CN104848883A - Sensor noise and fault judging method based on sparse representation - Google Patents

Abstract

The invention discloses a sensor noise and fault judging method based on sparse representation. The specific method includes the following steps: 1. an overcomplete atom library containing normal signals, noise and fault samples corresponding thereto is built through historical data; 2. based on a hypothesis that linear representation of unknown samples of some category can be effectively realized in a corresponding subspace by a plurality of samples of the category, sparse representation of a mixed signal collected by a sensor is performed with a built dictionary, i.e., atoms best matched with the signal to be decomposed is found out from the overcomplete atom library and are subjected to linear reconstruction to obtain a new representation mode; 3. a reconstruction error of the sample subjected to linear reconstruction is calculated, and an error of a new sample reconstructed by use of a data set of each kind of noise and fault samples is obtained; 4. data of the same kind of fault samples and data of the same kind of noise samples are used to perform training to calculate a corresponding reconstruction error value; and 5. noise and fault judgment is realized through calculation of the reconstruction error.

Description

A kind of based on the sensor noise of rarefaction representation and the method for discrimination of fault

Technical field

The present invention relates to gas sensor signal processing technology field, be specifically related to a kind of based on the sensor noise of rarefaction representation and the method for discrimination of fault.

Background technology

Sensor technology, as one of the three large ingredients of infotech, has obtained and has applied extremely widely.Sensor is equivalent to the sense organ organ of people, be used for perception, measure various physical quantity, chemical quantity changed into electrical signal form, send into computing machine or electronic circuit processes, the object reaching monitoring or control.Along with the fast development of sensor technology, sensor great variety of goods, precision improves a lot, and the use fan of sensor also increases rapidly, and noise and the failure problems of sensor also more and more come into one's own.

Along with Modernization System structure is increasingly sophisticated, scale constantly expands, automaticity is more and more higher, the requirement of people to the reliability of this type systematic and security is also just more and more higher, requirement system has stronger fault-tolerance and becomes more and more important, fault diagnosis is then the major prerequisites realizing faults-tolerant control, and a wherein the most key step is exactly how to carry out detection and the isolation of fault in time.Sensor, as the source of acquisition of information, plays significant role in modern control system.The operation of its direct influential system of measurement result, the correctness of impact analysis, decision-making, particularly in the large-scale complicated system such as chemical industry, Aerospace test, once sensor failure, consequence is by hardly imaginable.Sensor not only may produce fault in the course of the work, and also has self and outside noise, therefore how carries out difference to the noise of sensor and fault and seems particularly important.This not only can reduce the off-time, increases the security of system cloud gray model, reduces manufacturing cost, enterprise can also be made to avoid the massive losses of personnel and property, bring considerable economic benefit to enterprise.

In prior art, rarefaction representation grows up along with compressive sensing theory, it is risen the nineties in 20th century, the research of rarefaction representation can be summed up as the design of sparse dictionary usually, the design of dictionary is a major issue in rarefaction representation, is divided into orthogonal basis dictionary, crosses complete dictionary etc.Rarefaction representation based on the complete dictionary of mistake is the focus of current signal transacting area research, and the object of rarefaction representation is exactly select less atom to represent raw data in the complete dictionary of mistake, namely makes the number of this expression coefficient non-zero few as much as possible.At present, rarefaction representation has been widely applied to many aspects of the signal transacting such as denoising, compression, coding, parameter estimation, feature extraction, target identification.

Summary of the invention

For the problems referred to above, the present invention proposes a kind of based on the sensor noise of rarefaction representation and the method for discrimination of fault, to solve the problem that moment sensor noise and fault cannot differentiate, reduce disturbing factor during sensor fault diagnosis, better solve fault, guarantee system is better run.

The present invention proposes a kind of based on the sensor noise of rarefaction representation and the method for discrimination of fault.Comprise the following steps:

(1) simultaneously corresponding containing normal signal, noise and fault sample over-complete dictionary of atoms is constructed by historical data;

(2) the dictionary rarefaction representation of the mixed signal structure collected by sensor, namely finds out the atom that mates the most with signal to be decomposed and obtains a new expression way by linear reconstruction from over-complete dictionary of atoms;

(3) by the sample of linear reconstruction with based on δ _jdefinition and the mathematic interpolation reconstructed error of linear expression of set B;

(4) adopt similar fault sample data and similar noise sample data to carry out training respectively and calculate corresponding reconstructed error value m _i, m _j, i=1 ..., J, j=1 ..., J;

(5) utilize reconstructed error to realize the judgement of noise and fault.If the reconstructed error calculated levels off to m _i, then judge it is fault, if the reconstructed error calculated levels off to m _j, then judge it is noise, and corresponding failure and noise type can be judged.

