CN103514458A - Sensor fault distinguishing method based on combination of error correction codes and support vector machine - Google Patents

Abstract

The invention provides a sensor fault distinguishing method based on the combination of error correction codes and a support vector machine. The method comprises the steps that first, the error correction codes are generated, namely codes with an error correction function are used for coding types, an SVM is used as a classifier; second, initial features are extracted, namely the time domain and frequency domain information of sensor output signals is extracted, six time domain feature parameters and 3 frequency domain feature parameters are selected as primitive features; third, the features are extracted, namely lines of the error correction codes obtained in the first step are used for forming an SVM two-type classifier, samples with the initial features are input to each SVM for training, a decision function of each classifier is obtained, Sigmoid conversion is carried out on each decision function, and new features in a conversion space are obtained; fourth, fault mode classifying is achieved.

Description

The sensor fault discrimination method combining with support vector machine based on Error Correction of Coding

Technical field

The present invention relates to a kind of fault identification method of sensor, particularly a kind of sensor fault discrimination method combining with support vector machine based on Error Correction of Coding, belongs to Intelligent Information Processing field.

Background technology

Sensor is the sensitive element in surveying instrument, smart instrumentation and computerized information input media, is widely used in various control system, and as the window of understanding systematic procedure state, the accuracy of its measurement result directly affects the operation of system.The working environment of sensor is conventionally more severe, and they in use often can break down because of various reasons.When sensor breaks down, its output signal main manifestations is following several form: deviation, drift, impact, PERIODIC INTERFERENCE, short circuit, open circuit.After fault generation being detected, need to carry out certain online or off-line Fault Compensation for different sensor fault types, therefore, sensor fault is carried out to identification and just seem particularly important.

Sensor is carried out to fault identification and belong to pattern recognition problem.Wherein two the most key steps are feature extraction and pattern classification.The selection of feature is the basis of pattern classification with extracting, and directly affects the accuracy of classification results.Between the fault that sensor occurs and failure cause, there is nonlinear relationship, complicacy, randomness and ambiguity make to be difficult to represent by accurate mathematical model, extract this rule, need to adopt certain nonlinear transformation to make feature there is higher separability.

In recent years, neural network and support vector machine (SVM, Support Vector Machine), as the representative of Nonlinear Classifier, are widely used in pattern classification.Method based on neural network mostly depends on the statistical property in the abundant situation of sample, and sensor fault identification is a kind of typical small-sample learning problem, support vector machine, aspect solution small sample Data classification problem, has the features such as global optimum, simple in structure, Generalization Ability is strong.Compare with neural network, SVM has avoided locally optimal solution problem, has effectively overcome the problem of " dimension disaster ".

At present, common support vector machine Multiclass Classification mainly contains: support vector machine multi-class classification method, one-to-many support vector machine multi-class classification method, binary-tree support vector machine multi-class classification method and Error Correction of Coding support vector machine classification method one to one.Support vector machine classification algorithm training speed one to one, but when classification number increases, classification speed can be slack-off.One-to-many support vector machine sorting algorithm principle is simple, but classification accuracy is not high, and all will use all training samples while training at every turn, and training speed can decline.Binary-tree support vector machine Multiclass Classification classification speed is fast, but training speed is slower, has the phenomenon of accumulation in wrong minute.The generalization of Error Correction of Coding support vector machine classification method is better, and classification speed is fast.

If can follow the reportedly particular problem of sensor fault identification to build qualified Error Correction of Coding, Error Correction of Coding support vector machine classification method, for this problem, can be carried out to identification to fault fast and accurately.

Summary of the invention

In order to improve accuracy and the real-time of sensor fault identification, the invention provides a kind of sensor fault discrimination method combining with support vector machine based on Error Correction of Coding, in the method for the signal of 7 kinds of different modes of sensor, choose respectively time domain and frequency domain character vector as initial characteristics, thereby different fault modes is distinguished.

