CN103914064A - Industrial process fault diagnosis method based on multiple classifiers and D-S evidence fusion - Google Patents

Abstract

The invention discloses an industrial process fault diagnosis method based on multiple classifiers and D-S evidence fusion. The method comprises the steps that firstly, independent repeated sampling is conducted according to fault data in the industrial process; secondly, the multiple classifiers are applied to new training data, respective off-line modeling models are obtained, and meanwhile the properties of all the classifiers are represented in the form of a fusion matrix; thirdly, different types of elementary probability valuation functions are calculated according to the D-S evidence theory, decisions of the multiple classifiers are selectively integrated and synthesized according to the similarity index, a combined elementary probability valuation function is obtained, and a final classified diagnosis result is obtained by means of comparison. Compared with other methods in the prior art, the industrial process fault diagnosis method can greatly improve the diagnosis effect of the industrial process, shorten delayed diagnosis time and increase the diagnosis accuracy rate, improves the monitoring performance to a great extent, enhances the comprehension ability and operation confidence of process operators in the process, and is more beneficial to automatic implementation of the industrial process.

Description

Based on the industrial process method for diagnosing faults of multi-categorizer and D-S evidence fusion

Technical field

The invention belongs to industrial process control field, relate in particular to a kind of industrial process method for diagnosing faults based on multi-categorizer and Dempster-Shafer (D-S) evidence fusion.

Background technology

In recent years, the monitoring problem of industrial processes more and more obtains the extensive attention of industry member and academia.On the one hand, actual industrial process is because of its process complexity, and performance variable is many, has the stages such as non-linear, non-Gauss, dynamic, under single hypothesis, uses a certain method, and its monitoring effect has great limitation.On the other hand, if process is not well monitored, and to contingent diagnosing malfunction, work accident likely can occur, the lighter affects the quality of product, and severe one will cause the loss of life and property.Therefore, find better process monitoring method and carry out correctly forecasting and having become one of the study hotspot of industrial processes and problem in the urgent need to address in time.

Traditional Industrial Process Monitoring method, except the method based on mechanism model, adopts Multielement statistical analysis method mostly, such as pca method (PCA) and offset minimum binary method (PLS) etc.In the situation that mechanism model is difficult to obtain, the Multielement statistical analysis method based on data-driven has become the main stream approach of Industrial Process Monitoring.But, all there are some basic assumed conditions in traditional Multielement statistical analysis method, such as the assumed condition of pca method (PCA) is that data are obeyed independent same distribution, and suppose that process obeys linear Gaussian distribution, but real process relative complex, process may be that a part is linear, a part is non-linear or a part of non-Gauss's combination.Impossible and want to find one to be applicable to all-environment method.By contrast, method under multiple different assumed condition is carried out integrated, being information fusion method is processing the advantage that has himself aspect the monitoring of complex industrial process and fault diagnosis, and the present invention adopts the method to substitute original single Multielement statistical analysis method procedure fault is diagnosed.In order to improve the effect of fusion, increasing in sorter diversity, can first carry out repeated sampling pre-service to training data, and utilize certain similarity indices, selectively merge.Traditional monitoring method hypothesis process operation, under single condition, cannot meet the monitoring requirement of actual industrial process.Even the different operating condition of process is carried out respectively to modeling, also cannot reach satisfied monitoring effect.Because when new process data is monitored, need cohesive process knowledge to judge the condition of work of these data, and choose corresponding monitoring model, this has just strengthened the dependence of monitoring method to procedural knowledge greatly, is unfavorable for that the robotization of industrial process is implemented.

Summary of the invention

The object of the invention is to for the deficiencies in the prior art, a kind of industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion is provided.

The object of the invention is to be achieved through the following technical solutions: a kind of industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion, comprises the following steps:

(1) utilize the data of the normal operation of systematic collection process and the training sample set that the modeling of various fault data composition is used: assumed fault classification is C, adds a normal class, and total classification of modeling data is C+1, that is, and X _i=[x ₁; x ₂; ...; x _n], i=1,2 ..., C+1; Wherein X _i∈ R ^{n × m}, R is set of real numbers, R ^{n × m}represent that X meets the Two dimensional Distribution of n × m, the sample number that n is each class, m is process variable number; So complete training sample set is X and X=[X ₁; X ₂; ...; X _c+1], X ∈ R ^{((C+1) * n) × m}, deposit these data in historical data base;

(2) collect the other failure classes data different from training data, as off-line test data, C class, that is: Y altogether _j=[y ₁; y ₂; ...; y _n], j=1,2 ..., C, wherein Y _j∈ R ^{n × m}, and the N sample number that is each class, m is process variable number; So complete test sample book integrates as Y, i.e. Y=[Y ₁; Y ₂; ...; Y _c], Y ∈ R ^{(C*N) × m}, deposit these data in historical data base simultaneously;

(3) from database, call training sample X, adopt independent repeated sampling method to reset processing to each class data matrix, and ensure that reordering rule is consistent, obtain data matrix collection

(4) to data set carry out pre-service and normalization,, in each classification, the average that makes respectively each process variable is zero, and variance is 1, obtains new data matrix collection and is

(5) data set Y is carried out to pre-service and normalization, the average of all kinds of training samples that obtain according to step 4 and variance, make in each class, and the average of each process variable is zero, and variance is 1, obtains new data matrix collection and be

(6) number of selection sort device method is G, comprises without measure of supervision and has measure of supervision, calls different sorters, at training dataset set up different sorter models down, to calculate corresponding T without monitor model ²detection statistics limit with SPE statistic; Calculate corresponding label index to there being monitor model;

(7) in test data set under, utilize different sorter models and parameter thereof, the fusion matrix of each classifier methods of calculated off-line;

(8) according to the index of similarity proposing, calculate the similarity between different classifier methods, for selectivity fusion process is afterwards prepared;

(9) by modeling data and each model parameter, deposit in together in historical data base and real-time data base for subsequent use;

(10) collect new online process data, and it is carried out to pre-service and normalization;

(11) adopt respectively different sorter models to monitor, for without monitor model, set up statistic T 2 and SPE statistic, for there being monitor model, obtain corresponding tag along sort;

(12) by D-S evidence theory, utilize and merge the priori to different faults recall rate in matrix, calculate the compressive classification rate of current sample under all classifier methods, the index of similarity calculating before utilizing, selectively merge, and make last decision-making.

