CN104731083B - A kind of industrial method for diagnosing faults and application based on self-adaptive feature extraction - Google Patents

Abstract

The invention discloses a kind of industrial method for diagnosing faults based on self-adaptive feature extraction and application, belong to industrial process monitoring and diagnostic techniques field.First, data characteristic analysis is carried out to industrial gathered data, suitable feature extracting method is selected for different data characteristics.Secondly, failure modes are realized using Hidden Markov model methods.Diversity of the present invention for industrial data, such as linear, non-linear and non-Gaussian system feature, using self-adaptive feature extraction method, at utmost retain the purpose of effective information to reach, and using the extremely strong dynamic process time series modeling ability of HMM and time series pattern classification capacity, failure to industrial process is classified, therefore compared with other existing methods, due to having taken into full account data characteristic, the inventive method has accuracy rate higher for industrial fault diagnosis.

Description

A kind of industrial method for diagnosing faults and application based on self-adaptive feature extraction

Technical field

It is more particularly to a kind of based on self-adaptive feature extraction the invention belongs to industrial process monitoring and fault diagnosis field Industrial Fault Classification.

Background technology

With the growth of industrial process complexity, the Usefulness Pair of Industrial Process Monitoring and diagnosis is in guarantee production process peace Entirely, maintaining product quality and optimization product interests becomes to become more and more important, and feature extraction is important step in fault diagnosis.

Traditional feature extracting method has a lot, such as Principal Component Analysis (PCA), Kernel Principal Component Analysis (KPCA) and Independent Component Analysis (ICA).PCA can Effectively to process the linear relationship between variable, but it can not detect the nonlinear organization between variable.But in industrial process, become Non-linear relation between amount is generally existing.For solving this problem, KPCA is suggested.But not having one at present has The mode of effect selects the kernel function and parameter of KPCA.If with the treatment linear process of KPCA blindnesses, not only kernel function and ginseng Several selections is complicated and time-consuming, and it also greatly increases the complexity of calculating.And if data are non-gaussian, PCA The non-Gaussian feature of data cannot be extracted, therefore ICA is suggested.

For process monitoring and troubleshooting issue, traditional method for diagnosing faults carries out spy using single method mostly Levy extraction.Due to the complexity of modern industry data, single feature extracting method possibly cannot obtain good effect, such as adopt Feature extraction is carried out to data with PCA, although can effectively extract the linear character of data, but if data are non-linear junctions Structure, PCA cannot effectively extract the nonlinear organization between data.And in industrial process field, the non-linear and non-Gaussian system of data It is very universal.It is incorrect due to feature extraction, fault diagnosis may be caused to make a mistake.

The content of the invention

The purpose of the present invention is in view of the shortcomings of the prior art, there is provided a kind of industrial failure based on self-adaptive feature extraction Diagnostic method and application, before feature extraction is carried out, first the characteristic to data is analyzed, and is selected for different data characteristics Different feature extracting methods, the purpose of effective information is at utmost retained to reach.Therefore, this method is solving process industrial With very big advantage when there is non-gaussian and the complex data characteristic such as non-linear, it is possible to achieve effectively fault diagnosis.

The step of a kind of industrial Fault Classification based on self-adaptive feature extraction, the method, is as follows：

Step one：Off-line modeling, to the off-line data of industrial process collection, carries out data characteristic analysis, for different Data characteristic from suitable feature extracting method, using the feature after extraction as HMM (HMM) observation sequence Row, train HMM；

Step 2：Inline diagnosis, the data to online acquisition carry out data characteristic analysis, are selected for different data characteristics With suitable feature extracting method, after obtaining corresponding observation sequence, the HMM model that selection is most matched, so that failure judgement class Type.

Off-line modeling process described in step one is as follows：

6) X=[x are constituted using the Monitoring Data of industrial process collection₁,x₂,…,x_n]^T∈R^n×m, wherein m represent monitoring become The number of amount, n represents number of samples, x_i∈R^m, i=1 ..., n represents i-th sample；

2) data for gathering are carried out with data characteristic analysis, including nonlinear metric and non-Gaussian system measurement；

2.1) nonlinear metric

2.1.1 raw data set X ∈ R) are given^n×m, l region is classified as, each area sample number is

2.1.2 the confidential interval of each element in the correlation matrix of raw data set, the Correlation Moment of raw data set) are calculated Matrix representation is as follows：

The confidence limit of each element is calculated using the average and variance of each variable, is expressed as following form

(2)

Pca model is equal to according to residual error and abandons characteristic value sum

Exact boundary can be tried to achieve by following formula：

Exact boundary is

By comparing residual error and exact boundary in regional, it may be determined that relationship between variables, if all of residual error all Fall within the exact boundary estimated, then illustrate between the data variable it is linear relationship, when at least one residual error falls in essence When beyond true border, it is non-linear relation to be judged as between variable；

