 respectively. At least two neural networks are required in the performance of the present invention although the use of five (5) neural networks is preferred as will be explained in greater detail hereinafter.
A schematic representation of a typical neural network is shown in Figure 2 and comprises a layer of input processing units or "neurons"
connected to other layers of similar neurons through weighted connections or "synapses" in accordance with 20 the particular neural network model employed. The neural network internally develops algorithms of its own based on adjustments of the weighted connections through training.
The first or input layer of neurons is 25 referred to as the input neurons 22, whereas the neurons in the last layer are called the output neurons 24. The input neurons 22, and the output neurons 24 may be constructed from sequential digital simulators or a variety of conventional digital or 30 analog devices such as, for e~ample, operational amplifiers. Intermediate layers of neurons are referred to as inner or hidden neuron layers 26.
While only four hidden neurons are shown in a single hidden layer 26 in Figure 2, it will be understood that a substantially greater or lesser number of neurons and/or greater number of layers of hidden 5 neurons may be employed depending on the particular function assigned to such neural network. Each neuron in each layer is connected to each neuron in each adjacent layer. That is, each input neuron 22 is connected to each inner neuron 26 in an adjacent 10 inner layer. Likewise, each inner neuron 26 is connected to each neuron in the next adjacent inner layer which may comprise additional inner neurons 26. As shown in Figure 2, the ne~t layer may comprise the output neurons 24. Each neuron of the 15 output layer is connected to each neuron in the previous adjacent inner layer.
Each of the connections 27 between neurons contain weights or "synapses" (only some of the connections 27 are labeled in Figure 2 to avoid 20 confusion; however, numeral 27 is meant to include all connections 27). These weights may be implemented with digital computer simulators, variable resistances, or with amplifiers with variable gains, or with field effect transistor (FET) 25 connection control devices utilizing capacitors and the like. The connection weights 27 serve to reduce or increase the strength of the connections between the neurons. While the connection weights 27 are shown with single lines, it will be understood that 30 two individual lines may be employed to provide signal transmission in two directions, since this will be required during the training procedure. The value of the connection weight 27 may be any positive or negative value. When the weight is zero there is no effect in the connection between the two neurons.
The input neurons 22, inner neurons 26 and S output neurons 24 each comprise similar processing units which have one or more inputs and produce a single output signal. In accordance with the preferred embodiment, a conventional back propagation training algorithm is employed. Alternatively, other lO equivalent learning paradigms as known to those skilled in the art may be used. Back propagation requires that each neuron produce an output that is a continuous differentiable nonlinear or semi-linear function of its input. It is preferred that this 15 function, called a transfer function, be a sigmoid logistic non-linear function of the general form:
Yi ~ (1) 1 + e _ [~(Wj-X;) + ~]
Where Yi is the output of neuron i, ~(Wj-~) is the sum of the inputs to neuron i from the previous layer of neurons j, x; is the output of each neuron j in 25 the previous layer to neuron i, w; is the weight associated with each synapse connecting each neuron j in the previous layer to neuron i, and ~ is a bias similar in function to a threshold. The derivative of this function Yi with respect to its total input, 30 NETi ~ ~(wj-xj) + ~] is given by ~ Yi ' Yi-(l-Yi) (2) ~ NETi .
Thus, the requirement that the output is a differentiable function of the input is met. Other transfer functions could be used such as the hyperbolic tangent and the like.
The process of training a neural network to accurately calculate outputs involves adjusting the connection weights of each synapse 27 in a repetitive fashion based on known inputs until an output is produced in response to a particular set of inputs 10 which satisfies the training criteria or tolerance factor as exemplified in Figure 4, step E.
During training, the transfer function Yi remains the same for each neuron but the weights 27 are modified. Thus, the strengths of connectivity 15 are modified as a function of experience. The weights 27 are modified according to ~W~ Wj (3) 20 where ~Wj is the incremental adjustment to the e~isting weight wj, ~i is an error signal available to the neuron, and ~ is a constant of proportionality also called the learning rate.
The determination of the error signal ~i is 25 a recursive process that is propagated backward from the output neurons. First, input values are transmitted to the input neurons 22. This causes computations in accordance with Equation 1 or those of a similar transfer function to be transmitted 30 through the neural network of Figure 2 until an output value is produced. It should be noted from Figure 3 that the transfer function Yi cannot reach the e~treme limits of minus one or plus one without D-16~06 infinitely large weights. ~he calculated output of each output neuron 24 is then compared to the output desired or known to be correct from the training data. For output neurons the error signal is ~i - (Di-Yi) ~-Yi - (4) a NETi where Di is the desired output of the given output neuron. By substituting Equation 2 into Equation 4 using the sigmoid transfer function the error signal 10 for output neurons i can be restated as follows:
~i e (Di-Yi)(Yi)(l-yi) (5) For hidden neurons 26 there is no specific desired output from the measured data, so the error signal is determined recursively in terms of the 15 error signals in the output or successive hidden layer neurons k to which the hidden layer neurons directly connect and the weights of those connections. Thus, for non-output neurons ~i ' Yi(l-Yi)~(~k-Wk) (6) 20 where ~k is the error signal of respective output or successive hidden layer neurons k to which the hidden neuron i is connected and Wk is the weight between that neuron k and the hidden neuron i.
From Equation 3 it can be seen that the 25 learning rate ~ will affect how greatly the weights are changed each time the error signal ~; is - propagated. The larger ~, the larger the changes in the weights and the faster the learning rate. If, however, the learning rate is made too large the 30 system can oscillate during learning. Oscillation can be avoided even with large learning rates by using a momentum term a. Thus, ~Wi,n~ iYi ~ ~ ~Wi,n (7 may be used in place of Equation 3 where ~Wi n+l is the present adjustment of Wi and ~Wi n is the previous adiustment of Wi.
The constant a determines the effect ~f past weight changes ~Wi n on the current direction of movement in weights ~Wi n+l providing a kind of momentum in weights that effectively filters out high frequency oscillation in the weights.
Training is accomplished by first collecting sets of input and output data from many actual decarburization operations to be presented as training data in random order to the neural networks. Data is collected defining the initial 15 contents of the chemical constituents of a molten metal bath, the initial bath temperature and weight, the weights of the solid additions added during the blow period, the ratio of oxygen to diluent gas blown and the final temperature obtained whereas output 20 data includes the counts of oxygen and diluent gas injected into the bath. E~amples of solid additions used during decarburization are the flu~es such as lime, dolomitic lime or magnesia, the base material used as a source of iron units in the case of ferrous 25 metal refining, cobalt units in the case of cobalt base metal refining or nickel units in the case of nickel based metal refining, ferro-chrome, ferro-manganese, nickel and ferro-nickel. The parameter to be used as the inputs and the parameter 30 to be used as the outputs for each of the neural networks will vary based upon the function of the network.
Each of the neural networks 1 to 5 are assigned different functions and are trained to recognize and identify the requirements needed to perform such functions during the decarburization 5 operation. For example, the first neural network 1 is assigned the function of determining the gas, injection requirements, i.e. the counts of o~ygen at a preselected ratio of oxygen to diluent gas to reach a specified bath temperature from the initial 10 chemistry, temperature and weight of the bath 12 charged in the vessel 10. The second neural network 2 may be assigned the function of determining the gas injection requirements to reach a specified carbon content from the initial chemistry, 15 temperature and weight of the bath 12 charged in the vessel 10 using a preestablished gas ratio schedule.
A third neural network may be assigned the function of determining the carbon content in the molten metal bath after the gases have been injected 20 in satisfaction of the computation of either of the first two neural networks. The fourth neural network is assigned the function of computing the bath temperature and the fifth neural network computes the silicon, manganese, chromium, nickel, and molydenum 25 contents of the bath at the completion of the injection of oxygen for the preestablished ratio of o~ygen to diluent gas in accordance with either neural network 1 or 2 based upon the input data of the initial bath chemistry, temperature and weight, 30 the counts of oxygen injected and the ratio of o~ygen to diluent gas used. The input data of initial conditions may represent either the initial conditions when the molten metal is transferred to the refining vessel or the initial conditions esisting at the commencement of any process period i.e, blow period within a decarburization operation 5 as will be explained hereafter in greater detail.
Thus the neural networks 1-2 provide the decarburization oxygen counts required to decarburize the molten metal bath pursuant to the decarburization logic of Figure 5. The computer 18 follows the logic 10 requirements of Figure 5 in performing the decarburization operation in compliance with the computation of the neural networks 1-2 respectively.
For purposes of the sub]ect invention neural network 1 is used to determine the amount of o~ygen 15 required to be injected into the bath to reach a specified aim temperature level and has ten respective input neurons 22 for the initial conditions including the initial carbon, silicon, manganese, chromium, nickel and molybedenum contents 20 of the bath, the initial temperature and weight of the bath, the specified aim temperature of the bath and the ratio of o~ygen to diluent gas to be used.
An additional six input neurons are used for the weights of each of sig types of solid additions which 25 may be added during the blow period as hereinabove identified. Thus neural network 1 is constructed of sixteen input neurons 22, one output neuron 24 for indicating the counts of oxygen required to reach the specified aim temperature level and eight hidden or 30 inner neurons 26 in a single layer.
Neural network 2 is used to determine the amount of oxygen required to reach a specified carbon content, and similarly to network 1, has ten input neurons 22 for the initial carbon, silicon, manganese, chromium, nickel and molydenum constituents of the bath, the initial bath 5 temperature and weight, the desired aim carbon content and the ratio of o~ygen to diluent gas. An additional six input neurons are used for the si~
solid addition types which may be added during the blow period. Thus neural network 2 is constructed of 10 seventeen input neurons 22 and one output neuron 24 for indicating the counts of ogygen required to reach the specified aim carbon content and has eight hidden or inner neurons 26 in a single layer.
Neural network 3 is used to determine the 15 carbon content reached by injecting a specified amount of oxygen at a specified ratio of oxygen to diluent gas into known initial bath conditions and has respective input neurons 22 for the initial carbon, silicon, manganese, chromium, nickel and 20 molybdenum contents of the bath, the initial bath temperature and weight, the specified amounts of oxygen and diluent gases injected, and the ratio of o~ygen to diluent gas blown and the weights of each of the addition types added during the blow period.
25 A network with si~ types of additions is thus constructed of seventeen input neurons. The network has one output neuron for the carbon content resulting from the specified gas injection and has nine hidden neurons in a single layer.
Neural network 4 is used to determine the temperature reached by injecting a specified amount of oxygen at a specified ratio of oxygen to diluent gas into known initial bath conditions and has respective input neurons 22 for the initial carbon, silicon, manganese, chromium, nickel and molybdenum contents of the bath, the bath temperature and 5 weight, the weights of each of the addition types added during the blow period, the specified amounts of o~ygen and diluent gases injected, the elapsed time, and the ratio of oxygen to diluent gas blown.
A network with si~ types of additions is thus 10 constructed of eighteen input neurons. The network has one output neuron for the temperature resulting from the specified gas injection and has nine hidden neurons in a single layer.
Neural network 5 is used to determine the 15 silicon, manganese, chromium, nickel, and molybdenum contents of the bath following the injection of specified amounts of oxygen and diluent gases at a specified ratio of ogygen to diluent gas into known initial bath conditions. Neural network 5 has 20 respective input neurons for the initial carbon, silicon, manganese, chromium, nickel and molybdenum contents of the bath, the bath temperature and weight, the weights of each of the addition types added during the blow period, the specified amounts 25 of o~ygen and diluent gases injected and the ratio of oxygen to diluent gas blown. A network with sig types of additions is thus constructed of seventeen input neurons. The network has five output neurons for the silicon, manganese, chromium, nickel, and 30 molybdenum contents, respectively, resulting form the specified gas injection and has eleven hidden neurons in a single layer.