The invention has the beneficial effects as follows: the present invention is by carrying out respectively the analysis and modeling under different classifier methods to each fault data.Then, introduce index of similarity, utilize D-S means of proof to carry out the diagnostic message under distinct methods integrated and comprehensive, obtain last diagnostic result.Compare other current method for diagnosing faults, the present invention not only can improve the monitoring effect of industrial process greatly, reduce the delayed diagnosis time, increase the accuracy of diagnosis, and improve to a great extent the dependence of monitoring method to procedural knowledge, strengthened process operator to the understandability of process and operation confidence, the robotization that is more conducive to industrial process is implemented.

Brief description of the drawings

Fig. 1 be under the inventive method without measure of supervision (PCA) diagnostic graph of the fault data to TE process;

Fig. 2 is the diagnostic graph that has the fault data of measure of supervision (FDA) to TE process under the inventive method;

Fig. 3 is without measure of supervision (PCA) with there is the fusion matrix diagram of measure of supervision (FDA) under the inventive method.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

A kind of industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion of the present invention, the method is for the troubleshooting issue of industrial process, first utilize Distributed Control System (DCS) to collect the data under normal operating conditions, and various already present fault datas, then these data are carried out to unified diverse processing, carry out independent repeated sampling, obtain new training dataset, on this basis, call respectively different classifier methods, set up corresponding sorter model, and unsupervised method is set up to two monitoring and statistics amount T ²with SPE and corresponding statistics limit thereof and SPE _lim, to there being the method for supervision to set up label classification.Call test data set, utilize various sorter models, obtain the fusion matrix that comprises sorter classification performance, then all model parameters are deposited in database for subsequent use.When monitoring with fault diagnosis to new online data, first utilize different sorter monitoring models to monitor it, obtain corresponding monitoring result, utilize similarity indices, select suitable classifier methods, then merging on the basis of matrix, obtaining the state final decision of these data by D-S evidence theory.

A kind of industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion of the present invention, comprises the following steps:

The first step: utilize the data of the normal operation of systematic collection process and the training sample set that the modeling of various fault data composition is used: assumed fault classification is C, adds a normal class, and total classification of modeling data is C+1, that is, and X _i=[x ₁; x ₂; ...; x _n], i=1,2 ..., C+1.Wherein X _i∈ R ^{n × m}, m is process variable number, the sample number that n is each class.So complete training sample set is X and X=[X ₁; X ₂; ...; X _c+1], X ∈ R ^{((C+1) * n) × m}, deposit these data in historical data base.

Second step: collect the other failure classes data different from training data, as off-line test data, C class altogether, that is, and Y _j=[y ₁; y ₂; ...; y _n], j=1,2 ..., C, wherein Y _j∈ R ^{n × m}, and the N sample number that is each class, m is process variable number.As Y so complete test sample book integrates, Y=[Y ₁; Y ₂; ...; Y _c], Y ∈ R ^{(C*N) × m}, deposit these data in historical data base simultaneously.

The 3rd step: call normal data from database, adopt independent repeated sampling method to reset processing to data matrix, obtain data matrix collection

In order to increase the diversity of data, to improve final syncretizing effect, utilize the method for independent repeated sampling, data are processed, under statistical significance, the possibility that the data of large probability are adopted is large, so just can utilize the method to filter the sample of a part of small probability, i.e. a part of garbage.

The 4th step: to data matrix stack carry out pre-service and normalization,, in each class, the average that makes each process variable is zero, and variance is 1, obtains new data matrix collection and is

In historical data base, the process data collecting is carried out to pre-service, reject outlier and obvious coarse error information, in order to make the yardstick of process data can not have influence on the result of monitoring, data to different variablees are normalized respectively, the average that is each variable is zero, and variance is 1.Like this, the data of various process variable are just under identical yardstick, can not have influence on afterwards follow-up monitoring effect.

The 5th step: data set Y is carried out to pre-service and normalization, according to the average of all kinds of training samples in step 4 and variance, make in each class, the average of each process variable is zero, and variance is 1, obtains new data matrix collection and is

The 6th step: call different classifier methods, comprise without measure of supervision and have measure of supervision, at new data matrix collection set up different sorter models down, to construct corresponding T without monitor model ²detection statistics limit with SPE statistic; Construct corresponding label index to there being monitor model;

Because training dataset the information that comprises C+1 kind, therefore, in modeling, will process separately for each failure classes sample set, altogether needs C+1 time, below to process a class fault as example:

A), for without measure of supervision, concrete performing step is as follows:

1) analyze by PCA, can obtain the covariance matrix Σ ∈ R of data matrix ^{n × n}, unitary matrix U ∈ R ^{n × m}, eigenwert form diagonal matrix D ∈ R ^{m × m}as follows:

$\begin{array}{c}\mathrm{\Σ}=\stackrel{=}{X}{\stackrel{=}{X}}^T/(n-1)\\ \mathrm{\Σ}={\mathrm{UDU}}^T\\ D=\mathrm{diag}\left({\mathrm{\λ}}_i\right),i=1,...,m\\ U=[u_1,u_2,...,u_m]\end{array}---\left(15\right)$

