CN105116872A - Fault diagnosis method based on metric learning and time sequence during industrial process - Google Patents

Abstract

The present invention relates to a fault diagnosis method based on a metric learning and time sequence during an industrial process for solving the problems that a conventional fault diagnosis method is high in system cost, is difficult for on-line diagnosis, is difficult to distinguish to the fault types, etc. The method is realized by a step 1 of dividing the system faults into n types; a step 2 of preparing a training sample; a step 3 of pre-processing the training sample; a step 4 of carrying out the metric learning on the pre-processed training sample; a step 5 of calculating the distances between a real-time monitoring sample and n subclasses; a step 6 of according to the distances between the real-time monitoring sample and the n subclasses, adopting a KNN classification method to diagnose the faults, namely determining whether a system generates the faults and determining the types of the faults. The fault diagnosis method based on the metric learning and time sequence during the industrial process of the present invention is applied to the fault diagnosis field.

Description

Based on metric learning and seasonal effect in time series method for diagnosing faults in a kind of industrial process

Technical field

The present invention relates to based on metric learning and seasonal effect in time series method for diagnosing faults, particularly in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults.

Background technology

Fault diagnosis system relates to the numerous areas such as iron and steel, boiler, chemical industry, pharmacy, has become ingredient important in modern industry production.

Modern industry process is generally tending towards maximizing and there is the feature of modelling by mechanism difficulty, in order to moment supervisory control system running state, often carries out long-term measurement to the many state variables in production run, thus obtains a large amount of on-site supervision data.Much system cloud gray model information is comprised significant to evaluation system running status in a large amount of monitor data, but between each variable be not often independently but there is certain correlativity to each other, complicated correlativity makes technician be difficult to find with observation and experience the true cause producing fault.How avoiding the present situation of modelling by mechanism difficulty, directly carry out rationally these data, utilize design error failure diagnostic system efficiently, thus ensure the quality of final products, is one of significant challenge of facing of present industrial process control system.The good method for diagnosing faults based on data, effectively can simplify and utilize monitor data information, and obtains accurate, a rational Diagnostic Strategy.The current design based on the method for diagnosing faults of data has received to be paid attention to widely.

Existing method for diagnosing faults exists that system cost is high, inline diagnosis is difficult and fault type is difficult to the problems such as resolution, adopts and can effectively solve the problem based on metric learning and seasonal effect in time series method for diagnosing faults, have good Generalization Ability.

Summary of the invention

The object of the invention is have that system cost is high, inline diagnosis is difficult and fault type is difficult to the problems such as resolution to solve existing method for diagnosing faults, and based on metric learning and seasonal effect in time series method for diagnosing faults in a kind of industrial process proposed.

Above-mentioned goal of the invention is achieved through the following technical solutions:

Step one, analyze history monitor data, the system failure is divided into n type and n subclass by the difference according to system failure producing cause;

Normal data under step 2, acquisition system nominal situation and n kind fault condition running status is as training sample;

Step 3, seasonal effect in time series method is utilized to carry out pre-service for training sample; Wherein, described preprocess method refers to and adopts wavelet transformation to process training sample;

Step 4, carry out metric learning to through pretreated training sample, the metric matrix A of the standard that the metric matrix of generation system is measured as sample similarity and generation system;

Step 5, wavelet transformation is carried out to Real-Time Monitoring sample after, the standard utilizing sample similarity to measure asks for the distance of Real-Time Monitoring sample and n subclass;

Step 6, distance according to Real-Time Monitoring sample and n subclass, adopt KNN sorting technique to carry out fault diagnosis and namely judge whether system breaks down and fault type; Namely complete in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults.

Invention effect

The present invention in order to solve that existing method for diagnosing faults system cost is high, inline diagnosis is difficult and fault type is difficult to the problem differentiated, thus proposes based on metric learning and seasonal effect in time series method for diagnosing faults.The method mainly used has metric learning method, wavelet transformation and KNN classification.

Advantage of the present invention is:

1, improve fault diagnosis rate of accuracy reached to 5% ~ 10%, metric learning and seasonal effect in time series method are combined, whether both utilizations advantage is for break down and fault type can better be diagnosed.

2, improve real-time, the combination of two kinds of methods for fault diagnosis and failure modes work consuming time short, the speed of detection can be improved.

3, improve adaptability.

4, improve applicability, the combination of two kinds of methods make the design of fault diagnosis system and operation more simple, the significant increase generalization of the method.