Accompanying drawing explanation

In order to make the object, technical solutions and advantages of the present invention clearly, below in conjunction with accompanying drawingthe present invention is described in further detail, wherein:

fig. 1for flow chart element of the present invention figure.

Embodiment

Below in conjunction with accompanying drawing, elaborate embodiments of the present invention:

(1) rarefaction representation achieves significant development in signal field, shows lot of advantages, is also widely used in the every field such as recognition of face, image procossing, target following.The present invention adopts the method for rarefaction representation for distinguishing sensor noise and fault, can obtain good effect equally.In the present invention first by training structure sensor noise and fault sample dictionary.

Based on the structure of sensor noise and fault sample dictionary, set up the complete set of mistake of sensor each noise like and representation for fault in the course of the work exactly.When structure sensor noise and fault data collection, this data set comprises normal data, multiclass noise sample data and multiclass fault data, and the present invention concentrates the sample data choosing some to be used for building noise and fault sample dictionary from the training data of every kind.Wherein the expression sequence of each sample data is a column vector in noise or fault sample dictionary.

Particularly, sensor often kind of typical noise in the course of the work and fault sample data are utilized to set up sample set:

$A_j=\{a_j^1,a_j^2,\·\·\·,a_j^K\}---\left(1\right)$

Wherein, K represents jth noise like or fault sample number.The sample data of all noises and fault type is merged, forms noise and fault sample dictionary, following form:

$B=\∪\left\{A_j\right\}=\{a_1^1,a_1^2,...,a_1^K,a_2^1,a_2^2,...,a_J^K\}---\left(2\right)$

Wherein, j=1...J.J represents typical normal type and various noise and failure mode number sum in this course of work.Therefore, we can obtain the sample dictionary B that is made up of K × J noise and fault sample.

(2) in rarefaction representation, suppose that the unknown sample of a certain class can be carried out linear expression by such some samples in respective subspace effectively.Because in sensor actual mechanical process, often can comprise normal mode, a lot of fault mode and noise pattern, thus the noise of structure and fault sample dictionary also very large.Obtain new sample by linear reconstruction in this invention, formula is as follows:

Bψ≈F _unknow(3)

Wherein, F _unknowrepresent new samples, j=1 ..., J, k=1 ..., K is the coefficient vector of sample in corresponding B, it is exactly the reconstruction coefficients of jth kind noise or a fault kth sample.

(3) in actual experiment, j=1 ..., J, k=1 ..., only have l to be the coefficient (l < < K) of non-zero in K.In mathematics, such matrix is called the l rarefaction representation of sample.The number of nonzero coefficient can be used || ψ || ₀represent, so by minimizing || ψ || ₀obtain the sparse reconstruct of sample.The problem of this 0-norm is the problem of NP-hard.Here, by 1-norm minimum with solving rarefaction representation problem:

arg min||ψ|| ₁

s.t.||Bψ-F _unknow|| ₂≤ε (4)

Wherein, || || ₁represent 1-norm, ε > 0 represents a less numerical value.By noise and fault sample dictionary B and the sparse coefficient vector ψ calculated, new samples is obtained by reconstruct, and reconstruction formula is:

$F_{\mathrm{unknow}}\&ap;\mathrm{B\&psi;}=\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{B}^k{\mathrm{\&psi;}}^k---\left(5\right)$

Based on the sparse reconstruct of 1-norm minimum, ensure that this reconstruct is compacted, that is, pick out from noise and fault sample data centralization and have the sample of representative most to represent new samples and specimen reconstruct, so effectively disclose the principal character of new samples and immanent structure and noise by sparse reconstruction coefficients and reconstructed error, which sample of fault sample data centralization matches.