The sensor fault discrimination method combining with support vector machine based on Error Correction of Coding, comprises the following steps:

Step 1, generation Error Correction of Coding: adopt the coding with error correcting capability to encode to classification, using SVM as sorter, according to preset rules, by Hadamard matrix, obtain qualified Error Correction of Coding;

Step 2, initial characteristics extract: extract time domain and the frequency domain information of sensor output signal, choose 6 temporal signatures parameters and 3 frequency domain character parameters as primitive character;

Step 3, feature extraction; A SVM binary classifier of each row structure of the Error Correction of Coding being obtained by step 1, the sample with initial characteristics is inputted in each SVM and trained respectively, obtain the decision function of each sorter, decision function is carried out to Sigmoid conversion, obtain feature new in transformation space;

Step 4, fault mode classification: according to a SVM binary classifier of each row structure of step 3 encoder matrix, the training sample with new feature parameter is input in each SVM and is trained, test sample book is input in the SVM training, and each sorter is differentiated sample respectively; The output of sorter forms a binary sequence λ={ λ ₁, λ ₂..., λ _n, calculate the Hamming distance between this sequence and classification code word, the class of the code word representative that minor increment is corresponding is final differentiation result.

Uncorrelated between the row of encoder matrix in coding described in step 1, uncorrelated and not complementary between row.

Beneficial effect of the present invention: the present invention can distinguish sensor normal mode and deviation, drift, impact, PERIODIC INTERFERENCE, short circuit, 6 kinds of typical fault modes of open circuit; Real-time and classification accuracy rate are all guaranteed, and generalization is better, particularly under small sample input condition, embody larger advantage.

Accompanying drawing explanation

Fig. 1 is sensor fault discrimination method principle schematic;

Fig. 2 is sensor fault discrimination method process flow diagram.

Embodiment

With reference to accompanying drawing 1, the invention will be further described, and specific implementation step of the present invention is as follows:

The first step: Signal Pretreatment

Gather each the 50 groups of X of sensor output data under 7 kinds of states _ij(i=1,2 ..., 7j=1,2 ..., 50), for making the signal characteristic of extraction not be subject to the impact of amplitude, first signal is carried out to standardization:

${\stackrel{\‾}{X}}_{\mathrm{ij}}=\frac{X_{\mathrm{ij}}-E\left(X_{\mathrm{ij}}\right)}{D_{{\mathrm{\σ}}_{\mathrm{ij}}}}$

Wherein: X _ijthe sensor output signal that represents different mode, E (X _ij) be X _ijaverage,

for X _ijstandard deviation.

Second step: initial characteristics extracts

Extract the peak index, root-mean-square value, kurtosis index, skewness index, waveform index, nargin index, gravity frequency of preprocessed signal, all square frequency, frequency variance be as primitive character.The initial characteristics parameter that obtains every group of signal is Z _i={ z _i1, z _i2..., z _i9(i=1,2 ..., 350).

The 3rd step: generate Error Correction of Coding

Error Correction of Coding is obtained by Hadamard matrix.

(1) generate Hadamard matrix.If 2 ^j-1< k≤2 ^j, generating exponent number is 2 ^jhadamard matrix.

High-order Hadamard matrix is obtained by low order Hadamard matrix recursion:

$H_N=\left[\begin{array}{cc}{H}_{N/2}& H_{N/2}\\ H_{N/2}& -H_{N/2}\end{array}\right]$

Wherein, N=2 ^j,-H _n/2expression is to H _n/2in element get benefit.2 rank matrixes are $H_2=\left[\begin{array}{cc}0& 0\\ 0& 1\end{array}\right].$

(2) delete complete zero row of first row, obtain 2 ^j* (2 ^j-1) matrix;

(3) it is capable that matrix step (2) being obtained is got its front k, obtains required k * (2 ^j-1) Error Correction of Coding.

To sensor normal and fault totally 7 kinds of patterns classify, get k=7.According to the generation step of above-mentioned Error Correction of Coding, obtain corresponding encoder matrix and be:

$H_7=\left[\begin{array}{ccccccc}0& 0& 0& 0& 0& 0& 0\\ 1& 0& 1& 0& 1& 0& 1\\ 0& 1& 1& 0& 0& 1& 1\\ 1& 1& 0& 0& 1& 1& 0\\ 0& 0& 0& 1& 1& 1& 1\\ 1& 0& 1& 1& 0& 1& 0\\ 0& 1& 1& 1& 1& 0& 0\end{array}\right]$

The matrix now obtaining has been eliminated first row entirely for zero-sum is to the conditional shortcoming of classification number, has the separated feature separated with row of row simultaneously, can meet the requirement of Error Correction of Coding in multicategory classification problem.