The industrial process method for diagnosing faults of current acquisition widespread use, such as: principle component analysis, in the relevant information between dissimilar malfunction monitoring sample not being considered, so this type of method for diagnosing faults can only judge whether industrial system breaks down, but accurately can not judge the type broken down, so just cause finding out the very first time position of breaking down in the reason and industrial system that fault occurs.The fault that can not occur in resolution system in time may cause more serious industrial accident or huger economic loss.

In relevant information between dissimilar malfunction monitoring sample is considered by metric learning method, can in the type judging accurately to judge while whether industrial system breaks down to break down.Meanwhile, the metric matrix of trying to achieve in metric learning method can directly apply to on-line fault diagnosis after calculated off-line, and this greatly reduces the calculated amount of on-line fault diagnosis, reduces the cost of online system failure diagnosis.

Accompanying drawing explanation

Fig. 1 be embodiment one propose based on metric learning and seasonal effect in time series Troubleshooting Flowchart.

Embodiment

Embodiment one: based on metric learning and seasonal effect in time series method for diagnosing faults in a kind of industrial process of present embodiment, specifically prepare according to following steps:

Step one, analyze history monitor data, the system failure is divided into n type and n subclass by the difference according to system failure producing cause;

Normal data under step 2, acquisition system nominal situation and n kind fault condition running status is as training sample;

Step 3, seasonal effect in time series method is utilized to carry out pre-service for training sample; Wherein, described preprocess method refers to and adopts wavelet transformation to process training sample;

Step 4, carry out metric learning to through pretreated training sample, the metric matrix A of the standard that the metric matrix of generation system is measured as sample similarity and generation system;

Step 5, wavelet transformation is carried out to Real-Time Monitoring sample after, the standard (metric matrix of generation system) utilizing sample similarity to measure asks for the distance of Real-Time Monitoring sample and n subclass; Such as have 10 subclasses, each subclass has 100 monitor sample, and the distance of Real-Time Monitoring sample and 10 subclasses is respectively the distance of each Real-Time Monitoring sample respectively and between 1000 monitor sample;

Step 6, distance according to Real-Time Monitoring sample and n subclass, adopt KNN sorting technique to carry out fault diagnosis and namely judge whether system breaks down and fault type; Namely complete in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults.

Present embodiment effect:

Present embodiment in order to solve that existing method for diagnosing faults system cost is high, inline diagnosis is difficult and fault type is difficult to the problem differentiated, thus proposes based on metric learning and seasonal effect in time series method for diagnosing faults.The method mainly used has metric learning method, wavelet transformation and KNN classification.

The advantage of present embodiment is:

1, improve fault diagnosis rate of accuracy reached to 5% ~ 10%, metric learning and seasonal effect in time series method are combined, whether both utilizations advantage is for break down and fault type can better be diagnosed.

2, improve real-time, the combination of two kinds of methods for fault diagnosis and failure modes work consuming time short, the speed of detection can be improved.

3, improve adaptability.

4, improve applicability, the combination of two kinds of methods make the design of fault diagnosis system and operation more simple, the significant increase generalization of the method.

The industrial process method for diagnosing faults of current acquisition widespread use, such as: principle component analysis, in the relevant information between dissimilar malfunction monitoring sample not being considered, so this type of method for diagnosing faults can only judge whether industrial system breaks down, but accurately can not judge the type broken down, so just cause finding out the very first time position of breaking down in the reason and industrial system that fault occurs.The fault that can not occur in resolution system in time may cause more serious industrial accident or huger economic loss.

In relevant information between dissimilar malfunction monitoring sample is considered by metric learning method, can in the type judging accurately to judge while whether industrial system breaks down to break down.Meanwhile, the metric matrix of trying to achieve in metric learning method can directly apply to on-line fault diagnosis after calculated off-line, and this greatly reduces the calculated amount of on-line fault diagnosis, reduces the cost of online system failure diagnosis.