Then obtain on its basis matrix of loadings P ∈ R ^{m × k}, residual error matrix of loadings pivot composition t ∈ R ^{n × k}, residual matrix as follows:

$\begin{array}{c}P=[u_1,u_2,...,u_k]\\ \stackrel{\‾}{P}=[u_{k+1},u_{k+2},...,u_m]\\ t=\stackrel{=}{X}P\\ \stackrel{\‾}{C}=\stackrel{\‾}{P}{\stackrel{\‾}{P}}^T\end{array}---\left(16\right)$

Wherein, k is the pivot number of extracting, and mainly utilizes accumulative total variance contribution ratio (>80%) to calculate.Then construct T ²statistic also utilizes F distribution to provide monitoring and statistics limit to residual matrix set up SPE statistic and calculate its corresponding monitoring and statistics limit SPE _lim.

2) analyze by KPCA, utilize radial basis kernel function, raw data is passed through to High Dimensional Mapping, obtain the eigenwert of higher dimensional space, proper vector and score, and utilize accumulative total variance contribution ratio (>80%) to calculate pivot number d, obtain corresponding matrix of loadings, pivot PCA method described above.Same structure T ²statistic also utilizes F distribution to provide monitoring and statistics limit residual matrix is set up SPE statistic and calculated its corresponding monitoring and statistics limit SPE _lim.

3) analyze by ICA, can obtain the independent component matrix S ∈ R of this data matrix ^{r × n}, hybrid matrix A ∈ R ^{m × r}, separation matrix W ∈ R ^{r × m}and residual matrix as follows:

$\begin{array}{c}\stackrel{=}{X}=\mathrm{AS}+\stackrel{=}{E}\\ S=W\stackrel{=}{X}\\ \stackrel{=}{E}=\stackrel{=}{E}-\mathrm{AS}\end{array}---\left(17\right)$

Wherein, r is the independent component number of choosing.Then, structure I ²statistic also utilizes Density Estimator method to provide its corresponding monitoring and statistics limit that is:

$\hat{f}(I^2,H)=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}K\left(H^{-1/2}(I^2-I_i^2)\right).---\left(18\right)$

Wherein, K () is kernel function, is conventionally chosen for gaussian kernel form, and the bandwidth parameter matrix that H is kernel function can easy choice be diagonal angle form, I ²for the I of current sample ²statistics value, be the I of i training sample ²statistics value.Like this, we just can obtain I ²the probability density distribution information of statistic, thus can ask for easily its statistics limit under confidence degree value.For residual matrix the monitoring and statistics limit of structure SPE statistic;

On the basis of previous step to residual matrix set up SPE statistic and calculate its corresponding monitoring and statistics limit SPE _lim, that is:

${\mathrm{SPE}}_{\lim }=\stackrel{=}{E}{\stackrel{=}{E}}^T---\left(19\right)$

Wherein, SPE _limobeying parameter is the χ of g and h ²distribute,

$\begin{array}{c}g\·h=\mathrm{mean}\left(\mathrm{SPE}\right)\\ {2g}^{2}h=\mathrm{var}\left(\mathrm{SPE}\right)\end{array}---\left(20\right)$

Therefore, the monitoring and statistics limit of SPE statistic also can be obtained easily,

B) for there being measure of supervision, concrete performing step is as follows:

1) by Fei Sheer criterion method, find out all kinds of between most suitable projecting direction, and determine the position of the central point of each class;

2) by K-near neighbor method, set 5 Neighbor Points, add class label to modeling data;

3) by neural net method, the two-layer BP network that selection comprises three hidden nodes, hidden layer is selected tansig function, and output layer is selected purelin function, training network model.

The 7th step is for test data set call different classifier methods, calculated off-line G kind merges matrix.The form that merges matrix is as follows:

${\mathrm{CM}}^g=\left[\begin{array}{ccccccc}{N}_{11}^g& N_{12}^g& \·& \·& \·& N_{1C}^g& N_{1(C+1)}^g\\ N_{21}^g& N_{22}^g& \·& \·& \·& N_{2C}^g& N_{2(C+1)}^g\\ \·& \·& \·& & & \·& \·\\ \·& \·& & \·& & \·& \·\\ \·& \·& & & \·& \·& \·\\ N_{C1}^g& N_{C2}^g& \·& \·& \·& N_{\mathrm{CC}}^g& N_{C(C+1)}^g\end{array}\right]g=\mathrm{1,2},...,G---\left(21\right)$

Wherein, CM ^grepresent the fusion matrix of g sorter, and the G number that is sorter, row represents C class fault, and front C row represent different faults, and last row representative is normal, with for example, its meaning is the sample number that the sample that belongs to Equations of The Second Kind is classified device and is divided into the first kind.Concrete performing step below.

A) call unsupervised classifier methods, concrete performing step is as follows:

1) PCA method, according to all kinds of parameters that obtain in above-mentioned steps, calculates the T of 6 class fault samples under 6 kinds of classifier methods ²with SPE statistics value, utilize the index of associating PCA and two statistics of SPE to do failure reconfiguration, be called associating discriminant index λ _i, as follows:

${\mathrm{\λ}}_i=(1-{\mathrm{\α}}_i)\×T^2/T_{\lim .i}^2+{\mathrm{\α}}_i\×\mathrm{SPE}/{\mathrm{SPE}}_{\lim .i},i=\mathrm{1,2},...,C+1---\left(22\right)$

Wherein, α _ifor score t _icumulative proportion in ANOVA, T ²with the SPE statistical value that is test sample book, and SPE _lim.ibe respectively the detection statistics limit that step (6) modeling obtains, get reckling in C+1 the testing classification as certain sample;

2) KPCA method, with above-mentioned PCA method;

3) ICA method, similar with above-mentioned PCA method, but T ²statistic has changed I into ²statistic, failure reconfiguration method is afterwards similar.