The 4th step: feature extraction

Sensor signal feature and failure symptom have nonlinear relationship, and the method for employing based on Error Correction of Coding and support vector machine carried out nonlinear transformation to initial characteristics and strengthened signal characteristic.

By second step, obtain sample χ={ (x ₁, c ₁), (x ₂, c ₂) ..., (x _n, c _n), wherein, n=350 is sample size,

for initial characteristics vector, c _i∈ 1,2 ..., 7} is sample class label.

According to matrix H ₇a SVM binary classifier of each row structure.Wherein the composition of j SVM training sample is H ₇in j row value be 0 all samples are classified as the 1st class, all samples that are 1 value are classified as the 2nd class.

Sample is inputted respectively in 7 SVM and trained, obtain parameter alpha _iand b, decision function is:

$f\left(x_j\right)=\underset{i=1}{\overset{N_{\mathrm{sv}}}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{c}_{i}k(x_i,x_j)+b,j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n$

Wherein,

for RBF kernel function, get γ=1.N _svnumber for support vector.

Use Sigmoid function to carry out nonlinear transformation to decision function, obtain feature new in transformation space:

$Z_{\mathrm{ij}}=\mathrm{\&sigma;}(a_j{f}_j\left(x\right)+b_j)=\frac{1}{1+\exp [-(a_j{f}_j\left(x\right)+b_j)]},j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n$

Wherein, get a _j=1, b _j=1.

Finally obtain the new proper vector in high-dimensional feature space

(k=1,2 ..., n).

The 5th step: fault mode classification

By the 4th step, obtain sample wherein, n=350 is sample size,

for proper vector, c _i∈ 1,2 ..., 7} is sample class label.

Choose at random in every class sample 30 groups as training sample, 20 groups as test sample book.Same step 4, according to matrix H ₇a SVM binary classifier of each row structure, training sample is inputted in each sorter and is trained.Test sample book is input in the SVM training, 7 sorters are differentiated sample respectively, and the output of each sorter forms a binary sequence λ={ λ ₁, λ ₂..., λ ₇.The Hamming distance of sequence of calculation λ and 7 codings, the class of the code word representative that minor increment is corresponding is final differentiation result.

$c_i=\arg \min d(\mathrm{\&lambda;},H_i)=\underset{j=1}{\overset{l}{\mathrm{\&Sigma;}}}|{\mathrm{\&lambda;}}_j-H_{i,j}|,l=7i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,7$ 。

Claims (2)

1. the sensor fault discrimination method combining with support vector machine based on Error Correction of Coding, is characterized in that, comprises the following steps:

Step 1, generation Error Correction of Coding: adopt the coding with error correcting capability to encode to classification, using SVM as sorter, according to preset rules, by Hadamard matrix, obtain qualified Error Correction of Coding;

Step 2, initial characteristics extract: extract time domain and the frequency domain information of sensor output signal, choose 6 temporal signatures parameters and 3 frequency domain character parameters as primitive character;

Step 3, feature extraction; A SVM binary classifier of each row structure of the Error Correction of Coding being obtained by step 1, the sample with initial characteristics is inputted in each SVM and trained respectively, obtain the decision function of each sorter, decision function is carried out to Sigmoid conversion, obtain feature new in transformation space;

Step 4, fault mode classification: according to a SVM binary classifier of each row structure of step 3 encoder matrix, the training sample with new feature parameter is input in each SVM and is trained, test sample book is input in the SVM training, and each sorter is differentiated sample respectively; The output of sorter forms a binary sequence λ={ λ ₁, λ ₂..., λ _n, calculate the Hamming distance between this sequence and classification code word, the class of the code word representative that minor increment is corresponding is final differentiation result.

2. the sensor fault discrimination method combining with support vector machine based on Error Correction of Coding as claimed in claim 1, is characterized in that, uncorrelated between the row of encoder matrix in the coding described in step 1, uncorrelated and not complementary between row.

Images