Although a single layer of hidden neurons is used, it is within the scope of the present invention to use a greater or lesser number of hidden layers of neurons. The e~act configuration is best established 5 empirically. This applies to the number of hidden neurons within a hidden layer and the number of hidden layers chosen for each of the neural networks.
Input and output data from many actual decarburization operations are used to train the 10 neural networks with data separately collected to correspond to multiple process periods in each decarburization operation. Data is collected for each process period in which only one discreet ratio of oxygen to diluent gas is injected at any time in a 15 single process period. A process period is herein defined as the time between two consecutive samples of bath chemistry and temperature for a given decarburization operation, i.e., within a single heat. The time interval between samples may be short 20 or long in a random relationship. Thus the process periods have no defined time relationship or chronology. Pure diluent gas stirring may also be performed or the vessel may be idle during portions of the process period or additions may be added at 25 any time concurrent with any of these events during process periods from which the data is collected for purposes of training the neural networks. The data ,should be collected in such a way that the ranges of useful or expected input and output values are 30 represented. For instance, for AOD refining it is best to have initial carbon contents of from 0.1% to 1.8% in the molten metal as initial conditions for various process periods and have data for process periods using o~ygen to diluent gas ratios from 4 to 1 to ratios of 1 to 3. Pure diluent gas decarburization data would also be needed to 5 accurately model a practice which uses this technique. Preferably, at least 10 process periods of data should be collected at each oxygen to diluent gas ratio, although the accuracy of the neural network is enhanced by greater amounts of data.
An example of a block of input and output training data for the neural networks 1-5 is set forth in the following Table:
TABLE
INITIAL
ELAPSED COUNTSCOUNTS COUNTS INITIAL INITIAL INITIAL INITIAL INITIAL INITIAL INITIAL METAL
RATIO TIME 02 NZ AR TEMP ~Frc %si 7~R 7MN 7Ni 7MoWEIGHT
lbs 0.000 4.0000.000 64.00039.000 2884.00 1.300 0.25019.6800.620 6.3400.26 109333 10 3.000 8.000 209.00081.000 0.0002792.000 1.2400.24019.630 0.6406.370 0.25 109202 3.000 9.000300.000130.0000.000 2942.0001.080 0.09019.4800.600 6.4000.25 109700 1.00015.000344.000370.0000.000 2947.0000.800 0.08017.9201.330 6.9700.26 114794 3.00010.000412.000143.0000.000 2751.0001.200 0.17019.2400.610 6.4600.13 101000 0.000 6.0000.000 67.0000.000 2982.0000.680 0.09018.6600.560 6.5600.13 99808 15 3.00011.000 299.000142.000 0.0002778.000 0.6500.10017.360 1.4206.900 0.13 109985 1.00012.000243.000272.0000.000 2952.0000.450 0.10016.8001.160 6.9900.13 108157 0.000 4.0000.000 57.0000.000 2849.0000.160 0.21018.7700.610 6.9701.56 99667 3.00011.000406.000134.0000.000 2770.0001.120 0.19018.7800.610 6.9701.61 99607 0.000 5.0000.000 74.0000.000 2997.0000.620 0.10018.2500.550 7.0501.58 98491 20 3.00011.000 398.000165.000 0.0002690.000 0.6800.11017.150 1.3708.370 1.55 109798 1.000 8.000147.000173.0000.000 2980.0000.390 0.09016.3901.060 8.4601.57 108623 0.33323.000106.000209.000116.000 3037.0000.200 0.09016.1801.020 8.4901.56 108189 0.000 5.0000.000 68.0000.000 2772.0001.440 0.26018.2700.550 3.8700.19 106100 4.00012.000465.000139.0000.000 2680.0001.390 0.23018.4000.560 3.8500.19 106015 25 0.000 9.000 0.00088.000 0.0002971.000 0.9400.07018.040 0.5103.920 0.20 105093 3.00014.000456.000188.0000.000 2703.0001.030 0.09017.2801.750 7.8600.21 114993 1.000 9.000185.000204.0000.000 2972.0000.550 0.08016.7501.470 7.9600.21 113820 4.000 4.00034,000111.0000.000 2829.0001.550 0.17019.0700.540 6.5900.36 102667 4.00011.000331.000144.0000.000 2769.0001.520 0.13018.8600.540 6.6600.36 102379 30 0.000 5.000 0.00054.000 0.0002844.000 1.3900.18018.730 0.5704.280 0.34 101667 4.00011.000362.000122.0000.000 2752.0001.240 0.17018.7100.580 4.2900.35 101484 3.000 6.000194.00091.0000.000 2943.0000.850 0.17018.4500.540 4.2900.35 100824 3.000 6.000157.00077.0000.000 2860.0000.720 0.08016.9801.560 7.0000.36 109271 1.000 5.00091.000112.0000.000 2947.0000.540 0.08016.8601.560 7.0400.36 108943 35 0.33339.000 356.000759.000 149.000Z977.000 0.4100.08016.690 1.5407.060 0.36 108616 0.000 5.0000.000 55.0000.000 2840.0001.210 0.30018.6500.660 3.5502.10 96333 4.00011.000454.000142.0000.000 2746.0001.200 0.30018.6500.660 3.5502.08 96324 0.00012.0000.000 207.0000.000 3060.0000.690 0.30018.6500.660 3.5502.08 95832 3.00013.000458.000184.0000.000 2546.0000.690 0.10017.5301.390 8.4002.07 111824 40 1.000 9.000 191.000215.000 0.0002942.000 0.5300.07016.550 1.0908.530 2.07 110516 0.000 5.0000.000 72.0000.000 2826.0001.580 0.12019.0200.600 3.6300.39 104500 ~ o ~ O ~ ~) ~ ~ ~ 0 U') "~ ~ N ~ 7 ~ ~ ~ ~ ~ ~ ~1 1~ 0 0 0 0 0 0 0 0 0 -- -- -- -- -- O O O O O O O O O O O O O ~'J t'i N N ~'J O
J O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O
~ o ~ o cr~ o ~ o u ~ ~ ~D O ~D _ O ~ O ~ ~ U7 0 ~ ~' z z f~ ~ o Ln ~ o~ o ~ ~ ~ ~ o ~ I ~1 0 0 0 0 O ~ ~ ~ 1~ 0 0 0 0 ~ ~ 1~1~ 0 ~D ~ ~ ~ r~I~ r~I' ~ r7 0 0 CD ~
J O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O
z ~ ~ ~ ~ ~ U~ ~ _ O ~ U~ ~ O O O U~ O ~ O O ~D
~ o o - - o - - - o o - - - - o o - - - o o o o - - - - o o - - - o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o J ~ 0 ~ u~ ~ ~o o ~ 0 u7 u~ a~ 0 o o ~ 0 u~ ~ ~ o _ u > 0 ~o ~ o Ir~
Z ~ 0 ~ D 0 0 1~ 0 0 1~. ~ D 0 W 0 0 ~ D 0 0 ~ O 0 o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o ~r ~ ~ ~ 0 ~ a~ o o 0 o~ o ~ c~ 0 u:~ ~ 0 1~ ~ 0 a) 0 1~ o o o z ~ ~ o o o o - - o - - ~ o o o ~ o o o o - o - - o o o o ~ ~ - o o l~ o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o ~ o o o o o o o o o o o o o 0 o o o o o o o o o o ~ o o o o o o ~S ~r 0 0 ~t 0 U~ Ul 1' OJ O~ 0 CJ~ O ~) C7~ ~ 0 U7 Cr~
Z ~ ~ 0 0 ~ ~O ~ D ~ ~ O ~ Cr~ o u~ 0 r~ ~ ~ o ~ ~
-- -- o o o o o o -- o o o o o -- o -- o o -- ~ -- o o o o o . o o O o _ ~, C ", ~ ~ ~ 1~ _ N CC~ N 1~ o ~ O O ~ cr O -- 0 N O 0~ 0 t~l ~) O 1~ r~ a~ ~D O 'D N 1~ --~' Z ~ ~ 1' ~ ~ 1' 0 ~ u~ 0 r~) 0 0 ~ o ~ o ~c> ~ u~ ~ ~o ~ ~ ~ ~ ~D In ~ _ 0 'c ~ O ~ ~ O O ~ I~ C~ O r' o~ o 0 ~ ~ r' cr~ o r' G~ 0 ~ O~ O 1' 0 ~ O O
o ~ 1~ 'J ~ 0 ~ N ~ ) ~J ~ OJ ~ 0 ~J ~ ~ ~ N ~) ~ J ~ N C~l 0 ~'J 0 C~
-O
~ ~ ~ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1' 0 0 0 0 o~ O ~ O O o O O ~ O O o o o O ~ O 0 t~) O~ U') z ~ ~ J~ O O O O O O O O O O ~ O O O O O 1' 0 0 0 0 0 0 0 0 0 0 0 0 U~ O O O
g O~
~r _ z o o o o o o o o o o ~ o o o o o o o o o o o o o o o o o o l-- o o o ~ o ~o - ~ ~ ~
o ~-~ ~ v~ o o o o o r~ o o o o ~ o o o o o 1' o o o o o o 1-- o o o o o ~ o o o O ~ O In cn o - ~ ~
O ~u ~ o O ~ 0 0 ~0 0 o o o " o o o o o 0 0 0 0 0 0 0 ~ 0 0 0 0 0 ~ O O O
o ~J o o ~ o o ~ o o o o ~ o o o o o o o o o o o o l' o o o o o o o o o O ~ .D 0 ~ _ ~
u o u~ o ~n o ~ o -- t~J N 0 0 ~
Each network is trained using the standard back propagation paradigm. Training should use either a hyperbolic tangent, or preferably a sigmoid transfer function, a learning rate of 0.1 and a 5 momentum of zero for each neuron. Once the neural network is sufficiently trained, it is translated to a readily usable programming language such as C or BASIC or FORTRAN. The code in one of these languages is compiled and linked as necessary.
A flowchart indicative of the training operation is shown in Figure 4. Pursuant to Step A
the weights and offset are set to small random values between one and minus one. The collected training input and output data for a given process period are 15 then presented to the neural network input neurons 22 under training as indicated in Step B. After the input data is propagated through the inner layer of neurons 26 to the output neurons 24, an output 20 as shown in Step C is formed for each output neuron 24 20 based on the transfer function Yi described in Equation (1). The calculated output 20 from the output neurons 24 is compared in Step D to the output data of the given process period to develop an error signal 30 using Equations 5 and 6 for the output and 25 hidden neurons respectively. The error signal 30 is then compared to a preset tolerance factor in Step E. If the error signal 30 is larger than the tolerance factor, the error signal 30 as shown in Step F makes a backward pass through the network 30 using Equation 7 for adjusting the weights to the output and hidden neurons and each weight in Step A
is incrementally changed by ~Wi. Input data of another process period is presented and Steps B
through E are repeated until the error signal 30 is reduced to an acceptable level. When the error signal 30 is smaller than the preset tolerance factor S the training procedure pursuant to Step G is complete.
For purposes of verification the verification Steps H and I are followed in which test inputs are presented to generate outputs 20 as in Step C for comparison in Step D with known outputs.
10 The tolerance factor is an e~ternally determined standard for the desired accuracy of the neural network. The training is continued until the error signal is less than this tolerance. The simplest form of a tolerance is to assign a certain percentage 15 error for training to stop. A more practical form of tolerance is to test whether the neural network is in fact learning to generalize the relationships between the problem's inputs and outputs or whether it has begun to memorize those relationships for the 20 specific data with which it trains itself. After a periodic number of iterations the neural network is applied to the reserve or test data and its ability to estimate the desired output for that data is assessed. In the early stage of training the neural 25 network will learn to estimate the test outputs with increasing accuracy. After the neural network has completed generalization, it begins to increase its accuracy relative to the training data at the e~pense of its accuracy relative to the test data. At this 30 point the training is considered to have reached the optimum configuration or weights for general problem solvinq, and the training process is stopped. Each neural network 1-5 is trained in the aforementioned manner.