B) call unsupervised classifier methods, concrete performing step is as follows:

1) for Fei Sheer method of discrimination, calculate the Euclidean distance between test sample book and central point of all categories, label corresponding to that class of distance minimum is as final label;

2) for K-near neighbor method, calculate the distance between test sample book and known class label, obtain final label;

3) for neural net method, the network model that utilizes training to obtain, the output label of calculating test sample book.The 8th step, according to the index of similarity proposing, is calculated the similarity between different classifier methods, for selectivity fusion process is afterwards prepared.In order to weigh the similarity between different classifier methods, the present invention proposes a kind of index of similarity computing method based on merging matrix, as follows:

${\mathrm{corr}}_{\mathrm{ij}}=\frac{\mathrm{cov}(\mathrm{ci}\_m,\mathrm{cj}\_m)}{\sqrt{D(\mathrm{ci}\_m)}\sqrt{D(\mathrm{cj}\_m)}}---\left(23\right)$

Wherein, ci_m and cj_m representative be the fusion matrix of different classifier methods, D (ci_m) and D (cj_m) are respectively the variance yields corresponding to fusion matrix of different classifier methods, corr _ijwhat represent is two linear dependence sex index between sorter.

The 9th step deposits in historical data base and real-time data base for subsequent use by modeling data and each model parameter;

The tenth step is collected new process data, and it is carried out to pre-service and normalization;

For the data sample of newly collecting in process, except it is carried out pre-service, the model parameter while adopting modeling is in addition normalized this data point, deducts modeling average and divided by modeling standard deviation.

The 11 step adopts respectively different sorter models to monitor it, sets up statistic T ²with SPE and label, each method can obtain a decision-making about normal or fault so

A), for unsupervised method, set up corresponding monitoring and statistics amount as follows:

1) analyze for PCA

$\begin{array}{c}{t}_{\mathrm{new}}={\stackrel{=}{X}}_{\mathrm{new}}P\\ {\mathrm{SPE}}_{\mathrm{new}}={\left|\right|\stackrel{\‾}{C}{\stackrel{=}{X}}_{\mathrm{new}}\left|\right|}^2\\ {T_{\mathrm{new}}}^2={\left|\right|{D_r}^{-1/2}{t}_{\mathrm{new}}\left|\right|}^2=\left|\right|{D_r}^{-1/2}{P}^T{\stackrel{=}{X}}_{\mathrm{new}}\left|\right|\end{array}---\left(24\right)$

Wherein for the online new samples after normalization, t _newfor the pivot of new samples, for residual matrix, P is matrix of loadings, SPE _newfor the SPE statistics value of online new samples, || g|| represents 2-norm, T _new ²for the T of new samples ²statistics value, T is transpose of a matrix.

2) analyze for KPCA, with above-mentioned PCA process.

3) analyze for ICA

$\begin{array}{c}{s}_{\mathrm{new}}=W{\stackrel{=}{x}}_{\mathrm{new}}\\ {\stackrel{=}{e}}_{\mathrm{new}}={\stackrel{=}{x}}_{\mathrm{new}}\\ I_{\mathrm{new}}^2=s_{\mathrm{new}}^T{s}_{\mathrm{new}}\end{array},-{\mathrm{As}}_{\mathrm{new}}---\left(25\right)$

Wherein, for the new data after normalization, s _newfor the independent component vector extracting based on new data, for the I of new data ²statistic, continues for residual vector setting up SPE statistic is SPE _new:

${\mathrm{SPE}}_{\mathrm{new}}={\stackrel{=}{e}}_{\mathrm{new}}{\stackrel{=}{e}}_{\mathrm{new}}^T---\left(26\right)$

B), for the method that has supervision, obtain corresponding class label:

1) for Fei Sheer criterion method, calculate the Euclidean distance between new samples and central point of all categories, label corresponding to that classification of distance minimum is as final label;

2) for K-near neighbor method, calculate the distance method between new samples and known class label, obtain final label; 3) for neural net method, the network model that utilizes training to obtain, the output label of calculating new samples.

The 9th step, by D-S evidence theory, is utilized the priori of each method to different faults recall rate, calculates the comprehensive recall rate of current Monitoring Data under all classifier methods, and makes last decision-making

A) first call different classifier methods, calculate corresponding basic probability assignment function m _g(C _i), as follows:

$m_g\left(C_i\right)=\frac{N_{\mathrm{ij}}^g}{\underset{i=1}{\overset{G}{\mathrm{\Σ}}}{N}_{\mathrm{ij}}^g},g=\mathrm{1,2},...,G---\left(27\right)$

Wherein, refer to the element of the capable j row of i in the fusion matrix of g classifier methods, m _g(C _i) refer to g sorter sample is assigned to C _ithe probable value of class, is also basic probability assignment function value, and G is the number of selection sort device.

Fault category in assumed fault storehouse has C=1, and 2 ... L, taking PCA method as example, have:

$\begin{array}{c}{m}_{\mathrm{PCA}}\left(1\right)=p_{\mathrm{PCA}}\left(1\right);\\ m_{\mathrm{PCA}}\left(2\right)=p_{\mathrm{PCA}}\left(2\right);\\ \·\·\·\\ m_{\mathrm{PCA}}\left(L\right)=p_{\mathrm{PCA}}\left(L\right);\end{array}---\left(28\right)$

Wherein, m _pCA(1) that representative is the probable value p that PCA method is divided into sample the first kind _pCA(1), same m _pCA(L) that representative is the probable value p that PCA method is divided into sample L class _pCA(L).

