CN103514458B - The sensor fault discrimination method combined with support vector machine based on Error Correction of Coding - Google Patents

Abstract

The invention provides a kind of sensor fault discrimination method combined based on Error Correction of Coding with support vector machine.Step one, generation Error Correction of Coding: use the coding with error correcting capability classification to be encoded, using SVM as grader；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；Every string one SVM binary classifier of structure of the Error Correction of Coding obtained by step one, the sample with initial characteristics is inputted in each SVM and is trained respectively, obtain the decision function of each grader, decision function is carried out Sigmoid conversion, obtain feature new in transformation space；Step 4, fault mode classification.

Description

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

Technical field

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

Background technology

Sensor is the sensing element in measuring instrument, smart instrumentation and computerized information input equipment, 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 generally relatively more severe, and they break down the most because of various reasons.When the sensor fails, its output signal mainly shows as following several form: deviation, drift about, impact, PERIODIC INTERFERENCE, short circuit, open circuit.After detecting that fault occurs, it is required for different sensor fault types and carries out certain online or off-line Fault Compensation, therefore, sensor fault is carried out identification and is just particularly important.

Sensor is carried out fault identification and belongs to pattern recognition problem.Two the most the most key steps are characterized extraction and pattern classification.The selection of feature and the basis that extraction is pattern classification, directly affect the accuracy of classification results.Having non-linear relation between fault and failure cause that sensor occurs, complexity, randomness and ambiguity make it difficult to represent with accurate mathematical model, will extract this rule, then need to use certain nonlinear transformation to make feature have higher separability.

In recent years, neutral net and support vector machine (SVM, Support Vector Machine), as the representative of Nonlinear Classifier, are widely used in pattern classification.Method based on neutral net all rely on greatly sample fully in the case of statistical property, and sensor fault identification is a kind of typical small-sample learning problem, support vector machine solving in terms of Small Sample Database classification problem, has the features such as global optimum, simple in construction, Generalization Ability are strong.Compared with neutral net, SVM avoids locally optimal solution problem, the problem effectively overcoming " dimension disaster ".

At present, common support vector machine Multiclass Classification mainly has: 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 cassification Algorithm for Training speed one to one, but when classification number increases, classification speed can be slack-off.One-to-many support vector cassification algorithm principle is simple, but classification accuracy is the highest, and will use all training samples when training every time, and training speed can decline.Binary-tree support vector machine Multiclass Classification classification speed is fast, but training speed is relatively slow, there is the phenomenon of wrong point of accumulation.The generalization of Error Correction of Coding support vector machine classification method is preferable, and classification speed is fast.

If able to the particular problem according to sensor fault identification builds qualified Error Correction of Coding, Error Correction of Coding support vector machine classification method is used for this problem, then fast and accurately fault can be carried out identification.

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 combined based on Error Correction of Coding with support vector machine, for the signal of 7 kinds of different modes of sensor in the method, choose time domain respectively and frequency domain character is vectorial as initial characteristics, thus different fault modes is distinguished.

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

Step one, generation Error Correction of Coding: use the coding with error correcting capability that classification is encoded, using SVM as grader, according to preset rules, 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；Every string one SVM binary classifier of structure of the Error Correction of Coding obtained by step one, the sample with initial characteristics is inputted in each SVM and is trained respectively, obtain the decision function of each grader, decision function is carried out Sigmoid conversion, obtain feature new in transformation space；

Step 4, fault mode classification: construct a SVM binary classifier according to every string of step 3 encoder matrix, it is input to the training sample with new feature parameter in each SVM be trained, test sample is input in the SVM trained, and sample is differentiated by each grader respectively；Output one binary sequence λ={ λ of composition of grader₁, λ₂..., λ_n, calculating the Hamming distance between this sequence and classification code word, the class representated by code word that minimum range is corresponding is final differentiation result.

In coding described in step 1 uncorrelated between the row of encoder matrix, uncorrelated and the most complementary between row.

Beneficial effects of the present invention: the present invention can by sensor normal mode and deviation, drift about, impact, PERIODIC INTERFERENCE, short circuit, 6 kinds of typical fault modes of open circuit distinguish；Real-time and classification accuracy rate are obtained for guarantee, and generalization is preferable, particularly embodies bigger advantage under small sample input condition.

Accompanying drawing explanation

Fig. 1 is sensor fault discrimination method principle schematic；

Fig. 2 is sensor fault discrimination method flow chart.

Detailed description of the invention

Referring to the drawings 1, the invention will be further described, and it is as follows that the present invention implements step:

The first step: Signal Pretreatment

Gather each 50 groups of X of the sensor output data under 7 kinds of states_ij(i=1,2 ..., 7j=1,2 ..., 50), for making the signal characteristic of extraction not affected by amplitude, first it is standardized signal processing:

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

Wherein: X_ijRepresent the sensor output signal of different mode, E (X_ij) it is X_ijAverage,For X_ijStandard deviation.