The determination of the error signal 30 is a recursive process that starts by generating outputs 5 from the output neurons 24 based on feeding the collected data to the input neurons 22. The input neurons 22 cause a signal to be propagated forward through the neural network until an output signal is produced at the output neuron 24. From equation 3 it 10 can be seen that the learning rate ~ will effect how much the weights are changed each time an error signal is propagated. The larger ~, the larger the changes in the weights and the faster the learning rate at the possible expense of the accuracy that may 15 eventually be obtained.
The total population of collected input and output data should be randomly divided into two groups. The larger group should be used as training data for training the neural network with the 20 remaining smaller group of data used as test data for verification. One reasonable division is to use 75%
of the collected data for training purposes and to use the remaining 25% of the collected data as test data to verify the network's predictive accuracy.
25 The neural network should be trained until comparisons to the verification data show that the model's accuracy is not increasing. At this point, those skilled in the art will know that the network is no longer learning to generalize the problem, but 30 is rather memorizing the specific solutions for the training set of data. The learning process typically takes 10,000 to 500,000 presentations of process periods, i.e, presentations of individual sets of complete input and output data for a given process period, to the network for adjustment of its weights. The order of presenting the process periods 5 within the entire training set of data to the neural network for training should be randomly shuffled after each time the entire set has been presented to the network for training.
The sequence of using the trained neural 10 networks 1-5 is determined in accordance with the decarburization logic shown in Figure 4. The composition, weight and temperature of the bath at the time of transfer to the refining vessel is estimated or measured. The calculations of the solid 15 additions are independently calculated and do not form part of the present invention. The decarburization logic shown in Figure 9 is an illustrative example of the invention using neural networks 1-5 based on a predetermined initial 20 decarburization o~ygen to diluent gas setting and a predetermined oxygen to diluent gas decarburization ratio schedule. The e~ample of Figure 4 uses a preselected aim temperature level of 3050~F for a ratio of 4 to 1 oxygen to diluent gas and a ratio 25 schedule of 1, .333 and 0 for the successive aim carbon levels of .15%C, .05%C and .03%C
respectively. The decarburization logic establishes decision trees to determine when to use the neural networks 1-5.
Decarburization proceeds only if the carbon level is above the ultimate aim level of 0.03% C. If the bath temperature is less than 3050~F and calculated solid additions have yet to be added to the bath, a ration of 4 to 1 oxygen to diluent gas is selected and neural network 1 is activated to compute the oxygen counts necessary to raise the temperature 5 of the bath to the preselected level of 3050~F. Upon supplying o~ygen equal to the computed counts calculated by neural network 1 the neural networks 3, 4 and 5 are activated or fired to compute the updated conditions of carbon content, bath temperature and 10 metal chemistry upon completion of said injection.
Neural network 1 is again activated with the aforementioned outputs of neural networks 3, 4 and 5 as the new initial conditions and the required solid additions also used as new inputs to compute the 15 oxygen count necessary to raise the bath temperature to the preselected level of 3050~F while simultaneously adding said additions. O~ygen is injected at the preselected ratio of 4 to 1 while the said additions are added until the computed oxygen 20 counts are satisfied.
If the bath temperature is less than 3050~F
and no solid additions have yet to be added to the bath, a ratio of 4 to 1 o~ygen to diluent gas is selected and neural network 1 is activated to compute 25 the o~ygen counts necessary to raise the temperature of the bath to the preselected level of 3050~F. Upon supplying oxygen equal to the computed counts calculated by neural network 1 the neural networks 3, 4 and 5 are activated to compute updated conditions 30 of carbon content, bath temperature and metal chemistry.
..
If the bath temperature computed by neural networks 3, 4 and 5 equals or exceeds the predetermined aim temperature level of 3050~F a new ratio of oxygen to diluent gas is specified 5 corresponding to a ratio of 1/1, 1/3 or zero, respectively, with the determination based upon the temperature and carbon concentration such that if the temperature is between 3050~F and 3100~F and the carbon concentration exceeds .15% the ratio of 1/1 is 10 specified, whereas if the temperature is equal to or greater than 3050~F and the carbon content is between .08% and .15% a ratio of 1/3 is specified and finally if the temperature exceeds or equals 3050~F and the carbon content is less than .08% a zero ratio is 15 specified. For any of these conditions neural network 2 is activated, the appropriate oxygen to diluent gas ratio is chosen and the required oxygen gas counts are computed to reach the aim carbon level. Oxygen and/or diluent gas is then blown at 20 the specified ratio until the oxygen counts as computed by neural network 2 are satisfied. The neural networks 3, 4 and 5 are then activated after each successive step to update the bath chemistry, temperature and carbon content for the initial 25 condition of any subsequent decarburization.
An AOD process was run using a conventional thermodynamic model for predicting and controlling the decarburization process during the production of both ASTM 300 series and ASTM 400 series stainless 30 steels. Upon adjusting the constants in the model to attain optimal accuracy, the carbon content could be predicted with a standard deviation of 0.11% carbon for actual carbon contents between 0.1% and 0.3%.
Fourteen heats of stainless steels were sampled after the use of each ratio of oxygen to diluent gas to measure the bath chemistry and temperature. The 5 information was used for training the first neural network of the present invention. The trained neural network was then used to predict the carbon content at carbon contents between 0.1% and 0.3% carbon during the production of the same grades of stainless 10 steels. The carbon content prediction using the said neural network had a standard deviation of only 0.035% carbon.
THE ~FINING OF STEEL USING NEURAL NETWORKS
5 ~ield of the Invention This invention relates to an AOD process for decarburizing molten metal in the refining of steel and more particularly to an AOD process for decarburizing molten metal using neural networks to 10 control the decarburization operation.
Backqround of the Invention A process which has received wide acceptance in the steel industry for refining metal is the 15 argon-02ygen decarburization process also referred to as the "AOD" process. It is the purpose of AOD
refining to first remove carbon from a bath of metal, ne~t reduce any metals that may have oxidized during decarburization, and finally adjust the temperature 20 and chemistry of the bath before casting the metal into a product. Decarburization is achieved by injecting mixtures of 02ygen and inert gases in such a way as to favor the o~idation of carbon over the oxidation of other metal components present in the 25 bath. At progressively lower carbon contents during the process of decarburization progressively greater dilution of the 02ygen by inert gases is injected to favor the oxidation or removal of carbon.
Relationships between the bath weight, 30 chemistry, and temperature, the injections of o~ygen and inert gases, and the resultant changes in metal chemistry and temperature have been theorized to achieve both control and understanding of how to optimize the economics of the process. Thermodynamic models have tracked the general relationships between these parameters, but have limited accuracy and have not obviated the need for intermediate sampling of the bath temperature and chemistry in processing any 5 given heat of metal. Some theorists have adopted the approach that the decarburization reaction may be better understood, and hence controlled, by considering the chemical kinetics of the competing o~idations of carbon and the various metal species 10 present. It follows that approaches incorporating both thermodynamic and kinetic considerations have also been constructed. Finally, statistical approaches have been used to empirically model decarburization in an AOD converter.
The traditional modeling of the decarburization cycle of the AOD operation requires not only a comprehensive understanding of how to represent the thermodynamics and/or kinetics for use in a computer program, but also requires the 20 knowledge of many properties of the species involved in the reactions. For instance, normal thermodynamic modeling requires the knowledge of at least 25 pertinent interaction coefficients. The free enthalpies and entropies associated with each 25 potential reaction must also be known as well as a representative pressure exerted on the bubbles passing through and reacting with the bath. Kinetic models that are based on assumptions that diffusion, adsorption and desorption rates significantly affect 30 the relative extents to which the competing o~idation reactions occur are similarly dependent on accurate knowledge of these rates with respect to temperature and base composition. They must also be capable of modeling the surface areas, velocities of the bubbles relative to the surrounding liquid, and the residence times of the bubbles in the metal phase. Thus, the 5 modeling of decarburization based on chemical theories is subject to many items of data being all accurately measured. They also require a correct understanding of the mechanisms of the various reactions. Since models are deficient in at least 10 one of these two requirements, it is normal for known physical "constants" to be altered to make the results of the model fit actual results better. Due to the complexity of these models, great skill is required to adjust the parameters to improve the 15 overall accuracy of an entire population of results.
Often it is found that one particular solution or combination of adjusted constants is optimal for representing the results of only one particular set of working conditions. That is, solutions tend not 20 to be general, but rather geared to specific small sets of data for which they were adjusted.
In spite of the variety of approaches, inaccuracies remain and some form of measuring the carbon content during the decarburization process 25 step is normally required. This usually necessitates halting the process, withdrawing a metal sample, analyzing the carbon content and measuring the bath temperature before resuming. Lack of process control during decarburization not only necessitates extra 30 sampling, but precludes operation at the optimal conditions for cost reduction and production maximization.
A computerized system using "neural networks" benefits from the fact that a theoretical understanding of decarburization is not required.
Xnowledge of the physical properties of the species 5 and thermodynamic and kinetic reactions involved is also not required nor are the heat transfer properties of the reactor vessel required. Given the pertinent input parameters, a neural network can evaluate the input data and provide appropriate 10 output data for controlling the decarburization operation based upon the recognition of patterns between the input and output data which it has learned through a learning or training procedure involving the evaluation of random egamples presented 15 to the neural network thousands of times.
The processing of a computer to perform parallel distributive processing logic based upon neural models which simulate the operation of the human brain is, in general, referred to as "neural 20 networks". A neural network utilizes numerous nonlinear elements referred to as "neurons" to simulate the function of neurons in a human brain with each neuron representing a processing element.
Each processing element is connected to other 25 processing elements through a connecting weight or "synapse" which is combined by summation. The connecting weights are modified by adaptive learning from multiple egamples. Once trained, the neural network is capable of recognizing a pattern between 30 the input and output data which may be utilized, as hereinafter explained in detail, to provide information for controlling a decarburization operation without concern for the thermodynamic activity of the constituents in the bath and/or the kinetics of the reactions. The bath represents the mass of molten metal which is transferred to a 5 refractory lined vessel to be refined in accordance with the present invention.
Summary of the Invention In its broadest aspects, the present 10 invention is a method for refining steel by controlling the decarburization of a predetermined molten metal bath having a known composition of elements including carbon and having a known or estimated initial temperature and weight at the 15 outset of decarburization of a molten metal bath in a refractory vessel with said process of decarburization performed through the injection of -oxygen and a diluting gas into said bath under adjustable conditions of gas flow, comprising the 20 steps of:
(a) training a first neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, 25 weight and temperature at the outset of each process period, the gas ratio of o~ygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period, and the final temperature obtained at the conclusion of 30 each process period, until said first neural network is able to provide a substantially accurate output representing the counts of o~ygen required to be injected into said predetermined bath at any preselected gas ratio to cause the temperature of the bath to rise to a specified aim temperature level as a result of such gas injection;
(b) training a second neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, weight and temperature at the outset of the process 10 period, the gas ratio of oxygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period and the final carbon content obtained at the conclusion of each process period until the second neural 15 network is able to provide a substantially accurate output schedule of o~ygen counts to be injected into said predetermined bath to reduce the carbon level to a predetermined aim level in one or more successive stages corresponding to a preselected schedule of 20 ratios of oxygen to diluent gas;
(c) employing said first neural network to compute the o~ygen counts to be injected into said predetermined bath, from its known initial chemistry, weight and temperature at a first 25 preselected ratio of oxygen to diluent gas to raise the bath temperature to a specified aim temperature level;
(d) injecting o~ygen and diluent gas into said bath at said first preselected ratio until 30 the o~ygen counts computed by said first neural network are satisfied;
(e) employing said second neural network to provide an output schedule of oxygen counts to be injected into the bath from its known initial chemistry, weight and temperature to 5 successively reduce the carbon level in said bath to a predetermined aim carbon level in one or more stages corresponding to a preselected schedule of ratios of o~ygen to diluent gas; and (f) injecting oxygen and diluent gas 10 into said bath at said preselected schedule of oxygen counts corresponding to said output schedule as computed by said second neural network.