For other classifier methods, can obtain equally corresponding basic probability assignment function value, as follows:

$\begin{array}{c}{m}_{\mathrm{method}}\left(1\right)=p_{\mathrm{method}}\left(1\right);\\ m_{\mathrm{method}}\left(2\right)=p_{\mathrm{method}}\left(2\right);\\ \·\·\·\\ m_{\mathrm{method}}\left(L\right)=p_{\mathrm{method}}\left(L\right);\end{array}---\left(29\right)$

Wherein m _method(L) that representative is the probable value p that classifier methods menthod is divided into sample L class _method(L).

B) in same sampling instant, call the output under six kinds of classifier methods of selection, according to the index of similarity calculating before, selectively selection sort device method, picking out corresponding testing result is the probable value that fault occurs, utilize following D-S fusion rule, obtain last basic probability assignment function:

Wherein ⊕ represent orthogonal and, definition is as formula (29) as shown in, set A, B, C represent respectively different failure classes set, and A is the common factor of set B and C, η representative be that common factor is the joint probability assignment function value of empty set, m ₁and m ₂represent respectively first and second sorter, m _1.2(A) be the probable values after the associating of two kinds of classifier methods, and result after normalization.

$\begin{array}{c}{m}_{\mathrm{1,2},...,k}=m_1\⊕m_2\⊕...\⊕m_k\\ =(((m_1\⊕m_2)\⊕m_2)\⊕...\⊕m_k)\\ =((m_{\mathrm{1,2}}\⊕m_3)\⊕...\⊕m_k)\\ \·\·\·\end{array}---\left(32\right)$

M _{1,2 ..., K}what represent is K the probable value after Classifier combination, and it is the associating by first doing two methods, then with the 3rd method associating, by that analogy.

C) for the basic probability assignment function value after merging, select larger value, as last result, Final (A _i):

$\mathrm{Final}\left(A_i\right)=\underset{i}{\arg \max }\left[m_{\mathrm{1,2},...G}\left(A_i\right)\right]i=\mathrm{1,2},...,C+1---\left(33\right)$

Wherein m _{1,2 ... G}(A _i) what represent is joint probability assignment function total under G kind classifier methods.

Below in conjunction with the example of a concrete industrial process, validity of the present invention is described.The data of this process are from the experiment of U.S. TE (Tennessee Eastman---Tennessee-Yi Siman) chemical process, and prototype is an actual process flow process of Eastman chemical company.Whole TE process comprises 41 measurands and 12 performance variables (control variable), and wherein 41 measurands comprise 22 continuous coverage variablees and 19 composition measurement values, and they are sampled once for every 3 minutes.Comprising 21 batches of fault datas.In these faults, 16 is that oneself knows, 5 is 1301 of the unknown.Fault 1～7 is relevant with the step variation of process variable, as the variation of the temperature in of chilled water or charging composition.Fault 8～12 increases and matters a lot with the changeability of some process variable.Fault 13 is the slow drifts in reaction kinetics, and fault 14,15 is relevant with sticking valve with 21.Fault 16～20th is unknown.For this process is monitored, choose altogether 16 process variable, as shown in table 1:

Table 1: monitored variable explanation

Sequence number	Variable	Sequence number	Variable
				1	A charging (stream 1)	9	Separation of products actuator temperature
2	D charging (stream 2)	10	Product separator pressure
				3	E charging (stream 3)	11	Low discharge at the bottom of product separator tower (stream 10)
4	Combined feed (stream 4)	12	Stripper pressure
				5	Recirculating mass (stream 8)	13	Stripper temperature
6	Reactor feed speed (stream 6)	14	Stripper flow
				7	Temperature of reactor	15	Reactor cooling water outlet temperature
8	Mass rate of emission (stream 9)	16	Separation vessel cooling water outlet temperature

Next in conjunction with this detailed process, implementation step of the present invention is at length set forth:

1. gather normal processes data, gather various fault datas simultaneously, carry out pre-service, normalization

2. for training data, call different classifier methods, set up respectively confidence limit and the label of different sorter models definite corresponding statistic

Respectively to new data matrix carry out model foundation, taking the data modeling of a class as example:

1) carry out PCA analysis and modeling, choose 6 pivot compositions, obtain detailed pca model.Then construct T ²statistic also distributes and determines its corresponding monitoring and statistics limit with F.In like manner, utilize Ka Er side to distribute and can determine the monitoring confidence limit of SPE statistic.Here the degree of confidence that, we choose two statistics is 99%.

2) carry out KPCA analysis and modeling, choose 5 pivot compositions, obtain detailed KPCA model.Then construct T ²statistic also distributes and determines its corresponding monitoring and statistics limit with F.In like manner, utilize Ka Er side to distribute and can determine the monitoring confidence limit of SPE statistic.Here the degree of confidence that, we choose two statistics is 99%.

3) carry out ICA analysis and modeling, choose 4 independent components, obtain detailed ICA model parameter information, i.e. independent component information S ∈ R ^{4 × 960}, hybrid matrix A ∈ R ^{16 × 4}, separation matrix W ∈ R ^{4 × 16}and residual matrix then construct I ²statistic is also determined its corresponding monitoring and statistics limit by Density Estimator method.In like manner, can determine the confidence limit of SPE statistic.Here the degree of confidence that, we choose two statistics is 99%.

4) carry out the modeling of Fei Sheer criterion, find all kinds of optimal projecting directions, add class label to modeling data, label is 1 to 7, altogether 7 natural numbers.

5) carry out K-neighbour modeling, set 5 Neighbor Points, add class label to modeling data, label is 1 to 7, altogether 7 natural numbers.