Embodiment two: present embodiment and embodiment one unlike: step 4 vacuum metrics learns to be specially: metric learning the difference between weight analysis different faults categorical data, obtain metric matrix by training sample, have a clear superiority in on-line fault diagnosis aspect; Metric learning realizes the differentiation of fault type by trying to achieve a distance matrix accurately can portraying Sample Similarity;

(1) by metric learning obtain in all training samples with Real-Time Monitoring sample distance M _abe expressed as:

$M_A=\sqrt{{(x-y)}^{T}A(x-y)}$

Wherein, A is metric matrix, and x is through pretreated training sample, and y is Real-Time Monitoring sample; It is worthy of note that the metric matrix A tried to achieve in different industrial system is all not identical, the distribution of the abundant reflected sample of metric matrix energy obtained by respective systematic training sample, can improve the performance of sorter;

(2) basic thought asking for metric matrix is, if training sample and Real-Time Monitoring sample belong to the M that identical type is calculated by metric matrix _adistance is less;

If training sample and Real-Time Monitoring sample belong to the M that different types is calculated by metric matrix _adistance is larger; Problem is described below:

A∈R ^D×D

s.t. $\begin{array}{cc}{M^2}_A(x_i,x_j)\≤s& (i,j)\∈S\\ {M^2}_A(x_i,x_j)\≥b& (i,j)\∈D\end{array}$

Wherein, S represents the set of similar sample to composition, and D represents the set of inhomogeneity sample to composition, s and d is two given threshold values (usually get s be about 1 d be about 100);

(3) initial metric matrix A is established ₀for unit battle array, d be inhomogeneity sample between distance threshold values, s be similar sample between distance threshold values, correlation parameter λ _ijinitial value is 0, correlation parameter ξ _{c (i, j)}when (i, j) belongs to S respectively, initial value is s, correlation parameter ξ _{c (i, j)}when (i, j) belongs to D respectively, initial value is b, γ is the slack variable chosen, and chooses a kind of alternative manner based on Gradient Descent and calculates metric matrix A.Other step and parameter identical with embodiment one.

Embodiment three: present embodiment and embodiment one or two unlike: choose a kind of alternative manner based on Gradient Descent and the concrete iterative process that metric matrix A calculates be described below:

(1) a pair constraint (i, j) is selected;

(2) p=(x is calculated _i-x _j) ^ta _k(x _i-x _j);

(3) when (i, j) ∈ S gets δ=1, otherwise δ=-1 is got;

(4) calculate $a=min({\mathrm{\λ}}_{ij},\frac{\mathrm{\δ}}{2}(\frac{1}{p}-\frac{\mathrm{\γ}}{{\mathrm{\ξ}}_{c(i,j)}}\left)\right);$

(5) calculate $\mathrm{\β}=\frac{\mathrm{\δ}a}{(1-\mathrm{\δ}ap)};$

(6) calculate ${\mathrm{\ξ}}_{c(i,j)}=\frac{{\mathrm{\γ\ξ}}_{c(i,j)}}{\mathrm{\γ}+{\mathrm{\δa\ξ}}_{c(i,j)}};$

(7) λ is calculated _ij=λ _ij-a

(8) A is calculated _k+1=A _k+ β A _k(x _i-x _j) (x _i-x _j) ^ta _k;

Wherein, γ is slack variable; δ does not have implication, is the intermediate quantity of definition in above-mentioned steps (3); A _krefer to the value of kth time iteration vacuum metrics matrix A;

Choose a pair constraint (i, j) and carry out iteration with regard to repeating step (1) ~ (8), until metric matrix A restrains, then obtain the metric matrix A of objective matrix and generation system.Other step and parameter identical with embodiment one or two.

Embodiment four: one of present embodiment and embodiment one to three unlike: in step 5, wavelet transformation detailed process is:

Because the method for diagnosing faults based on learning distance metric is more responsive to the gap on sample values, but in industrial processes the data sample of fault condition and the difference of the data sample of non-fault operating mode to be sometimes presented as in vibration frequency, Oscillation Amplitude and variation tendency different; Adopt wavelet transformation to carry out pre-service to sample data and better can carry out feature extraction, the accuracy rate of raising method fault diagnosis;

(1) wavelet mother function is defined in a kind of function that limited interval and wavelet mother function mean value are zero, and have frequency and the amplitude of limited duration and sudden change, waveform can be irregular, also can be asymmetric; Obtain corresponding wavelet function by flexible and translation wavelet mother function, be expressed as follows:

${\mathrm{\ψ}}_{a,b}\left(t\right)=\frac{1}{\sqrt{a}}\mathrm{\ψ}\left(\frac{t-b}{a}\right)$

In formula, a and b is constant value, a, b ∈ R, and a>0, a are scale factor, and b is shift factor; Scale factor a realizes the coupling to different frequency signals by flexible wavelet basis function, obtains the frequency characteristic of signal; Shift factor b realizes the traversal analysis to signal by making wavelet basis function along time shaft translation, obtains the temporal information of signal; ψ is wavelet mother function; T is the time;