Brief Description of the Drawings Further advantages of the present invention will become apparent from the following detailed description of the invention when read in conjunction with the accompanying drawings of which:
Figure 1 is a general schematic diagram of a 20 decarburization system which utilizes the present invention;
Figure 2 is a schematic diagram of the type of neural network used in the present invention;
Figure 3 illustrates the preferred type of 25 transfer function used in training the neural network of Figure 2 in accordance with the training technique of Figure 4;
Figure 4 is a flowchart of the training technique for training a neural network in accordance 30 with the present invention; and Figure 5 is the preferred decarburization logic for the carrying out the process of decarburization in accordance with the present inventlon.
DescriPtion of a Preferred Embodiment The decarburization system as shown in Figure 1 includes a refractory lined vessel 10 charged with a predetermined mass of molten metal 12 having a known composition including carbon and other alloying constituents such as chromium, nickel, 10 manganese, silicon, iron and molybdenum in the production of steel particularly stainless steel, or nickel or cobalt based alloys. The weights of the liquid metal charged into the vessel is measured or estimated. The weight of solid additions, if any, 15 are independently computed, using conventional methods well known to those skilled in the art, to adjust the bath chemistry and weight to desired levels. Also the initial bath temperature is either estimated or measured. Conventional apparatus is 20 available to weigh the liquid metal charged into the vessel and to measure the temperature of the bath.
The flow of oxygen from a source (not shown) is regulated by a conventional o~ygen flow controller 14. Likewise, the flow of diluting gas from a source 25 (not shown) is regulated by a conventional gas flow controller 15. The gases are combined and injected directly into the melt 12 through a conventional tuyere assembly 16 or another suitable gas injector.
Following decarburization the molten metal 30 bath is reduced, finished and tapped with all of the finishing steps, including reduction, practiced in a conventional manner. The method of decarburization is achieved in accordance with the present invention by the injection of o~ygen and diluent gas, preferably subsurfacely, alone or in combination with a supply of o~ygen and/or a diluent gas blown from 5 above the bath. Alternatively, all oxygen and diluent qas, if any, may be blown onto the bath from above its surface. The diluent gas may be selected from the group consisting of argon, nitrogen and carbon dioxide. The metal bath is heated through the 10 exothermic oxidation reactions which take place during decarburization. If e~tra heat is needed, solid additions are added to the molten bath generally through the addition of aluminum and/or silicon with oxygen subsequently supplied to the bath 15 to oxidize those additions. The control of the slag chemistry is independent of the present invention.
The heat or bath of molten metal is generally blown at the maximum gas flow rate obtainable for the refining vessel and heat size 20 which is roughly 500 to 4,000 cubic feet per hour of total gas flow per ton of metal refining capacity for an AOD vessel and keeping the ratio of oxygen flow rate to the flow rate of diluent gas relatively high, preferably between 3:1 and 10:1, until the refractory 25 is threatened by high temperature. A given amount of o~ygen injected into the vessel is defined for purposes of the present invention as a count of oxygen or o~ygen "count". Likewise, a given amount of argon or other diluent gas to be injected into the 30 vessel is defined as a "count" of diluent gas.
A set of flowmeters 19 and 19' and a set of integrators 25 and 25' are used to measure the counts of oxygen and diluent gases injected into the bath 12. The ratio of oxygen to diluent gas is controlled by adjusting the flow of each gas through their respective flow controllers which can be manually or 5 automatically adjusted under the direction of the computer 18. The computer 18 is programmed to perform the decarburization logic as outlined in Figure 5 in conjunction with the selective operation of a plurality of neural networks numbered 1-5, 10 respectively. At least two neural networks are required in the performance of the present invention although the use of five (5) neural networks is preferred as will be explained in greater detail hereinafter.
A schematic representation of a typical neural network is shown in Figure 2 and comprises a layer of input processing units or "neurons"
connected to other layers of similar neurons through weighted connections or "synapses" in accordance with 20 the particular neural network model employed. The neural network internally develops algorithms of its own based on adjustments of the weighted connections through training.
The first or input layer of neurons is 25 referred to as the input neurons 22, whereas the neurons in the last layer are called the output neurons 24. The input neurons 22, and the output neurons 24 may be constructed from sequential digital simulators or a variety of conventional digital or 30 analog devices such as, for e~ample, operational amplifiers. Intermediate layers of neurons are referred to as inner or hidden neuron layers 26.
While only four hidden neurons are shown in a single hidden layer 26 in Figure 2, it will be understood that a substantially greater or lesser number of neurons and/or greater number of layers of hidden 5 neurons may be employed depending on the particular function assigned to such neural network. Each neuron in each layer is connected to each neuron in each adjacent layer. That is, each input neuron 22 is connected to each inner neuron 26 in an adjacent 10 inner layer. Likewise, each inner neuron 26 is connected to each neuron in the next adjacent inner layer which may comprise additional inner neurons 26. As shown in Figure 2, the ne~t layer may comprise the output neurons 24. Each neuron of the 15 output layer is connected to each neuron in the previous adjacent inner layer.
Each of the connections 27 between neurons contain weights or "synapses" (only some of the connections 27 are labeled in Figure 2 to avoid 20 confusion; however, numeral 27 is meant to include all connections 27). These weights may be implemented with digital computer simulators, variable resistances, or with amplifiers with variable gains, or with field effect transistor (FET) 25 connection control devices utilizing capacitors and the like. The connection weights 27 serve to reduce or increase the strength of the connections between the neurons. While the connection weights 27 are shown with single lines, it will be understood that 30 two individual lines may be employed to provide signal transmission in two directions, since this will be required during the training procedure. The value of the connection weight 27 may be any positive or negative value. When the weight is zero there is no effect in the connection between the two neurons.
The input neurons 22, inner neurons 26 and S output neurons 24 each comprise similar processing units which have one or more inputs and produce a single output signal. In accordance with the preferred embodiment, a conventional back propagation training algorithm is employed. Alternatively, other lO equivalent learning paradigms as known to those skilled in the art may be used. Back propagation requires that each neuron produce an output that is a continuous differentiable nonlinear or semi-linear function of its input. It is preferred that this 15 function, called a transfer function, be a sigmoid logistic non-linear function of the general form:
Yi ~ (1) 1 + e _ [~(Wj-X;) + ~]
Where Yi is the output of neuron i, ~(Wj-~) is the sum of the inputs to neuron i from the previous layer of neurons j, x; is the output of each neuron j in 25 the previous layer to neuron i, w; is the weight associated with each synapse connecting each neuron j in the previous layer to neuron i, and ~ is a bias similar in function to a threshold. The derivative of this function Yi with respect to its total input, 30 NETi ~ ~(wj-xj) + ~] is given by ~ Yi ' Yi-(l-Yi) (2) ~ NETi .
Thus, the requirement that the output is a differentiable function of the input is met. Other transfer functions could be used such as the hyperbolic tangent and the like.
The process of training a neural network to accurately calculate outputs involves adjusting the connection weights of each synapse 27 in a repetitive fashion based on known inputs until an output is produced in response to a particular set of inputs 10 which satisfies the training criteria or tolerance factor as exemplified in Figure 4, step E.
During training, the transfer function Yi remains the same for each neuron but the weights 27 are modified. Thus, the strengths of connectivity 15 are modified as a function of experience. The weights 27 are modified according to ~W~ Wj (3) 20 where ~Wj is the incremental adjustment to the e~isting weight wj, ~i is an error signal available to the neuron, and ~ is a constant of proportionality also called the learning rate.
The determination of the error signal ~i is 25 a recursive process that is propagated backward from the output neurons. First, input values are transmitted to the input neurons 22. This causes computations in accordance with Equation 1 or those of a similar transfer function to be transmitted 30 through the neural network of Figure 2 until an output value is produced. It should be noted from Figure 3 that the transfer function Yi cannot reach the e~treme limits of minus one or plus one without D-16~06 infinitely large weights. ~he calculated output of each output neuron 24 is then compared to the output desired or known to be correct from the training data. For output neurons the error signal is ~i - (Di-Yi) ~-Yi - (4) a NETi where Di is the desired output of the given output neuron. By substituting Equation 2 into Equation 4 using the sigmoid transfer function the error signal 10 for output neurons i can be restated as follows:
~i e (Di-Yi)(Yi)(l-yi) (5) For hidden neurons 26 there is no specific desired output from the measured data, so the error signal is determined recursively in terms of the 15 error signals in the output or successive hidden layer neurons k to which the hidden layer neurons directly connect and the weights of those connections. Thus, for non-output neurons ~i ' Yi(l-Yi)~(~k-Wk) (6) 20 where ~k is the error signal of respective output or successive hidden layer neurons k to which the hidden neuron i is connected and Wk is the weight between that neuron k and the hidden neuron i.
From Equation 3 it can be seen that the 25 learning rate ~ will affect how greatly the weights are changed each time the error signal ~; is - propagated. The larger ~, the larger the changes in the weights and the faster the learning rate. If, however, the learning rate is made too large the 30 system can oscillate during learning. Oscillation can be avoided even with large learning rates by using a momentum term a. Thus, ~Wi,n~ iYi ~ ~ ~Wi,n (7 may be used in place of Equation 3 where ~Wi n+l is the present adjustment of Wi and ~Wi n is the previous adiustment of Wi.
The constant a determines the effect ~f past weight changes ~Wi n on the current direction of movement in weights ~Wi n+l providing a kind of momentum in weights that effectively filters out high frequency oscillation in the weights.
Training is accomplished by first collecting sets of input and output data from many actual decarburization operations to be presented as training data in random order to the neural networks. Data is collected defining the initial 15 contents of the chemical constituents of a molten metal bath, the initial bath temperature and weight, the weights of the solid additions added during the blow period, the ratio of oxygen to diluent gas blown and the final temperature obtained whereas output 20 data includes the counts of oxygen and diluent gas injected into the bath. E~amples of solid additions used during decarburization are the flu~es such as lime, dolomitic lime or magnesia, the base material used as a source of iron units in the case of ferrous 25 metal refining, cobalt units in the case of cobalt base metal refining or nickel units in the case of nickel based metal refining, ferro-chrome, ferro-manganese, nickel and ferro-nickel. The parameter to be used as the inputs and the parameter 30 to be used as the outputs for each of the neural networks will vary based upon the function of the network.
Each of the neural networks 1 to 5 are assigned different functions and are trained to recognize and identify the requirements needed to perform such functions during the decarburization 5 operation. For example, the first neural network 1 is assigned the function of determining the gas, injection requirements, i.e. the counts of o~ygen at a preselected ratio of oxygen to diluent gas to reach a specified bath temperature from the initial 10 chemistry, temperature and weight of the bath 12 charged in the vessel 10. The second neural network 2 may be assigned the function of determining the gas injection requirements to reach a specified carbon content from the initial chemistry, 15 temperature and weight of the bath 12 charged in the vessel 10 using a preestablished gas ratio schedule.
A third neural network may be assigned the function of determining the carbon content in the molten metal bath after the gases have been injected 20 in satisfaction of the computation of either of the first two neural networks. The fourth neural network is assigned the function of computing the bath temperature and the fifth neural network computes the silicon, manganese, chromium, nickel, and molydenum 25 contents of the bath at the completion of the injection of oxygen for the preestablished ratio of o~ygen to diluent gas in accordance with either neural network 1 or 2 based upon the input data of the initial bath chemistry, temperature and weight, 30 the counts of oxygen injected and the ratio of o~ygen to diluent gas used. The input data of initial conditions may represent either the initial conditions when the molten metal is transferred to the refining vessel or the initial conditions esisting at the commencement of any process period i.e, blow period within a decarburization operation 5 as will be explained hereafter in greater detail.