6) carry out neural net model establishing, select to have the two-layer BP network of three hidden nodes, multicategory classification problem is converted into two class classification, number of tags is 0 and 1.

3. call off-line test data, according to the sorter model of above-mentioned foundation, merged accordingly matrix.

4. obtain current online data information, and it is carried out to pre-service and normalization

In order to test the validity of new method, respectively normal sample and fault sample are tested.Choose at random a certain process data, and utilize the normalized parameter under each classifier methods to process it.Choose a kind of typical fault and test, equally it is normalized.

5. on-line fault diagnosis

First training data is carried out to modeling, select fault 1,2,5,6,8,12 6 kind, unsupervised method and have diagnostic result (part) that the method for supervision obtains respectively as depicted in figs. 1 and 2.As can be seen from the figure, single classifier methods is made good diagnosis to fault as PCA and Fei Sheer criterion.Then, call test sample book, obtain merging matrix, select equally partial results as PCA and Fei Sheer criterion, as shown in Figure 3, illustrate that single classifier methods is higher to the recall rate of some fault, responsive not to some; According to the index of similarity computing formula proposing, calculate the diversity between six kinds of classifier methods selecting, as shown in table 2, it is to provide foundation for the process of last fusion.

Table 2: the indicator gauge of relative index under various classifier methods

Finally, new samples is diagnosed, new method and the contrast of the diagnostic result of single classifier methods are as shown in table 3:

Table 3: the diagnosis effect contrast table of the inventive method and single classifier methods

Can obviously find out, new method has successfully detected the fault of process and has diagnosed out accurately very much corresponding fault, had less sample delay and higher accuracy rate of diagnosis simultaneously.

Above-described embodiment is used for the present invention that explains, instead of limits the invention, and in the protection domain of spirit of the present invention and claim, any amendment and change that the present invention is made, all fall into protection scope of the present invention.

Claims (7)

1. the industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion, is characterized in that, comprises the following steps:

(1) utilize the data of the normal operation of systematic collection process and the training sample set that the modeling of various fault data composition is used: assumed fault classification is C, adds a normal class, and total classification of modeling data is C+1, that is, and X _i=[x ₁; x ₂; ...; x _n], i=1,2 ..., C+1; Wherein X _i∈ R ^{n × m}, R is set of real numbers, R ^{n × m}represent that X meets the Two dimensional Distribution of n × m, the sample number that n is each class, m is process variable number; So complete training sample set is X and X=[X ₁; X ₂; ...; X _c+1], X ∈ R ^{((C+1) * n) × m}, deposit these data in historical data base;

(2) collect the other failure classes data different from training data, as off-line test data, C class, that is: Y altogether _j=[y ₁; y ₂; ...; y _n], j=1,2 ..., C, wherein Y _j∈ R ^{n × m}, and the N sample number that is each class, m is process variable number; So complete test sample book integrates as Y, i.e. Y=[Y ₁; Y ₂; ...; Y _c], Y ∈ R ^{(C*N) × m}, deposit these data in historical data base simultaneously;

(3) from database, call training sample X, adopt independent repeated sampling method to reset processing to each class data matrix, and ensure that reordering rule is consistent, obtain data matrix collection

(4) to data set carry out pre-service and normalization,, in each classification, the average that makes respectively each process variable is zero, and variance is 1, obtains new data matrix collection and is

(5) data set Y is carried out to pre-service and normalization, the average of all kinds of training samples that obtain according to step 4 and variance, make in each class, and the average of each process variable is zero, and variance is 1, obtains new data matrix collection and be

(6) number of selection sort device method is G, comprises without measure of supervision and has measure of supervision, calls different sorters, at training dataset set up different sorter models down, to calculate corresponding T without monitor model ²detection statistics limit with SPE statistic; Calculate corresponding label index to there being monitor model;

(7) in test data set under, utilize different sorter models and parameter thereof, the fusion matrix of each classifier methods of calculated off-line;

(8) according to the index of similarity proposing, calculate the similarity between different classifier methods, for selectivity fusion process is afterwards prepared;

(9) by modeling data and each model parameter, deposit in together in historical data base and real-time data base for subsequent use;

(10) collect new online process data, and it is carried out to pre-service and normalization;

(11) adopt respectively different sorter models to monitor, for without monitor model, set up statistic T 2 and SPE statistic, for there being monitor model, obtain corresponding tag along sort;

(12) by D-S evidence theory, utilize and merge the priori to different faults recall rate in matrix, calculate the compressive classification rate of current sample under all classifier methods, the index of similarity calculating before utilizing, selectively merge, and make last decision-making.

2. the industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion according to claim 1, it is characterized in that, described step 3 is specially: to equating with number of samples, the natural number array that is n is carried out random repeated sampling, obtain a new natural number array, according to this array, each categorical data in training sample set X is rearranged, reconstitute new data matrix

3. the industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion according to claim 1, it is characterized in that, described step 6 is specially: according to the characteristics of objects of selecting, there is linearity in order can to process simultaneously, non-linear, non-Gauss's process, selecting G is 6, and different multi-categorizer methods specifically comprise: without measure of supervision: pivot analysis (PCA), core pivot element analysis (KPCA), Independent component analysis (ICA); There is measure of supervision: Fei Sheer criterion (FDA), k-neighbour (KNN), neural network (BP-ANN);

(a), for unsupervised method, specific implementation step is as follows:

(1) analyze by PCA, obtain the covariance matrix Σ ∈ R of data matrix ^{n × n}, unitary matrix U ∈ R ^{n × m}, eigenwert form diagonal matrix D ∈ R ^{m × m}as follows:

$\begin{array}{c}\mathrm{\Σ}=\stackrel{=}{X}{\stackrel{=}{X}}^T/(n-1)\\ \mathrm{\Σ}={\mathrm{UDU}}^T\\ D=\mathrm{diag}\left({\mathrm{\λ}}_i\right),i=1,...,m\\ U=[u_1,u_2,...,u_m]\end{array}---\left(1\right)$

Wherein, represent new data matrix collection, Σ represents covariance matrix, and U represents unitary matrix, and n represents number of training, and m is variable number, and T is transpose of a matrix, D representation feature value λ _ithe diagonal matrix forming, and its diagonal element is according to descending tactic, diag (g) represents the amount in bracket press diagonal line arrangement, u _mrepresent m column vector that forms U;

Then obtain on its basis matrix of loadings P ∈ R ^{m × k}, residual error matrix of loadings pivot composition t ∈ R ^{n × k}, residual matrix as follows:

$\begin{array}{c}P=[u_1,u_2,...,u_k]\\ \stackrel{\‾}{P}=[u_{k+1},u_{k+2},...,u_m]\\ t=\stackrel{=}{X}P\\ \stackrel{\‾}{C}=\stackrel{\‾}{P}{\stackrel{\‾}{P}}^T\end{array}---\left(2\right)$

Wherein, k is the pivot number of extracting, and mainly utilizes accumulative total variance contribution ratio (>80%) to calculate, and meanwhile, then constructs T ²statistic also utilizes F distribution to provide detection statistics limit to residual matrix set up SPE statistic and calculate its corresponding detection statistics limit SPE _lim;

(2) analyze by KPCA, utilize radial basis kernel function, raw data is passed through to High Dimensional Mapping, obtain the eigenwert of higher dimensional space, proper vector and score, and utilize accumulative total variance contribution ratio (>80%) to calculate pivot number k, obtain corresponding matrix of loadings, pivot PCA method described above;

Same structure T ²statistic also utilizes F distribution to provide monitoring and statistics limit residual matrix is set up SPE statistic and calculated its corresponding monitoring and statistics limit SPE _lim;

(3) analyze by ICA, can obtain the independent component matrix S ∈ R of this data matrix ^{r × n}, hybrid matrix A ∈ R ^{m × r}, separation matrix W ∈ R ^{r × m}and residual matrix as follows:

$\begin{array}{c}\stackrel{=}{X}=\mathrm{AS}+\stackrel{=}{E}\\ S=W\stackrel{=}{X}\\ \stackrel{=}{E}=\stackrel{=}{E}-\mathrm{AS}\end{array}---\left(3\right)$

Wherein, r is the independent component number of choosing; Then, structure I ²statistic also utilizes Density Estimator method to provide its corresponding monitoring and statistics limit to residual matrix set up SPE statistic and calculate its corresponding monitoring and statistics limit SPE _lim;

(b) for there being measure of supervision, specific implementation step is as follows:

(1) by Fei Sheer criterion method, find out all kinds of between most suitable projecting direction, and determine the position of the central point of each class;

(2) by K-near neighbor method, set 5 Neighbor Points, add class label to modeling data;

(3) by neural net method, the two-layer BP network that selection comprises three hidden nodes, hidden layer is selected tansig function, and output layer is selected purelin function, training network model.

4. the industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion according to claim 1, is characterized in that, described step 7 is specially: for test data set

(a) call unsupervised classifier methods, concrete performing step is as follows:

(1) PCA method, according to all kinds of parameters that obtain in above-mentioned steps, calculates the T of 6 class fault samples under 6 kinds of classifier methods ²with SPE statistics value, utilize the index of combining these two statistics to do failure reconfiguration, that is, and associating discriminant index λ _i, as follows:

${\mathrm{\λ}}_i=(1-{\mathrm{\α}}_i)\×T^2/T_{\lim .i}^2+{\mathrm{\α}}_i\×\mathrm{SPE}/{\mathrm{SPE}}_{\lim .i},i=\mathrm{1,2},...,C+1---\left(4\right)$

Wherein, α _ifor score t _icumulative proportion in ANOVA, i is different classifier methods, T ²with the SPE statistical value that is test sample book, and SPE _lim.ibe respectively the detection statistics limit that step 6 modeling obtains, get reckling in C+1 the testing classification as certain sample, finally obtain the matrix of a C* (C+1);

(2) KPCA method, with above-mentioned PCA method;

(3) ICA method, similar with above-mentioned PCA method, but T ²statistic has changed I into ²statistic, failure reconfiguration method is afterwards similarly, obtains equally the matrix of a C* (C+1);

(b) call unsupervised classifier methods, concrete performing step is as follows:

(1) for Fei Sheer method of discrimination, calculate the Euclidean distance between test sample book and all kinds of central point, label corresponding to that classification of distance minimum, as final label, obtains the matrix of a C* (C+1) equally;

(2) for K-near neighbor method, calculate the distance method between test sample book and known class label, obtain final label, obtain equally the matrix of a C* (C+1);

(3) for neural net method, utilize and train the network model obtaining, calculate the output label of test sample book, obtain equally the matrix of a C* (C+1).

5. the industrial process method for diagnosing faults based on multi-categorizer and D-S evidence according to claim 1, it is characterized in that, described step 8 is specially: in order to weigh the similarity between different classifier methods, adopt the index of similarity computing method based on merging matrix, as follows:

${\mathrm{corr}}_{\mathrm{ij}}=\frac{\mathrm{cov}(\mathrm{ci}\_m,\mathrm{cj}\_m)}{\sqrt{D(\mathrm{ci}\_m)}\sqrt{D(\mathrm{cj}\_m)}}---\left(5\right)$

Wherein, ci_m and cj_m representative be the fusion matrix of different classifier methods, D (ci_m) and D (cj_m) are respectively the variance yields corresponding to fusion matrix of different classifier methods, corr _ijwhat represent is two linear dependence sex index between sorter.