(2) wavelet transformation defining x (t) is as follows:

${\mathrm{WT}}_{a,b}=\frac{1}{\sqrt{a}}{\∫}_{R}x\left(t\right){\mathrm{\ψ}}^{*}\left(\frac{t-b}{a}\right)dt$

Wherein, WT _a,bbe wavelet basis function, * is the meaning of convolution, is a kind of mathematical operation, common practise

The numerical value of wavelet transformation is produced by signal different piece under the different scale factor, adopt a heavy little conversion, namely scale factor is not changed, only change shift factor completes matlab software and carries the time traversal of wavelet mother function cgau4 to sample data (which data is sample data refer to), carries out metric learning can obtain gratifying fault diagnosis result with gained wavelet transform result.Other step and parameter identical with one of embodiment one to three.

Embodiment five: one of present embodiment and embodiment one to four unlike: step 6 KNN sorting technique is specially:

(1) on the basis of trying to achieve metric matrix between sample data, KNN sorting technique is adopted to carry out inline diagnosis to Real-Time Monitoring sample, KNN algorithm is the abbreviated form of k-vicinity method (k-NearestNeighboralgorithm), KNN algorithm primitive rule is the distance calculating Real-Time Monitoring sample and n subclass, find in all training samples with Real-Time Monitoring sample distance M _ak minimum sample; Wherein, k=k ₁+ k ₂+ k ₃+ ...+k _i+ ... k _n

(2) i-th subclass schedule of samples in k sample is shown as k _i, i=1,2 ..., c, definition discriminant function is as follows:

g _i(x)＝k _i,i＝1,2,…,c

Real-Time Monitoring sample is marked as type i, if g _i(x)>=g _jx (), i ≠ j, k-contiguous method generally adopts k to be that odd number is to avoid the equal problem being difficult to decision-making of the discriminant function of the sample occurring two kinds in n kind fault type.Other step and parameter identical with one of embodiment one to four.

Following examples are adopted to verify beneficial effect of the present invention:

Embodiment one:

Based on metric learning and seasonal effect in time series method for diagnosing faults in a kind of industrial process of the present embodiment, specifically prepare according to following steps:

Step one, analyze history monitor data, the system failure is divided into n type and n subclass by the difference according to system failure producing cause;

Normal data under step 2, acquisition system nominal situation and n kind fault condition running status is as training sample;

Tennessee-Yi Siman (TennesseeEastman, TE) process is the simulation reconstruction of true chemical process, is widely used in checking and the contrast of method for diagnosing faults as industry benchmark testing process;

TE process contains 8 reactants and 5 formants, is respectively reacting furnace, condenser, compressor, seperator and stripping tower; Fully describe system cloud gray model information containing 52 observational variables and 21 kinds of operation troubless in TE process, 52 observational variables are 11 manipulated variables and 41 process variable respectively; In 21 kinds of operation troubless, fault 1-15,21 is 16 kinds of known fault types, and fault 16-20 is 5 kinds of unknown failure types;

TE process data is concentrated and is comprised 22 groups of training datas and 22 groups of test datas, respectively corresponding 21 kinds of failure operation data and failure-free operation data.Wherein 22 groups of training datas have recorded the state of 52 observational variables in 24 continuous throughout the twenty-four hour24s; 22 groups of test datas are online record 52 observational variables states in 48 continuous throughout the twenty-four hour24s, and corresponding failure starts to add after 9 hours in operation simultaneously.The sampling interval of 22 groups of training datas and 22 groups of test datas is 3 minutes.TE process data collects can be downloaded in http://brahms.scs.uiuc.edu, and TE process FORTRAN code is also in online announcement simultaneously.