Thus the neural networks 1-2 provide the decarburization oxygen counts required to decarburize the molten metal bath pursuant to the decarburization logic of Figure 5. The computer 18 follows the logic 10 requirements of Figure 5 in performing the decarburization operation in compliance with the computation of the neural networks 1-2 respectively.
For purposes of the sub]ect invention neural network 1 is used to determine the amount of o~ygen 15 required to be injected into the bath to reach a specified aim temperature level and has ten respective input neurons 22 for the initial conditions including the initial carbon, silicon, manganese, chromium, nickel and molybedenum contents 20 of the bath, the initial temperature and weight of the bath, the specified aim temperature of the bath and the ratio of o~ygen to diluent gas to be used.
An additional six input neurons are used for the weights of each of sig types of solid additions which 25 may be added during the blow period as hereinabove identified. Thus neural network 1 is constructed of sixteen input neurons 22, one output neuron 24 for indicating the counts of oxygen required to reach the specified aim temperature level and eight hidden or 30 inner neurons 26 in a single layer.
Neural network 2 is used to determine the amount of oxygen required to reach a specified carbon content, and similarly to network 1, has ten input neurons 22 for the initial carbon, silicon, manganese, chromium, nickel and molydenum constituents of the bath, the initial bath 5 temperature and weight, the desired aim carbon content and the ratio of o~ygen to diluent gas. An additional six input neurons are used for the si~
solid addition types which may be added during the blow period. Thus neural network 2 is constructed of 10 seventeen input neurons 22 and one output neuron 24 for indicating the counts of ogygen required to reach the specified aim carbon content and has eight hidden or inner neurons 26 in a single layer.
Neural network 3 is used to determine the 15 carbon content reached by injecting a specified amount of oxygen at a specified ratio of oxygen to diluent gas into known initial bath conditions and has respective input neurons 22 for the initial carbon, silicon, manganese, chromium, nickel and 20 molybdenum contents of the bath, the initial bath temperature and weight, the specified amounts of oxygen and diluent gases injected, and the ratio of o~ygen to diluent gas blown and the weights of each of the addition types added during the blow period.
25 A network with si~ types of additions is thus constructed of seventeen input neurons. The network has one output neuron for the carbon content resulting from the specified gas injection and has nine hidden neurons in a single layer.
Neural network 4 is used to determine the temperature reached by injecting a specified amount of oxygen at a specified ratio of oxygen to diluent gas into known initial bath conditions and has respective input neurons 22 for the initial carbon, silicon, manganese, chromium, nickel and molybdenum contents of the bath, the bath temperature and 5 weight, the weights of each of the addition types added during the blow period, the specified amounts of o~ygen and diluent gases injected, the elapsed time, and the ratio of oxygen to diluent gas blown.
A network with si~ types of additions is thus 10 constructed of eighteen input neurons. The network has one output neuron for the temperature resulting from the specified gas injection and has nine hidden neurons in a single layer.
Neural network 5 is used to determine the 15 silicon, manganese, chromium, nickel, and molybdenum contents of the bath following the injection of specified amounts of oxygen and diluent gases at a specified ratio of ogygen to diluent gas into known initial bath conditions. Neural network 5 has 20 respective input neurons for the initial carbon, silicon, manganese, chromium, nickel and molybdenum contents of the bath, the bath temperature and weight, the weights of each of the addition types added during the blow period, the specified amounts 25 of o~ygen and diluent gases injected and the ratio of oxygen to diluent gas blown. A network with sig types of additions is thus constructed of seventeen input neurons. The network has five output neurons for the silicon, manganese, chromium, nickel, and 30 molybdenum contents, respectively, resulting form the specified gas injection and has eleven hidden neurons in a single layer.
Although a single layer of hidden neurons is used, it is within the scope of the present invention to use a greater or lesser number of hidden layers of neurons. The e~act configuration is best established 5 empirically. This applies to the number of hidden neurons within a hidden layer and the number of hidden layers chosen for each of the neural networks.
Input and output data from many actual decarburization operations are used to train the 10 neural networks with data separately collected to correspond to multiple process periods in each decarburization operation. Data is collected for each process period in which only one discreet ratio of oxygen to diluent gas is injected at any time in a 15 single process period. A process period is herein defined as the time between two consecutive samples of bath chemistry and temperature for a given decarburization operation, i.e., within a single heat. The time interval between samples may be short 20 or long in a random relationship. Thus the process periods have no defined time relationship or chronology. Pure diluent gas stirring may also be performed or the vessel may be idle during portions of the process period or additions may be added at 25 any time concurrent with any of these events during process periods from which the data is collected for purposes of training the neural networks. The data ,should be collected in such a way that the ranges of useful or expected input and output values are 30 represented. For instance, for AOD refining it is best to have initial carbon contents of from 0.1% to 1.8% in the molten metal as initial conditions for various process periods and have data for process periods using o~ygen to diluent gas ratios from 4 to 1 to ratios of 1 to 3. Pure diluent gas decarburization data would also be needed to 5 accurately model a practice which uses this technique. Preferably, at least 10 process periods of data should be collected at each oxygen to diluent gas ratio, although the accuracy of the neural network is enhanced by greater amounts of data.
An example of a block of input and output training data for the neural networks 1-5 is set forth in the following Table:
TABLE
INITIAL
ELAPSED COUNTSCOUNTS COUNTS INITIAL INITIAL INITIAL INITIAL INITIAL INITIAL INITIAL METAL
RATIO TIME 02 NZ AR TEMP ~Frc %si 7~R 7MN 7Ni 7MoWEIGHT
lbs 0.000 4.0000.000 64.00039.000 2884.00 1.300 0.25019.6800.620 6.3400.26 109333 10 3.000 8.000 209.00081.000 0.0002792.000 1.2400.24019.630 0.6406.370 0.25 109202 3.000 9.000300.000130.0000.000 2942.0001.080 0.09019.4800.600 6.4000.25 109700 1.00015.000344.000370.0000.000 2947.0000.800 0.08017.9201.330 6.9700.26 114794 3.00010.000412.000143.0000.000 2751.0001.200 0.17019.2400.610 6.4600.13 101000 0.000 6.0000.000 67.0000.000 2982.0000.680 0.09018.6600.560 6.5600.13 99808 15 3.00011.000 299.000142.000 0.0002778.000 0.6500.10017.360 1.4206.900 0.13 109985 1.00012.000243.000272.0000.000 2952.0000.450 0.10016.8001.160 6.9900.13 108157 0.000 4.0000.000 57.0000.000 2849.0000.160 0.21018.7700.610 6.9701.56 99667 3.00011.000406.000134.0000.000 2770.0001.120 0.19018.7800.610 6.9701.61 99607 0.000 5.0000.000 74.0000.000 2997.0000.620 0.10018.2500.550 7.0501.58 98491 20 3.00011.000 398.000165.000 0.0002690.000 0.6800.11017.150 1.3708.370 1.55 109798 1.000 8.000147.000173.0000.000 2980.0000.390 0.09016.3901.060 8.4601.57 108623 0.33323.000106.000209.000116.000 3037.0000.200 0.09016.1801.020 8.4901.56 108189 0.000 5.0000.000 68.0000.000 2772.0001.440 0.26018.2700.550 3.8700.19 106100 4.00012.000465.000139.0000.000 2680.0001.390 0.23018.4000.560 3.8500.19 106015 25 0.000 9.000 0.00088.000 0.0002971.000 0.9400.07018.040 0.5103.920 0.20 105093 3.00014.000456.000188.0000.000 2703.0001.030 0.09017.2801.750 7.8600.21 114993 1.000 9.000185.000204.0000.000 2972.0000.550 0.08016.7501.470 7.9600.21 113820 4.000 4.00034,000111.0000.000 2829.0001.550 0.17019.0700.540 6.5900.36 102667 4.00011.000331.000144.0000.000 2769.0001.520 0.13018.8600.540 6.6600.36 102379 30 0.000 5.000 0.00054.000 0.0002844.000 1.3900.18018.730 0.5704.280 0.34 101667 4.00011.000362.000122.0000.000 2752.0001.240 0.17018.7100.580 4.2900.35 101484 3.000 6.000194.00091.0000.000 2943.0000.850 0.17018.4500.540 4.2900.35 100824 3.000 6.000157.00077.0000.000 2860.0000.720 0.08016.9801.560 7.0000.36 109271 1.000 5.00091.000112.0000.000 2947.0000.540 0.08016.8601.560 7.0400.36 108943 35 0.33339.000 356.000759.000 149.000Z977.000 0.4100.08016.690 1.5407.060 0.36 108616 0.000 5.0000.000 55.0000.000 2840.0001.210 0.30018.6500.660 3.5502.10 96333 4.00011.000454.000142.0000.000 2746.0001.200 0.30018.6500.660 3.5502.08 96324 0.00012.0000.000 207.0000.000 3060.0000.690 0.30018.6500.660 3.5502.08 95832 3.00013.000458.000184.0000.000 2546.0000.690 0.10017.5301.390 8.4002.07 111824 40 1.000 9.000 191.000215.000 0.0002942.000 0.5300.07016.550 1.0908.530 2.07 110516 0.000 5.0000.000 72.0000.000 2826.0001.580 0.12019.0200.600 3.6300.39 104500 ~ o ~ O ~ ~) ~ ~ ~ 0 U') "~ ~ N ~ 7 ~ ~ ~ ~ ~ ~ ~1 1~ 0 0 0 0 0 0 0 0 0 -- -- -- -- -- O O O O O O O O O O O O O ~'J t'i N N ~'J O
J O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O
~ o ~ o cr~ o ~ o u ~ ~ ~D O ~D _ O ~ O ~ ~ U7 0 ~ ~' z z f~ ~ o Ln ~ o~ o ~ ~ ~ ~ o ~ I ~1 0 0 0 0 O ~ ~ ~ 1~ 0 0 0 0 ~ ~ 1~1~ 0 ~D ~ ~ ~ r~I~ r~I' ~ r7 0 0 CD ~
J O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O O
z ~ ~ ~ ~ ~ U~ ~ _ O ~ U~ ~ O O O U~ O ~ O O ~D
~ o o - - o - - - o o - - - - o o - - - o o o o - - - - o o - - - o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o J ~ 0 ~ u~ ~ ~o o ~ 0 u7 u~ a~ 0 o o ~ 0 u~ ~ ~ o _ u > 0 ~o ~ o Ir~
Z ~ 0 ~ D 0 0 1~ 0 0 1~. ~ D 0 W 0 0 ~ D 0 0 ~ O 0 o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o ~r ~ ~ ~ 0 ~ a~ o o 0 o~ o ~ c~ 0 u:~ ~ 0 1~ ~ 0 a) 0 1~ o o o z ~ ~ o o o o - - o - - ~ o o o ~ o o o o - o - - o o o o ~ ~ - o o l~ o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o ~ o o o o o o o o o o o o o 0 o o o o o o o o o o ~ o o o o o o ~S ~r 0 0 ~t 0 U~ Ul 1' OJ O~ 0 CJ~ O ~) C7~ ~ 0 U7 Cr~
Z ~ ~ 0 0 ~ ~O ~ D ~ ~ O ~ Cr~ o u~ 0 r~ ~ ~ o ~ ~
-- -- o o o o o o -- o o o o o -- o -- o o -- ~ -- o o o o o . o o O o _ ~, C ", ~ ~ ~ 1~ _ N CC~ N 1~ o ~ O O ~ cr O -- 0 N O 0~ 0 t~l ~) O 1~ r~ a~ ~D O 'D N 1~ --~' Z ~ ~ 1' ~ ~ 1' 0 ~ u~ 0 r~) 0 0 ~ o ~ o ~c> ~ u~ ~ ~o ~ ~ ~ ~ ~D In ~ _ 0 'c ~ O ~ ~ O O ~ I~ C~ O r' o~ o 0 ~ ~ r' cr~ o r' G~ 0 ~ O~ O 1' 0 ~ O O
o ~ 1~ 'J ~ 0 ~ N ~ ) ~J ~ OJ ~ 0 ~J ~ ~ ~ N ~) ~ J ~ N C~l 0 ~'J 0 C~
-O
~ ~ ~ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1' 0 0 0 0 o~ O ~ O O o O O ~ O O o o o O ~ O 0 t~) O~ U') z ~ ~ J~ O O O O O O O O O O ~ O O O O O 1' 0 0 0 0 0 0 0 0 0 0 0 0 U~ O O O
g O~
~r _ z o o o o o o o o o o ~ o o o o o o o o o o o o o o o o o o l-- o o o ~ o ~o - ~ ~ ~
o ~-~ ~ v~ o o o o o r~ o o o o ~ o o o o o 1' o o o o o o 1-- o o o o o ~ o o o O ~ O In cn o - ~ ~
O ~u ~ o O ~ 0 0 ~0 0 o o o " o o o o o 0 0 0 0 0 0 0 ~ 0 0 0 0 0 ~ O O O
o ~J o o ~ o o ~ o o o o ~ o o o o o o o o o o o o l' o o o o o o o o o O ~ .D 0 ~ _ ~
u o u~ o ~n o ~ o -- t~J N 0 0 ~
Each network is trained using the standard back propagation paradigm. Training should use either a hyperbolic tangent, or preferably a sigmoid transfer function, a learning rate of 0.1 and a 5 momentum of zero for each neuron. Once the neural network is sufficiently trained, it is translated to a readily usable programming language such as C or BASIC or FORTRAN. The code in one of these languages is compiled and linked as necessary.