6. the industrial process method for diagnosing faults based on multi-categorizer and D-S evidence according to claim 1, is characterized in that, described step 11 is specially: for the new data after normalization adopt respectively different models to monitor it;

(a), for unsupervised method, set up corresponding monitoring and statistics amount as follows:

(1) analyze for PCA

$\begin{array}{c}{t}_{\mathrm{new}}={\stackrel{=}{X}}_{\mathrm{new}}P\\ {\mathrm{SPE}}_{\mathrm{new}}={\left|\right|\stackrel{\‾}{C}{\stackrel{=}{X}}_{\mathrm{new}}\left|\right|}^2\\ {T_{\mathrm{new}}}^2={\left|\right|{D_r}^{-1/2}{t}_{\mathrm{new}}\left|\right|}^2=\left|\right|{D_r}^{-1/2}{P}^T{\stackrel{=}{X}}_{\mathrm{new}}\left|\right|\end{array}---\left(6\right)$

Wherein, for the new data after normalization, t _newfor the pivot of new data, for residual matrix, P is matrix of loadings, SPE _newfor the SPE statistics value of new data, || g|| represents 2-norm, D _rrepresent the diagonal matrix being formed by a front r eigenwert, T _new ²for the T of new data ²statistics value, T is transpose of a matrix;

(2) analyze for KPCA, with above-mentioned PCA process;

(3) analyze for ICA:

$\begin{array}{c}{s}_{\mathrm{new}}=W{\stackrel{=}{x}}_{\mathrm{new}}\\ {\stackrel{=}{e}}_{\mathrm{new}}={\stackrel{=}{x}}_{\mathrm{new}}\\ I_{\mathrm{new}}^2=s_{\mathrm{new}}^T{s}_{\mathrm{new}}\end{array},-{\mathrm{As}}_{\mathrm{new}}---\left(7\right)$

Wherein, for the new data after normalization, s _newfor the independent component vector extracting based on new data, for the I of new data ²statistic, continues for residual vector setting up SPE statistic is SPE _new:

${\mathrm{SPE}}_{\mathrm{new}}={\stackrel{=}{e}}_{\mathrm{new}}{\stackrel{=}{e}}_{\mathrm{new}}^T---\left(8\right)$

(b), for the method that has supervision, obtain corresponding class label:

(1) for Fei Sheer criterion method, calculate the Euclidean distance between new samples and central point of all categories, label corresponding to that classification of distance minimum is as final label;

(2) for K-near neighbor method, calculate the distance method between new samples and known class label, obtain final label;

(3) for neural net method, the network model that utilizes training to obtain, the output label of calculating new samples.

7. the industrial process method for diagnosing faults based on multi-categorizer and D-S evidence fusion according to claim 1, is characterized in that, described step 12 is specially:

(a) first call different classifier methods, calculate corresponding basic probability assignment function, as follows:

$m_g\left(C_i\right)=\frac{N_{\mathrm{ij}}^g}{\underset{i=1}{\overset{G}{\mathrm{\Σ}}}{N}_{\mathrm{ij}}^g},g=\mathrm{1,2},...,G---\left(9\right)$

Wherein, refer to the element of the capable j row of i in the fusion matrix of g classifier methods, m _g(C _i) refer to g sorter sample assigned to the probable value of Ci class, be also basic probability assignment function value, G is the number of selection sort device;

Fault category in assumed fault storehouse has C=1, and 2 ... L kind, for different method methods, has:

$\begin{array}{c}{m}_{\mathrm{method}}\left(1\right)=p_{\mathrm{method}}\left(1\right);\\ m_{\mathrm{method}}\left(2\right)=p_{\mathrm{method}}\left(2\right);\\ \·\·\·\\ m_{\mathrm{method}}\left(L\right)=p_{\mathrm{method}}\left(L\right);\end{array}---\left(10\right)$

Wherein, m _method(1) that representative is the probable value p that method method is assigned to sample the first kind _method(1), same m _method(L) that representative is the probable value p that method method is assigned to sample L class _method(L);

(b) in same sampling instant, call the output under six kinds of classifier methods of selection, according to the index of similarity calculating before, selectively selection sort device method, picking out corresponding testing result is the probable value that fault occurs, utilize following D-S fusion rule, obtain last basic probability assignment function:

Wherein, ⊕ represent orthogonal and, as shown in Equation (11), wherein set A, B, C represent respectively different failure classes set in definition, and A is the common factor of set B and C, η representative be that to occur simultaneously be the joint probability assignment function value of empty set, m ₁and m ₂represent respectively first and second sorter, m _1.2(A) be the probable values after the associating of two kinds of classifier methods, and result after normalization;

$\begin{array}{c}{m}_{\mathrm{1,2},...,k}=m_1\⊕m_2\⊕...\⊕m_k\\ =(((m_1\⊕m_2)\⊕m_2)\⊕...\⊕m_k)\\ =((m_{\mathrm{1,2}}\⊕m_3)\⊕...\⊕m_k)\\ \·\·\·\end{array}---\left(13\right)$

M _{1,2 ..., K}what represent is K the probable value after Classifier combination, and it is the associating by first doing two methods, then with the 3rd method associating, by that analogy; m _krepresent K sorter;

C) for the basic probability assignment function value after merging, select larger value, as last result Final (A _i):

$\mathrm{Final}\left(A_i\right)=\underset{i}{\arg \max }\left[m_{\mathrm{1,2},...G}\left(A_i\right)\right]i=\mathrm{1,2},...,C+1---\left(14\right)$

Wherein, m _{1,2 ... G}(A _i) what represent is joint probability assignment function total under G kind classifier methods.