This emulation object has been the validation verification based on metric learning and seasonal effect in time series method for diagnosing faults, chooses TE process data and concentrates fault 2, fault 4, fault 7 and fault 11 service data as standard.The fault diagnosis effect of the comparatively Fei Sheer discriminant analysis method of maturation in method and conventional fault diagnosis method that contrast this patent proposes;

Step 3, seasonal effect in time series method is utilized to carry out pre-service for training sample; Wherein, described preprocess method refers to and adopts wavelet transformation to process training sample;

Step 4, carry out metric learning to through pretreated training sample, the metric matrix A of the standard that the metric matrix of generation system is measured as sample similarity and generation system;

Metric learning the difference between weight analysis different faults categorical data, obtains metric matrix by training sample, has a clear superiority in on-line fault diagnosis aspect; Metric learning realizes the differentiation of fault type by trying to achieve a distance matrix accurately can portraying Sample Similarity;

(1) by metric learning obtain in all training samples with Real-Time Monitoring sample distance M _abe expressed as:

$M_A=\sqrt{{(x-y)}^{T}A(x-y)}$

Wherein, A is metric matrix, and x is through pretreated training sample, and y is Real-Time Monitoring sample; It is worthy of note that the metric matrix A tried to achieve in different industrial system is all not identical, the distribution of the abundant reflected sample of metric matrix energy obtained by respective systematic training sample, can improve the performance of sorter;

(2) basic thought asking for metric matrix is, if training sample and Real-Time Monitoring sample belong to the M that identical type is calculated by metric matrix _adistance is less;

If training sample and Real-Time Monitoring sample belong to the M that different types is calculated by metric matrix _adistance is larger; Problem is described below:

A∈R ^D×D

s.t. $\begin{array}{cc}{M^2}_A(x_i,x_j)\≤s& (i,j)\∈S\\ {M^2}_A(x_i,x_j)\≥b& (i,j)\∈D\end{array}$

Wherein, S represents the set of similar sample to composition, and D represents the set of inhomogeneity sample to composition, s and d is two given threshold values (usually get s be about 1 d be about 100);

(3) initial metric matrix A is established ₀for unit battle array, d be inhomogeneity sample between distance threshold values, s be similar sample between distance threshold values, correlation parameter λ _ijinitial value is 0, correlation parameter ξ _{c (i, j)}when (i, j) belongs to S respectively, initial value is s, correlation parameter ξ _{c (i, j)}when (i, j) belongs to D respectively, initial value is b, γ is the slack variable chosen, and chooses a kind of alternative manner based on Gradient Descent and calculates metric matrix A.

Choose a kind of alternative manner based on Gradient Descent to be described below the concrete iterative process that metric matrix A calculates:

(1) a pair constraint (i, j) is selected;

(2) p=(x is calculated _i-x _j) ^ta _k(x _i-x _j);

(3) when (i, j) ∈ S gets δ=1, otherwise δ=-1 is got;

(4) calculate $a=min({\mathrm{\λ}}_{ij},\frac{\mathrm{\δ}}{2}(\frac{1}{p}-\frac{\mathrm{\γ}}{{\mathrm{\ξ}}_{c(i,j)}}\left)\right);$

(5) calculate $\mathrm{\β}=\frac{\mathrm{\δa}}{(1-\mathrm{\δap})};$

(6) calculate ${\mathrm{\ξ}}_{c(i,j)}=\frac{{\mathrm{\γ\ξ}}_{c(i,j)}}{\mathrm{\γ}+{\mathrm{\δa\ξ}}_{c(i,j)}};$

(7) λ is calculated _ij=λ _ij-a;

(8) A is calculated _k+1=A _k+ β A _k(x _i-x _j) (x _i-x _j) ^ta _k;

Wherein, γ is slack variable; δ does not have implication, is the intermediate quantity of definition in above-mentioned steps (3); A _krefer to the value of kth time iteration vacuum metrics matrix A;

Choose a pair constraint (i, j) and carry out iteration with regard to repeating step (1) ~ (8), until metric matrix A restrains, then obtain the metric matrix A of objective matrix and generation system.

Step 5, wavelet transformation is carried out to Real-Time Monitoring sample after, the standard (metric matrix of generation system) utilizing sample similarity to measure asks for the distance of Real-Time Monitoring sample and n subclass; Such as have 10 subclasses, each subclass has 100 monitor sample, and the distance of Real-Time Monitoring sample and 10 subclasses is respectively the distance of each Real-Time Monitoring sample respectively and between 1000 monitor sample;

Because the method for diagnosing faults based on learning distance metric is more responsive to the gap on sample values, but in industrial processes the data sample of fault condition and the difference of the data sample of non-fault operating mode to be sometimes presented as in vibration frequency, Oscillation Amplitude and variation tendency different.Adopt wavelet transformation to carry out pre-service to sample data and better can carry out feature extraction, the accuracy rate of raising method fault diagnosis;