Second step: initial characteristics extracts

Extract the peak index of preprocessed signal, root-mean-square value, kurtosis index, skewness index, waveform index, margin index, gravity frequency, mean square frequency, frequency variance as primitive character.The initial characteristics parameter often being organized signal is Z_i={z_i1, z_i2..., z_i9(i=1,2 ..., 350).

3rd step: generate Error Correction of Coding

Error Correction of Coding is obtained by Hadamard matrix.

(1) Hadamard matrix is generated.If 2^j-1＜ k≤2^j, then 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/2Represent H_N/2In element take 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) matrix obtaining step (2) takes its front k row, then obtain required k × (2^j-1) Error Correction of Coding.

Sensor is normally classified with fault 7 kinds of patterns totally, takes k=7.According to the generation step of above-mentioned Error Correction of Coding, obtaining corresponding encoder matrix is:

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

The matrix now obtained eliminates first row and is all zero-sum shortcoming conditional to classification number, has the feature that row separates and row separate, it is possible to meet the requirement of Error Correction of Coding in multicategory classification problem simultaneously.

4th step: feature extraction

Sensor signal features and failure symptom have non-linear relation, use method based on Error Correction of Coding Yu support vector machine that initial characteristics is carried out nonlinear transformation and carry out enhancing signal feature.

Sample χ={ (x is obtained by second step₁, c₁), (x₂, c₂) ..., (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 H₇Every string structure one SVM binary classifier.Wherein the composition of jth SVM training sample is H₇Middle jth row value be 0 all samples be classified as the 1st class, all samples that value is 1 are classified as the 2nd class.

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

$f\left(x_j\right)=\underset{i=1}{\overset{N_{\mathrm{sv}}}{\mathrm{\Σ}}}{\mathrm{\α}}_i{c}_{i}k(x_i,x_j)+b,j=\mathrm{1,2},\·\·\·,n$

Wherein,For RBF kernel function, take γ=1.N_svFor supporting the number of vector.

Use Sigmoid function that decision function is carried out nonlinear transformation, obtain feature new in transformation space:

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

Wherein, a is taken_j=1, b_j=1.

Finally obtain the new characteristic vector in high-dimensional feature space(k=1,2 ..., n).

5th step: fault mode classification

Sample is obtained by the 4th stepWherein, n=350 is sample size,It is characterized vector, c_i∈ 1,2 ..., 7} is sample class label.

Randomly selecting 30 groups in every class sample as training sample, 20 groups as test sample.Same step 4, according to matrix H₇Every string structure one SVM binary classifier, training sample is inputted in each grader and is trained.Test sample being input in the SVM trained, sample is differentiated by 7 graders respectively, output one binary sequence λ={ λ of composition of each grader₁, λ₂..., λ₇}.Sequence of calculation λ and the Hamming distance of 7 codings, the class representated by code word that minimum range is corresponding is final differentiation result.

$c_i=\arg \min d(\mathrm{\λ},H_i)=\underset{j=1}{\overset{l}{\mathrm{\Σ}}}|{\mathrm{\λ}}_j-H_{i,j}|,l=7i=\mathrm{1,2},\·\·\·,7$ 。

Claims (2)

1. the sensor fault discrimination method combined with support vector machine based on Error Correction of Coding, its feature exists In, comprise the following steps:

Step one, generation Error Correction of Coding: use the coding with error correcting capability to encode classification, will prop up Hold vector machine SVM as grader, according to preset rules, Hadamard Hadamard matrix met The Error Correction of Coding of condition；

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

Step 3, feature extraction；Every string of the Error Correction of Coding obtained by step one constructs one and supports vector Machine SVM binary classifier, inputs in each support vector machines to enter respectively by the sample with initial characteristics Row training, obtains the decision function of each grader, decision function is carried out Sigmoid conversion, is become Change feature new in space；

Step 4, fault mode classification: construct one according to every string of step 3 encoder matrix and support vector Machine SVM binary classifier, is input to each support vector machines by the training sample with new feature parameter In be trained, test sample is input in the support vector machines trained, and each grader is the most right Sample differentiates；Output one binary sequence λ={ λ of composition of grader₁, λ₂..., λ_n, calculate this sequence and Hamming distance between classification code word, the class representated by code word that minimum range is corresponding is final differentiation knot Really.

2. the sensor fault combined with support vector machine based on Error Correction of Coding as claimed in claim 1 Discrimination method, it is characterised in that uncorrelated between the row of encoder matrix in the coding described in step 1, arranges it Between uncorrelated and the most complementary.