A flowchart indicative of the training operation is shown in Figure 4. Pursuant to Step A
the weights and offset are set to small random values between one and minus one. The collected training input and output data for a given process period are 15 then presented to the neural network input neurons 22 under training as indicated in Step B. After the input data is propagated through the inner layer of neurons 26 to the output neurons 24, an output 20 as shown in Step C is formed for each output neuron 24 20 based on the transfer function Yi described in Equation (1). The calculated output 20 from the output neurons 24 is compared in Step D to the output data of the given process period to develop an error signal 30 using Equations 5 and 6 for the output and 25 hidden neurons respectively. The error signal 30 is then compared to a preset tolerance factor in Step E. If the error signal 30 is larger than the tolerance factor, the error signal 30 as shown in Step F makes a backward pass through the network 30 using Equation 7 for adjusting the weights to the output and hidden neurons and each weight in Step A
is incrementally changed by ~Wi. Input data of another process period is presented and Steps B
through E are repeated until the error signal 30 is reduced to an acceptable level. When the error signal 30 is smaller than the preset tolerance factor S the training procedure pursuant to Step G is complete.
For purposes of verification the verification Steps H and I are followed in which test inputs are presented to generate outputs 20 as in Step C for comparison in Step D with known outputs.
10 The tolerance factor is an e~ternally determined standard for the desired accuracy of the neural network. The training is continued until the error signal is less than this tolerance. The simplest form of a tolerance is to assign a certain percentage 15 error for training to stop. A more practical form of tolerance is to test whether the neural network is in fact learning to generalize the relationships between the problem's inputs and outputs or whether it has begun to memorize those relationships for the 20 specific data with which it trains itself. After a periodic number of iterations the neural network is applied to the reserve or test data and its ability to estimate the desired output for that data is assessed. In the early stage of training the neural 25 network will learn to estimate the test outputs with increasing accuracy. After the neural network has completed generalization, it begins to increase its accuracy relative to the training data at the e~pense of its accuracy relative to the test data. At this 30 point the training is considered to have reached the optimum configuration or weights for general problem solvinq, and the training process is stopped. Each neural network 1-5 is trained in the aforementioned manner.
The determination of the error signal 30 is a recursive process that starts by generating outputs 5 from the output neurons 24 based on feeding the collected data to the input neurons 22. The input neurons 22 cause a signal to be propagated forward through the neural network until an output signal is produced at the output neuron 24. From equation 3 it 10 can be seen that the learning rate ~ will effect how much the weights are changed each time an error signal is propagated. The larger ~, the larger the changes in the weights and the faster the learning rate at the possible expense of the accuracy that may 15 eventually be obtained.
The total population of collected input and output data should be randomly divided into two groups. The larger group should be used as training data for training the neural network with the 20 remaining smaller group of data used as test data for verification. One reasonable division is to use 75%
of the collected data for training purposes and to use the remaining 25% of the collected data as test data to verify the network's predictive accuracy.
25 The neural network should be trained until comparisons to the verification data show that the model's accuracy is not increasing. At this point, those skilled in the art will know that the network is no longer learning to generalize the problem, but 30 is rather memorizing the specific solutions for the training set of data. The learning process typically takes 10,000 to 500,000 presentations of process periods, i.e, presentations of individual sets of complete input and output data for a given process period, to the network for adjustment of its weights. The order of presenting the process periods 5 within the entire training set of data to the neural network for training should be randomly shuffled after each time the entire set has been presented to the network for training.
The sequence of using the trained neural 10 networks 1-5 is determined in accordance with the decarburization logic shown in Figure 4. The composition, weight and temperature of the bath at the time of transfer to the refining vessel is estimated or measured. The calculations of the solid 15 additions are independently calculated and do not form part of the present invention. The decarburization logic shown in Figure 9 is an illustrative example of the invention using neural networks 1-5 based on a predetermined initial 20 decarburization o~ygen to diluent gas setting and a predetermined oxygen to diluent gas decarburization ratio schedule. The e~ample of Figure 4 uses a preselected aim temperature level of 3050~F for a ratio of 4 to 1 oxygen to diluent gas and a ratio 25 schedule of 1, .333 and 0 for the successive aim carbon levels of .15%C, .05%C and .03%C
respectively. The decarburization logic establishes decision trees to determine when to use the neural networks 1-5.
Decarburization proceeds only if the carbon level is above the ultimate aim level of 0.03% C. If the bath temperature is less than 3050~F and calculated solid additions have yet to be added to the bath, a ration of 4 to 1 oxygen to diluent gas is selected and neural network 1 is activated to compute the oxygen counts necessary to raise the temperature 5 of the bath to the preselected level of 3050~F. Upon supplying o~ygen equal to the computed counts calculated by neural network 1 the neural networks 3, 4 and 5 are activated or fired to compute the updated conditions of carbon content, bath temperature and 10 metal chemistry upon completion of said injection.
Neural network 1 is again activated with the aforementioned outputs of neural networks 3, 4 and 5 as the new initial conditions and the required solid additions also used as new inputs to compute the 15 oxygen count necessary to raise the bath temperature to the preselected level of 3050~F while simultaneously adding said additions. O~ygen is injected at the preselected ratio of 4 to 1 while the said additions are added until the computed oxygen 20 counts are satisfied.
If the bath temperature is less than 3050~F
and no solid additions have yet to be added to the bath, a ratio of 4 to 1 o~ygen to diluent gas is selected and neural network 1 is activated to compute 25 the o~ygen counts necessary to raise the temperature of the bath to the preselected level of 3050~F. Upon supplying oxygen equal to the computed counts calculated by neural network 1 the neural networks 3, 4 and 5 are activated to compute updated conditions 30 of carbon content, bath temperature and metal chemistry.
..
If the bath temperature computed by neural networks 3, 4 and 5 equals or exceeds the predetermined aim temperature level of 3050~F a new ratio of oxygen to diluent gas is specified 5 corresponding to a ratio of 1/1, 1/3 or zero, respectively, with the determination based upon the temperature and carbon concentration such that if the temperature is between 3050~F and 3100~F and the carbon concentration exceeds .15% the ratio of 1/1 is 10 specified, whereas if the temperature is equal to or greater than 3050~F and the carbon content is between .08% and .15% a ratio of 1/3 is specified and finally if the temperature exceeds or equals 3050~F and the carbon content is less than .08% a zero ratio is 15 specified. For any of these conditions neural network 2 is activated, the appropriate oxygen to diluent gas ratio is chosen and the required oxygen gas counts are computed to reach the aim carbon level. Oxygen and/or diluent gas is then blown at 20 the specified ratio until the oxygen counts as computed by neural network 2 are satisfied. The neural networks 3, 4 and 5 are then activated after each successive step to update the bath chemistry, temperature and carbon content for the initial 25 condition of any subsequent decarburization.
An AOD process was run using a conventional thermodynamic model for predicting and controlling the decarburization process during the production of both ASTM 300 series and ASTM 400 series stainless 30 steels. Upon adjusting the constants in the model to attain optimal accuracy, the carbon content could be predicted with a standard deviation of 0.11% carbon for actual carbon contents between 0.1% and 0.3%.
Fourteen heats of stainless steels were sampled after the use of each ratio of oxygen to diluent gas to measure the bath chemistry and temperature. The 5 information was used for training the first neural network of the present invention. The trained neural network was then used to predict the carbon content at carbon contents between 0.1% and 0.3% carbon during the production of the same grades of stainless 10 steels. The carbon content prediction using the said neural network had a standard deviation of only 0.035% carbon.
Claims (14)
1. A method for refining steel by controlling the decarburization of a predetermined molten metal bath having a known composition of elements including carbon and having a known or estimated initial temperature and weight at the outset of decarburization of a molten metal bath in a refractory vessel with a process of decarburization performed through the injection of oxygen and a diluting gas into said bath under adjustable conditions of gas flow, comprising the steps of:
(a) training a first neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, weight and temperature at the outset of each process period, the gas ratio of oxygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period, and the final temperature obtained at the conclusion of each process period, until said first neural network is able to provide a substantially accurate output representing the counts of oxygen required to be injected into said predetermined bath at any preselected gas ratio to cause the temperature of the bath to rise to a specified aim temperature level as a result of such gas injection;
(b) training a second neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, weight and temperature at the outset of each process period, the gas ratio of oxygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period and the final carbon content obtained at the conclusion of each process period until said second neural network is able to provide a substantially accurate output schedule of oxygen counts to be injected into said predetermined bath to reduce the carbon level to a predetermined aim level in one or more successive stages corresponding to a preselected schedule of ratios of oxygen to diluent gas;
(c) employing said first neural network to compute the oxygen counts to be injected into said predetermined bath, from its known initial chemistry, weight and temperature at a first preselected ratio of oxygen to diluent gas to raise the bath temperature to a specified aim temperature level.
(d) injecting oxygen and diluent gas into said bath at said first preselected ratio until the oxygen counts computed by said first neural network are satisfied;
(e) employing said second neural network to provide an output schedule of oxygen counts to be injected into said predetermined bath from its known initial chemistry, weight and temperature to successively reduce the carbon level in said bath to a predetermined aim carbon level in one or more stages corresponding to a preselected schedule of ratios of oxygen to diluent gas;
(f) injecting oxygen and diluent gas into said bath at said preselected schedule of oxygen counts corresponding to said output schedule as computed by said second neural network;
(g) training a third neural network to analyze data from the bath chemistry, weight and temperature at the outset of each process period, the weight of each solid addition, if any, made during each process period, the counts of oxygen injected during each process period, the corresponding ratio of oxygen to diluent gas used during each process period and the resulting carbon content at the conclusion of each process period for the purpose of predicting an output representing the carbon content that would be obtained as a result of such oxygen injection; and (h) employing said third neural network to compute the carbon content in the bath upon completion of the injection of oxygen intended as a result of computations performed in at least one of the steps (c) and (e).