(1) wavelet mother function is defined in a kind of function that limited interval and wavelet mother function mean value are zero, and have frequency and the amplitude of limited duration and sudden change, waveform can be irregular, also can be asymmetric; Obtain corresponding wavelet function by flexible and translation wavelet mother function, be expressed as follows:

${\mathrm{\ψ}}_{a,b}\left(t\right)=\frac{1}{\sqrt{a}}\mathrm{\ψ}\left(\frac{t-b}{a}\right)$

In formula, a and b is constant value, a, b ∈ R, and a>0, a are scale factor, and b is shift factor; Scale factor a realizes the coupling to different frequency signals by flexible wavelet basis function, obtains the frequency characteristic of signal; Shift factor b realizes the traversal analysis to signal by making wavelet basis function along time shaft translation, obtains the temporal information of signal; ψ is wavelet mother function; T is the time;

(2) wavelet transformation defining x (t) is as follows:

${\mathrm{WT}}_{a,b}=\frac{1}{\sqrt{a}}{\∫}_{R}x\left(t\right){\mathrm{\ψ}}^{*}\left(\frac{t-b}{a}\right)dt$

Wherein, WT _a,bbe wavelet basis function, * is the meaning of convolution, is a kind of mathematical operation, common practise

The numerical value of wavelet transformation is produced by signal different piece under the different scale factor, adopt a heavy little conversion, namely scale factor is not changed, only change shift factor completes matlab software and carries the time traversal of wavelet mother function cgau4 to sample data (which data is sample data refer to), carries out metric learning can obtain gratifying fault diagnosis result with gained wavelet transform result.

Step 6, distance according to Real-Time Monitoring sample and n subclass, adopt KNN sorting technique to carry out fault diagnosis and namely judge whether system breaks down and fault type;

(1) on the basis of trying to achieve metric matrix between sample data, KNN sorting technique is adopted to carry out inline diagnosis to Real-Time Monitoring sample, KNN algorithm is the abbreviated form of k-vicinity method (k-NearestNeighboralgorithm), KNN algorithm primitive rule is the distance calculating Real-Time Monitoring sample and n subclass, find in all training samples with Real-Time Monitoring sample distance M _ak minimum sample; Wherein, k=k ₁+ k ₂+ k ₃+ ...+k _i+ ... k _n

(2) i-th subclass schedule of samples in k sample is shown as k _i, i=1,2 ..., c, definition discriminant function is as follows:

g _i(x)＝k _i,i＝1,2,…,c

Real-Time Monitoring sample is marked as type i, if g _i(x)>=g _jx (), i ≠ j, k-contiguous method generally adopts k to be that odd number is to avoid the equal problem being difficult to decision-making of the discriminant function of the sample occurring two kinds in n kind fault type.

Following table is simulation result

The present invention also can have other various embodiments; when not deviating from the present invention's spirit and essence thereof; those skilled in the art are when making various corresponding change and distortion according to the present invention, but these change accordingly and are out of shape the protection domain that all should belong to the claim appended by the present invention.

Claims (5)

1. in industrial process based on metric learning and a seasonal effect in time series method for diagnosing faults, it is characterized in that specifically carrying out according to following steps based on metric learning and seasonal effect in time series method for diagnosing faults in a kind of industrial process:

Step one, analyze history monitor data, the system failure is divided into n type and n subclass by the difference according to system failure producing cause;

Normal data under step 2, acquisition system nominal situation and n kind fault condition running status is as training sample;

Step 3, seasonal effect in time series method is utilized to carry out pre-service for training sample; Wherein, described preprocess method refers to and adopts wavelet transformation to process training sample;

Step 4, carry out metric learning to through pretreated training sample, the metric matrix A of the standard that the metric matrix of generation system is measured as sample similarity and generation system;

Step 5, wavelet transformation is carried out to Real-Time Monitoring sample after, the standard utilizing sample similarity to measure asks for the distance of Real-Time Monitoring sample and n subclass;

Step 6, distance according to Real-Time Monitoring sample and n subclass, adopt KNN sorting technique to carry out fault diagnosis and namely judge whether system breaks down and fault type; Namely complete in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults.