(4) when the sparse reconstruct of test sample book completes, formula (4) is adopted to calculate its sparse coefficient ψ to all noises and fault sample data acquisition B.For each noise or fault mode j, define its characteristic of correspondence function δ _j.Wherein δ _jbe defined as and only retain jth kind noise in noise and fault sample data acquisition B, sparse coefficient corresponding to fault mode, simultaneously by the sparse coefficient corresponding to other noises and fault mode sample all assignment be 0.Based on δ _jdefinition and set B, utilize reconstructed error to judge that new samples belongs to which kind of noise pattern, fault mode or other patterns, reconstructed error computing formula is as follows:

r _j(F _unknow)＝||F _unknow-Bδ _j(ψ)|| ₂,j＝1,...,J (6)

Wherein, r _j(F _unknow) represent error when utilizing jth kind noise or fault mode sample data to be reconstructed new samples, obtain utilization successively by the method and often plant the error that noise and fault sample data set be reconstructed new samples.Then adopt similar fault sample data and similar noise sample data to carry out training respectively and calculate corresponding reconstructed error value m _i, m _j, i=1 ..., J, j=1 ..., J.Unknown sample based on a certain class can be carried out this hypothesis of linear expression by such some samples in respective subspace effectively, think to adopt when sparse reconstruct being carried out to new samples with noise like and fault sample data to obtain reconstructed error be minimum.So the present invention can not only realize the judgement of noise and fault, and can judge it is noise and fault type.

(5) normal signal, noise signal and fault-signal three kinds in this invention, is had.According to above analysis, can determine that sampled signal is that noise interferences or sensor there occurs fault.If normal signal then directly does not need to analyze; If the reconstructed error that sampled signal after reconstruction, calculates is minimum, if the reconstructed error calculated levels off to m _i, then judge it is fault; If the reconstructed error calculated levels off to m _j, then judge it is noise, and corresponding failure and noise type can be judged.

Claims (6)

1., based on the sensor noise of rarefaction representation and a method of discrimination for fault, it is characterized in that, comprise the steps:

Step one: set up simultaneously corresponding containing normal signal, noise and fault sample over-complete dictionary of atoms by historical data;

Step 2: the dictionary rarefaction representation of the mixed signal structure collected by sensor, namely finds out the atom that mates the most with signal to be decomposed and obtains a new expression way by linear reconstruction from over-complete dictionary of atoms;

Step 3: by the sample of linear reconstruction with based on δ _jdefinition and the mathematic interpolation reconstructed error of linear expression of set B;

Step 4: adopt similar fault sample data and similar noise sample data to carry out training and calculate corresponding reconstructed error value m _i, m _j, i=1 ..., J, j=1 ..., J;

Step 5: realize noise and fault distinguishing by comparing reconstructed error.

2. according to claim 1, a kind of based on the sensor noise of rarefaction representation and the method for discrimination of fault, it is characterized in that: in step one, in the process constructing noise and fault sample dictionary, because industrial process is often containing the many fault modes and the noise pattern that comprise normal mode, need to train mass data, so the noise of the present invention's structure and fault sample dictionary need very large.

3. according to claim 1, a kind of based on the sensor noise of rarefaction representation and the method for discrimination of fault, it is characterized in that: in step 2, need when rarefaction representation the unknown sample supposing a certain class effectively can be carried out linear expression by such some samples in respective subspace, and be from over-complete dictionary of atoms, find out the atom that mates the most with signal to be decomposed obtain a new expression way by linear reconstruction.

4. according to claim 1, a kind ofly to it is characterized in that: in step 3 based on the sensor noise of rarefaction representation and the method for discrimination of fault, for each noise pattern or fault mode j, define its characteristic of correspondence function δ _j.Wherein δ _jbe defined as and only retain jth kind noise or the sparse coefficient corresponding to fault mode in noise and fault sample data acquisition B, simultaneously by the sparse coefficient corresponding to other noises or fault mode sample all assignment be 0.Based on δ _jdefinition and set B, utilize linear reconstruction sample and based on δ _jdefinition and the mathematic interpolation reconstructed error of linear expression of set B realize to noise and fault differentiation.

5. according to claim 1, a kind ofly to it is characterized in that: in step 4 based on the sensor noise of rarefaction representation and the method for discrimination of fault, adopt similar fault sample data and similar noise sample data to carry out training and calculate corresponding reconstructed error value m _i, m _j, i=1 ..., J, j=1 ..., J; At calculating m _i, m _jtime, need that training is carried out to mass data and calculate, to the training result averaged of every type.

6. according to claim 1, a kind of based on the sensor noise of rarefaction representation and the method for discrimination of fault, it is characterized in that: in step 5, when the sparse reconstruct of test sample book completes, when calculating reconstructed error, when adopting similar noise sample data and similar fault sample data to carry out sparse reconstruct to new samples in the present invention to obtain reconstructed error be minimum, if the reconstructed error calculated levels off to m _i, then judge it is fault, if the reconstructed error calculated levels off to m _j, then judge it is noise, and corresponding failure and noise type can be judged.