(a) training a first neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, weight and temperature at the outset of each process period, the gas ratio of oxygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period, and the final temperature obtained at the conclusion of each process period, until said first neural network is able to provide a substantially accurate output representing the counts of oxygen required to be injected into said predetermined bath at any preselected gas ratio to cause the temperature of the bath to rise to a specified aim temperature level as a result of such gas injection;
(b) training a second neural network to analyze input and output data representative of many process periods of one or more decarburization operations, from data including the bath chemistry, weight and temperature at the outset of each process period, the gas ratio of oxygen to diluent gas used during each process period, the counts of oxygen injected into the bath for each process period and the final carbon content obtained at the conclusion of each process period until said second neural network is able to provide a substantially accurate output schedule of oxygen counts to be injected into said predetermined bath to reduce the carbon level to a predetermined aim level in one or more successive stages corresponding to a preselected schedule of ratios of oxygen to diluent gas;
(c) employing said first neural network to compute the oxygen counts to be injected into said predetermined bath, from its known initial chemistry, weight and temperature at a first preselected ratio of oxygen to diluent gas to raise the bath temperature to a specified aim temperature level.
(d) injecting oxygen and diluent gas into said bath at said first preselected ratio until the oxygen counts computed by said first neural network are satisfied;
(e) employing said second neural network to provide an output schedule of oxygen counts to be injected into said predetermined bath from its known initial chemistry, weight and temperature to successively reduce the carbon level in said bath to a predetermined aim carbon level in one or more stages corresponding to a preselected schedule of ratios of oxygen to diluent gas;
(f) injecting oxygen and diluent gas into said bath at said preselected schedule of oxygen counts corresponding to said output schedule as computed by said second neural network;
(g) training a third neural network to analyze data from the bath chemistry, weight and temperature at the outset of each process period, the weight of each solid addition, if any, made during each process period, the counts of oxygen injected during each process period, the corresponding ratio of oxygen to diluent gas used during each process period and the resulting carbon content at the conclusion of each process period for the purpose of predicting an output representing the carbon content that would be obtained as a result of such oxygen injection; and (h) employing said third neural network to compute the carbon content in the bath upon completion of the injection of oxygen intended as a result of computations performed in at least one of the steps (c) and (e).
2. A method as defined in claim 1 wherein said known composition of elements is selected from the class consisting essentially of carbon, iron, silicon, chromium, manganese, nickel and molybdenum.
3. A method as defined in claim 2 wherein said oxygen and diluent gas are injected into said bath subsurfacely.
4. A method as defined in claim 3 wherein said diluent gas is selected from the group consisting of argon, nitrogen and carbon dioxide.
5. A method as defined in claim 4 wherein said first neural network is trained and used in step (c) prior to the use of said second neural network in step (e).
6. A method as defined in claim 4 wherein at least 10 process periods of data are collected for each oxygen to diluent gas ratio.
7. A method as defined in claim 6 further comprising adding solid additions to said bath during decarburization.
8. A method as defined in claim 7 wherein said solid additions are selected from the group consisting of lime, dolomitic lime, magnesia, ferro-chrome, ferromanganese, nickel and ferro-nickel.
9. A method as defined in claim 7 wherein said data applied to train said first and second neural networks further comprises the weights of any solid additions added during each of said process periods for use in training said neural networks based on actual conditions of operation using solid additions.
10. A method as defined in claim 9 wherein said first, second, and/or third neural networks have a multiple number of input neurons to receive said input data, one layer of output neurons and at least one layer of hidden neurons with each neuron in each layer interconnected to each neuron in an adjacent layer through adjustable weights.
11. A method as defined in claim 10 wherein each neural network is trained by comparing the output generated from its output neurons to the output data for a corresponding process period or set of process periods; generating an error signal from such comparison, comparing said error signal to a predetermined tolerance factor and modifying the weights between neuron layers until said error signal is equal to or below said tolerance factor.
12. A method as defined in claim 11 wherein the output of the neural network under training is tested against test data to verify the accuracy of the neural network output.
13. A method as defined in claim 1 further comprising the steps of:
training a fourth neural network to analyze data from the bath chemistry, weight and temperature at the outset of each process period, the weight of each solid addition, if any, made during each process period, the counts of oxygen injected during each process period, the corresponding ratio of oxygen to diluent gas used during each process period, and the resulting temperature at the conclusion of each process period for the purpose of providing an output representing the temperature reached as a result of such oxygen injection; and employing said fourth neural network to compute the temperature of the bath upon completion of the injection of oxygen.
training a fourth neural network to analyze data from the bath chemistry, weight and temperature at the outset of each process period, the weight of each solid addition, if any, made during each process period, the counts of oxygen injected during each process period, the corresponding ratio of oxygen to diluent gas used during each process period, and the resulting temperature at the conclusion of each process period for the purpose of providing an output representing the temperature reached as a result of such oxygen injection; and employing said fourth neural network to compute the temperature of the bath upon completion of the injection of oxygen.
14. A method as defined in claim 14 further comprising the steps of:
training a fifth neural network to analyze data from the bath chemistry, weight and temperature at the outset of each process period, the weight of each solid addition, if any, made during each process period, the counts of oxygen injected during each process period, the corresponding ratio of oxygen to diluent gas used during each process period and the resulting chemistry at the conclusion of each process period for the purpose of providing an output representing the chemistry content of the bath as a result of such oxygen injection; and employing said fifth neural network to compute the chemistry content of the bath upon completion of the injection of oxygen.
training a fifth neural network to analyze data from the bath chemistry, weight and temperature at the outset of each process period, the weight of each solid addition, if any, made during each process period, the counts of oxygen injected during each process period, the corresponding ratio of oxygen to diluent gas used during each process period and the resulting chemistry at the conclusion of each process period for the purpose of providing an output representing the chemistry content of the bath as a result of such oxygen injection; and employing said fifth neural network to compute the chemistry content of the bath upon completion of the injection of oxygen.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US07/802,046 | 1991-12-03 | ||
| US07/802,046 US5327357A (en) | 1991-12-03 | 1991-12-03 | Method of decarburizing molten metal in the refining of steel using neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CA2084396A1 CA2084396A1 (en) | 1993-06-04 |
| CA2084396C true CA2084396C (en) | 1998-07-28 |
Family
ID=25182697
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA002084396A Expired - Fee Related CA2084396C (en) | 1991-12-03 | 1992-12-02 | Method of decarburizing molten metal in the refining of steel using neural networks |
Country Status (10)
| Country | Link |
|---|---|
| US (1) | US5327357A (en) |
| EP (1) | EP0545379B1 (en) |
| KR (1) | KR0148273B1 (en) |
| CN (1) | CN1037455C (en) |
| BR (1) | BR9204824A (en) |
| CA (1) | CA2084396C (en) |
| DE (1) | DE69209622T2 (en) |
| ES (1) | ES2085539T3 (en) |
| MX (1) | MX9206989A (en) |
| ZA (1) | ZA929352B (en) |
Families Citing this family (73)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE19547010C2 (en) * | 1994-12-19 | 2001-05-31 | Siemens Ag | Method and device for monitoring the process sequence during beam generation according to the oxygen inflation method |
| US5746511A (en) * | 1996-01-03 | 1998-05-05 | Rosemount Inc. | Temperature transmitter with on-line calibration using johnson noise |
| US6539267B1 (en) | 1996-03-28 | 2003-03-25 | Rosemount Inc. | Device in a process system for determining statistical parameter |
| US7949495B2 (en) | 1996-03-28 | 2011-05-24 | Rosemount, Inc. | Process variable transmitter with diagnostics |
| US7085610B2 (en) | 1996-03-28 | 2006-08-01 | Fisher-Rosemount Systems, Inc. | Root cause diagnostics |
| US6017143A (en) * | 1996-03-28 | 2000-01-25 | Rosemount Inc. | Device in a process system for detecting events |
| US6907383B2 (en) | 1996-03-28 | 2005-06-14 | Rosemount Inc. | Flow diagnostic system |
| US6654697B1 (en) | 1996-03-28 | 2003-11-25 | Rosemount Inc. | Flow measurement with diagnostics |
| US7254518B2 (en) * | 1996-03-28 | 2007-08-07 | Rosemount Inc. | Pressure transmitter with diagnostics |
| US8290721B2 (en) | 1996-03-28 | 2012-10-16 | Rosemount Inc. | Flow measurement diagnostics |
| US6754601B1 (en) | 1996-11-07 | 2004-06-22 | Rosemount Inc. | Diagnostics for resistive elements of process devices |
| US5828567A (en) * | 1996-11-07 | 1998-10-27 | Rosemount Inc. | Diagnostics for resistance based transmitter |
| US6434504B1 (en) | 1996-11-07 | 2002-08-13 | Rosemount Inc. | Resistance based process control device diagnostics |
| US6449574B1 (en) | 1996-11-07 | 2002-09-10 | Micro Motion, Inc. | Resistance based process control device diagnostics |
| US6519546B1 (en) | 1996-11-07 | 2003-02-11 | Rosemount Inc. | Auto correcting temperature transmitter with resistance based sensor |
| US6601005B1 (en) | 1996-11-07 | 2003-07-29 | Rosemount Inc. | Process device diagnostics using process variable sensor signal |
| US5956663A (en) * | 1996-11-07 | 1999-09-21 | Rosemount, Inc. | Signal processing technique which separates signal components in a sensor for sensor diagnostics |
| EP0948759B1 (en) * | 1996-12-31 | 2002-08-07 | Rosemount Inc. | Device in a process system for validating a control signal from a field device |
| DE69818494T2 (en) | 1997-10-13 | 2004-07-01 | Rosemount Inc., Eden Prairie | Transmission method for field devices in industrial processes |
| DE19748310C1 (en) * | 1997-10-31 | 1998-12-17 | Siemens Ag | Controlling formation of foam slag in an electric arc furnace |
| US6615149B1 (en) | 1998-12-10 | 2003-09-02 | Rosemount Inc. | Spectral diagnostics in a magnetic flow meter |