2. according to claim 1 in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults, it is characterized in that: the study of step 4 vacuum metrics is specially:

(1) by metric learning obtain in all training samples with Real-Time Monitoring sample distance M _abe expressed as:

$M_A=\sqrt{{(x-y)}^{T}A(x-y)}$

Wherein, A is metric matrix, and x is through pretreated training sample, and y is Real-Time Monitoring sample;

(2) if training sample and Real-Time Monitoring sample belong to the M that identical type is calculated by metric matrix _adistance is less;

If training sample and Real-Time Monitoring sample belong to the M that different types is calculated by metric matrix _adistance is larger; Problem is described below:

A∈R ^D×D

$\begin{array}{cc}s.t.& \begin{array}{cc}{M^2}_A(x_i,x_j)\≤s& (i,j)\∈S\\ {M^2}_A(x_i,x_j)\≥b& (i,j)\∈D\end{array}\end{array}$

Wherein, S represents the set of similar sample to composition, and D represents the set of inhomogeneity sample to composition, s and d is two given threshold values;

(3) initial metric matrix A is established ₀for unit battle array, d be inhomogeneity sample between distance threshold values, s be similar sample between distance threshold values, correlation parameter λ _ijinitial value is 0, correlation parameter ξ _{c (i, j)}when (i, j) belongs to S respectively, initial value is s, correlation parameter ξ _{c (i, j)}when (i, j) belongs to D respectively, initial value is b, γ is the slack variable chosen, and chooses a kind of alternative manner based on Gradient Descent and calculates metric matrix A.

3. according to claim 2 in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults, it is characterized in that: choose a kind of alternative manner based on Gradient Descent and the concrete iterative process that metric matrix A calculates is described below:

(1) a pair constraint (i, j) is selected;

(2) p=(x is calculated _i-x _j) ^ta _k(x _i-x _j);

(3) when (i, j) ∈ S gets δ=1, otherwise δ=-1 is got;

(4) calculate $a=min({\mathrm{\λ}}_{ij},\frac{\mathrm{\δ}}{2}(\frac{1}{p}-\frac{\mathrm{\γ}}{{\mathrm{\ξ}}_{c(i,j)}}\left)\right);$

(5) calculate $\mathrm{\β}=\frac{\mathrm{\δ}a}{(1-\mathrm{\δ}ap)};$

(6) calculate ${\mathrm{\ξ}}_{c(i,j)}=\frac{{\mathrm{\γ\ξ}}_{c(i,j)}}{\mathrm{\γ}+{\mathrm{\δa\ξ}}_{c(i,j)}};$

(7) λ is calculated _ij=λ _ij-a;

(8) A is calculated _k+1=A _k+ β A _k(x _i-x _j) (x _i-x _j) ^ta _k;

Wherein, γ is slack variable; A _krefer to the value of kth time iteration vacuum metrics matrix A;

Choose a pair constraint (i, j) and carry out iteration with regard to repeating step (1) ~ (8), until metric matrix A restrains, then obtain the metric matrix A of objective matrix and generation system.

4. according to claim 3 in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults, it is characterized in that: in step 5, wavelet transformation detailed process is:

(1) wavelet mother function is defined in a kind of function that limited interval and wavelet mother function mean value are zero, obtains corresponding wavelet function, be expressed as follows by flexible and translation wavelet mother function:

${\mathrm{\ψ}}_{a,b}\left(t\right)=\frac{1}{\sqrt{a}}\mathrm{\ψ}\left(\frac{t-b}{a}\right)$

In formula, a and b is constant value, a, b ∈ R, and a>0, a are scale factor, and b is shift factor; ψ is wavelet mother function; T is the time;

(2) wavelet transformation defining x (t) is as follows:

${\mathrm{WT}}_{a,b}=\frac{1}{\sqrt{a}}{\∫}_{R}x\left(t\right){\mathrm{\ψ}}^{*}\left(\frac{t-b}{a}\right)dt$

Wherein, WT _a,bit is wavelet basis function.

5. according to claim 4 in a kind of industrial process based on metric learning and seasonal effect in time series method for diagnosing faults, it is characterized in that: step 6 KNN sorting technique is specially:

(1) KNN algorithm is k-vicinity method, and KNN algorithm primitive rule is the distance calculating Real-Time Monitoring sample and n subclass, find in all training samples with Real-Time Monitoring sample distance M _ak minimum sample; Wherein,

k＝k ₁+k ₂+k ₃+...+k _i+...k _n

(2) i-th subclass schedule of samples in k sample is shown as k _i, i=1,2 ..., c, definition discriminant function is as follows:

g _i(x)＝k _i,i＝1,2,…,c

Real-Time Monitoring sample is marked as type i, if g _i(x)>=g _j(x), i ≠ j.