| US6611775B1 (en) | 1998-12-10 | 2003-08-26 | Rosemount Inc. | Electrode leakage diagnostics in a magnetic flow meter |
| US6298454B1 (en) | 1999-02-22 | 2001-10-02 | Fisher-Rosemount Systems, Inc. | Diagnostics in a process control system |
| US8044793B2 (en) | 2001-03-01 | 2011-10-25 | Fisher-Rosemount Systems, Inc. | Integrated device alerts in a process control system |
| US6633782B1 (en) | 1999-02-22 | 2003-10-14 | Fisher-Rosemount Systems, Inc. | Diagnostic expert in a process control system |
| WO2000068654A1 (en) | 1999-05-11 | 2000-11-16 | Georgia Tech Research Corporation | Laser doppler vibrometer for remote assessment of structural components |
| US6356191B1 (en) | 1999-06-17 | 2002-03-12 | Rosemount Inc. | Error compensation for a process fluid temperature transmitter |
| US7010459B2 (en) | 1999-06-25 | 2006-03-07 | Rosemount Inc. | Process device diagnostics using process variable sensor signal |
| EP1247268B2 (en) | 1999-07-01 | 2009-08-05 | Rosemount Inc. | Low power two-wire self validating temperature transmitter |
| US6505517B1 (en) | 1999-07-23 | 2003-01-14 | Rosemount Inc. | High accuracy signal processing for magnetic flowmeter |
| US6701274B1 (en) | 1999-08-27 | 2004-03-02 | Rosemount Inc. | Prediction of error magnitude in a pressure transmitter |
| US6556145B1 (en) | 1999-09-24 | 2003-04-29 | Rosemount Inc. | Two-wire fluid temperature transmitter with thermocouple diagnostics |
| US6442536B1 (en) | 2000-01-18 | 2002-08-27 | Praxair Technology, Inc. | Method for predicting flammability limits of complex mixtures |
| AU2001285629A1 (en) * | 2000-08-11 | 2002-02-25 | Dofasco Inc. | Desulphurization reagent control method and system |
| US6735484B1 (en) | 2000-09-20 | 2004-05-11 | Fargo Electronics, Inc. | Printer with a process diagnostics system for detecting events |
| US8073967B2 (en) | 2002-04-15 | 2011-12-06 | Fisher-Rosemount Systems, Inc. | Web services-based communications for use with process control systems |
| US7720727B2 (en) | 2001-03-01 | 2010-05-18 | Fisher-Rosemount Systems, Inc. | Economic calculations in process control system |
| US6970003B2 (en) | 2001-03-05 | 2005-11-29 | Rosemount Inc. | Electronics board life prediction of microprocessor-based transmitters |
| US6629059B2 (en) | 2001-05-14 | 2003-09-30 | Fisher-Rosemount Systems, Inc. | Hand held diagnostic and communication device with automatic bus detection |
| TW577929B (en) * | 2001-07-02 | 2004-03-01 | Nippon Steel Corp | Decarburization refining method of chromium containing molten steel |
| US6772036B2 (en) | 2001-08-30 | 2004-08-03 | Fisher-Rosemount Systems, Inc. | Control system using process model |
| AT411068B (en) * | 2001-11-13 | 2003-09-25 | Voest Alpine Ind Anlagen | METHOD FOR PRODUCING A METAL MELT IN A LODGE TECHNICAL PLANT |
| CA2480363A1 (en) | 2002-03-27 | 2003-10-09 | Praxair Technology, Inc. | Luminescence sensing system for welding |
| FR2838508B1 (en) * | 2002-04-15 | 2004-11-26 | Air Liquide | PROCESS FOR PRODUCING LIQUID METAL IN AN ELECTRIC OVEN |
| US9201420B2 (en) | 2005-04-08 | 2015-12-01 | Rosemount, Inc. | Method and apparatus for performing a function in a process plant using monitoring data with criticality evaluation data |
| US8005647B2 (en) | 2005-04-08 | 2011-08-23 | Rosemount, Inc. | Method and apparatus for monitoring and performing corrective measures in a process plant using monitoring data with corrective measures data |
| US8112565B2 (en) | 2005-06-08 | 2012-02-07 | Fisher-Rosemount Systems, Inc. | Multi-protocol field device interface with automatic bus detection |
| US20070068225A1 (en) | 2005-09-29 | 2007-03-29 | Brown Gregory C | Leak detector for process valve |
| US7953501B2 (en) | 2006-09-25 | 2011-05-31 | Fisher-Rosemount Systems, Inc. | Industrial process control loop monitor |
| US8788070B2 (en) | 2006-09-26 | 2014-07-22 | Rosemount Inc. | Automatic field device service adviser |
| WO2008042290A2 (en) | 2006-09-29 | 2008-04-10 | Rosemount Inc. | Magnetic flowmeter with verification |
| US8898036B2 (en) | 2007-08-06 | 2014-11-25 | Rosemount Inc. | Process variable transmitter with acceleration sensor |
| US8301676B2 (en) | 2007-08-23 | 2012-10-30 | Fisher-Rosemount Systems, Inc. | Field device with capability of calculating digital filter coefficients |
| US7702401B2 (en) | 2007-09-05 | 2010-04-20 | Fisher-Rosemount Systems, Inc. | System for preserving and displaying process control data associated with an abnormal situation |
| US8055479B2 (en) | 2007-10-10 | 2011-11-08 | Fisher-Rosemount Systems, Inc. | Simplified algorithm for abnormal situation prevention in load following applications including plugged line diagnostics in a dynamic process |
| US7921734B2 (en) | 2009-05-12 | 2011-04-12 | Rosemount Inc. | System to detect poor process ground connections |
| CN102033978B (en) * | 2010-09-19 | 2012-07-25 | 首钢总公司 | Method for forecasting and producing narrow hardenability strip steel by hardenability |
| US9207670B2 (en) | 2011-03-21 | 2015-12-08 | Rosemount Inc. | Degrading sensor detection implemented within a transmitter |
| US9927788B2 (en) | 2011-05-19 | 2018-03-27 | Fisher-Rosemount Systems, Inc. | Software lockout coordination between a process control system and an asset management system |
| CN103031398B (en) * | 2011-09-30 | 2014-04-02 | 鞍钢股份有限公司 | Converter smelting end point carbon content forecasting device and forecasting method |
| CN102690923B (en) * | 2012-06-13 | 2013-11-06 | 鞍钢股份有限公司 | Method for forecasting sublance process carbon content of converter |
| US9052240B2 (en) | 2012-06-29 | 2015-06-09 | Rosemount Inc. | Industrial process temperature transmitter with sensor stress diagnostics |
| US9207129B2 (en) | 2012-09-27 | 2015-12-08 | Rosemount Inc. | Process variable transmitter with EMF detection and correction |
| US9602122B2 (en) | 2012-09-28 | 2017-03-21 | Rosemount Inc. | Process variable measurement noise diagnostic |
| CN106339020B (en) * | 2015-07-16 | 2018-06-05 | 广东兴发铝业有限公司 | Aluminium shape surface oxidation automatic control system based on neutral net |
| US11200489B2 (en) | 2018-01-30 | 2021-12-14 | Imubit Israel Ltd. | Controller training based on historical data |
| CN112912884B (en) * | 2018-10-30 | 2023-11-21 | 株式会社力森诺科 | Material designing apparatus, material designing method, and material designing program |
| BR112022014906A2 (en) * | 2020-02-06 | 2022-09-27 | Jfe Steel Corp | DECARBURATION END POINT DETERMINATION METHOD, DECARBURATION END POINT DETERMINATION DEVICE, SECONDARY REFINING OPERATION METHOD FOR STEEL MANUFACTURING AND METHOD FOR PRODUCING CAST STEEL |
| CN111353656B (en) * | 2020-03-23 | 2021-05-07 | 大连理工大学 | Steel enterprise oxygen load prediction method based on production plan |
| CN111985682B (en) * | 2020-07-13 | 2024-03-22 | 中石化宁波工程有限公司 | Furnace temperature prediction method of coal water slurry gasification furnace based on neural network |
| CN113061683B (en) * | 2021-03-16 | 2022-04-26 | 马鞍山钢铁股份有限公司 | Automatic matching method for converter end point oxygen and converter end point reblowing times quality factor |
| CN113343576B (en) * | 2021-06-22 | 2022-03-11 | 燕山大学 | Prediction method of calcium yield in calcium treatment process based on deep neural network |
| CN114611844B (en) * | 2022-05-11 | 2022-08-05 | 北京科技大学 | Method and system for determining alloy addition amount in converter tapping process |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US3816720A (en) * | 1971-11-01 | 1974-06-11 | Union Carbide Corp | Process for the decarburization of molten metal |
| US3754894A (en) * | 1972-04-20 | 1973-08-28 | Joslyn Mfg & Supply Co | Nitrogen control in argon oxygen refining of molten metal |
| JPH0232679A (en) * | 1988-07-22 | 1990-02-02 | Hitachi Ltd | Method and device for data communication by neural net |
| US5003490A (en) * | 1988-10-07 | 1991-03-26 | Hughes Aircraft Company | Neural network signal processor |
-
1991
- 1991-12-03 US US07/802,046 patent/US5327357A/en not_active Expired - Lifetime
-
1992
- 1992-12-02 ZA ZA929352A patent/ZA929352B/en unknown
- 1992-12-02 CN CN92115190A patent/CN1037455C/en not_active Expired - Fee Related
- 1992-12-02 ES ES92120555T patent/ES2085539T3/en not_active Expired - Lifetime
- 1992-12-02 EP EP92120555A patent/EP0545379B1/en not_active Expired - Lifetime
- 1992-12-02 CA CA002084396A patent/CA2084396C/en not_active Expired - Fee Related
- 1992-12-02 DE DE69209622T patent/DE69209622T2/en not_active Expired - Fee Related
- 1992-12-03 BR BR9204824A patent/BR9204824A/en not_active IP Right Cessation
- 1992-12-03 MX MX9206989A patent/MX9206989A/en not_active IP Right Cessation
- 1992-12-03 KR KR1019920023161A patent/KR0148273B1/en not_active IP Right Cessation
Also Published As
| Publication number | Publication date |
|---|---|
| US5327357A (en) | 1994-07-05 |
| BR9204824A (en) | 1993-06-08 |
| KR930013177A (en) | 1993-07-21 |
| DE69209622D1 (en) | 1996-05-09 |
| CN1074244A (en) | 1993-07-14 |
| CN1037455C (en) | 1998-02-18 |
| ES2085539T3 (en) | 1996-06-01 |
| DE69209622T2 (en) | 1996-10-02 |
| EP0545379A1 (en) | 1993-06-09 |
| KR0148273B1 (en) | 1998-11-02 |
| CA2084396A1 (en) | 1993-06-04 |
| MX9206989A (en) | 1994-05-31 |
| ZA929352B (en) | 1993-06-04 |
| EP0545379B1 (en) | 1996-04-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CA2084396C (en) | Method of decarburizing molten metal in the refining of steel using neural networks | |
| AU645699B2 (en) | Method of estimating material of steel product | |
| CN106096788B (en) | Converter steelmaking process cost control method and system based on PSO _ ELM neural network | |
| US3614682A (en) | Digital computer control of polymerization process | |
| ATE295557T1 (en) | METHOD FOR CONTROLLING AND/OR REGULATING A TECHNICAL PROCESS | |
| Çöl et al. | An artificial neural network model for toughness properties in microalloyed steel in consideration of industrial production conditions | |
| AT411068B (en) | METHOD FOR PRODUCING A METAL MELT IN A LODGE TECHNICAL PLANT | |
| CN102766730A (en) | Method for on-line prediction of total decarbonization amount of molten steel in circulating vacuum degassing method | |
| JP2000144229A (en) | Method for predicting slopping in converter and device therefor | |
| JPH06264129A (en) | Method for controlling end point of steelmaking in converter | |
| CN100507018C (en) | Method for determining oxygen blowing amount and cold material feeding amount in RH refining process | |
| JPH05195035A (en) | Device for controlling blowing converter | |
| JPH0665623A (en) | Method for estimating carbon content in molten steel during blowing in converter | |
| JPH08269518A (en) | Method for guiding to charging condition of treating agent in pretreatment operation of molten iron | |
| JPH0673428A (en) | Method for estimating carbon concentration in steel tapped from converter | |
| KR970010980B1 (en) | Temperature and components in furnace bath using neural network | |
| JPH0641625A (en) | Method for estimating concentration of phosphorus tapped from converter | |
| JP2985643B2 (en) | Method of estimating carbon concentration in molten steel using RH type vacuum chamber | |
| Terpák et al. | The Mathematical Model for Indirect Measurement of Carbon Concentration in the Steelmaking Process and its Utilization in Process Control | |
| CN113987761A (en) | Method and device for predicting temperature of molten pool in converter steelmaking blowing process | |
| Irving et al. | Optimal control of the argon-oxygen decarburising steelmaking process | |
| Schlautmann et al. | Model-based dynamic monitoring and end-point control of converter process | |
| CN116434856A (en) | Converter oxygen supply prediction method based on sectional oxygen decarburization efficiency | |
| JPH07268433A (en) | Method for controlling end point in steel making process of converter | |
| Koštial et al. | Blast furnace knowledge-based control by means of simulation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| EEER | Examination request | ||
| MKLA | Lapsed |