2.2) non-Gaussian system measurement

Using normal probability paper principle, the non-Gaussian system of inspection data, normal probability paper be by perpendicular to transverse axis, the longitudinal axis The ruled paper that some straight lines are constituted, transverse axis is, by evenly divided, to represent observation x, such as following formula：

The longitudinal axis represents normal distribution cumulative probability value, and the longitudinal axis is by non-evenly divided, the purpose is to make Normal Distribution Figure of the observation on normal probability paper be in straight line, will substantially be bent for Non-Gaussian Distribution；

3) data characteristic according to inspection carries out feature extraction from appropriate method, if being linear to use PCA between variable Carry out feature extraction, if be between variable it is nonlinear carry out feature extraction with KPCA, if between variable being being entered with ICA for non-Gaussian system Row feature extraction；

4) using the feature after extraction as the observation sequence O of HMM, HMM is trained, is obtained Its parameter lambda=(A, B, Π)；

Wherein, A is hidden state transition probability matrix, describes the transition probability between each state in HMM model, is seen Following formula (7)~(8),

In formula,It is abbreviated as a_i,j, representing in t, state is S_iUnder conditions of, it is S in t+1 moment states_jIt is general Rate；

B is observer state transition probability matrix, describes probability distribution of the Observable state under each hidden state, is seen Following formula (9)~(10),

In formula,It is abbreviated as b_j,k, representing in t, hidden state is S_jUnder the conditions of, observer state is V_kProbability；

Π is initial state probabilities matrix.

Online classification process described in step 2 is as follows：

16) online acquisition measurement data y ∈ R^m；

17) data are carried out with specificity analysis, the suitable feature extracting method of selection, PCA, KPCA or ICA；

18) after obtaining corresponding observation sequence, the HMM model most matched with it using forwards algorithms selection, so as to judge Fault type.

The data of described industrial failure have complex data characteristic, including linear, non-linear, non-gaussian.

Described industrial failure is blast furnace ironmaking process failure.

A kind of method described in basis is used for blast furnace ironmaking process fault diagnosis.

The present invention has following advantage：

1., present invention firstly provides a kind of industrial Fault Classification based on self-adaptive feature extraction, realize to complicated mistake The fault diagnosis of journey；

2. the present invention can be complicated for industrial process data characteristic, and the data to different qualities are carried using different features Method is taken, at utmost retains effective information, so as to improve the accuracy rate of fault diagnosis.

Brief description of the drawings

Fig. 1 is the FB(flow block) of the inventive method.

Specific embodiment

A kind of industrial Fault Classification based on self-adaptive feature extraction proposed by the present invention, its FB(flow block) such as Fig. 1 It is shown, including following steps：

Step one：Off-line modeling

7) Monitoring Data under sensor collection N kind operating modes, including a kind of nominal situation, and N-1 kind failures are assumed, and Every kind of Monitoring Data constitutes X=[x₁,x₂,…,x_n]^T∈R^n×m, wherein m represents the number of monitoring variable, and n represents number of samples, x_i∈R^m, i=1 ..., n represents i-th sample；

2) data characteristic analysis, including nonlinear metric and non-Gaussian system degree are carried out respectively to the N kinds Monitoring Data for gathering Amount；

2.1) nonlinear metric

2.1.1 raw data set X ∈ R) are given^n×m, l region is classified as, each area sample number is

2.1.2 the confidential interval of each element of correlation matrix) is calculated.The correlation matrix of raw data set can be represented such as Under：

According to Tables 1 and 2, the confidence limit of each element can be calculated using the average and variance of each variable, can be with table It is shown as following form：

Table 1：The calculating of average confidence limit

Table 2：The calculating of variance confidence limit

Pca model is equal to according to residual error and abandons characteristic value sum

Exact boundary can be tried to achieve by following formula：

Exact boundary is

By comparing residual error and exact boundary in regional, it may be determined that relationship between variables.If all of residual error all Fall within the exact boundary estimated, then illustrate between the data variable it is linear relationship, when at least one residual error falls in essence When beyond true border, it is non-linear relation to be judged as between variable.

2.2) non-Gaussian system measurement

Using normal probability paper principle, the non-Gaussian system of inspection data.Normal probability paper is whether a kind of inspection is totally The easy instrument relatively directly perceived of normal distribution.It is the ruled paper by being constituted perpendicular to some straight lines of transverse axis, the longitudinal axis.Transverse axis is By evenly divided, observation x, such as following formula are represented：

The longitudinal axis represents normal distribution cumulative probability value.The longitudinal axis is by non-evenly divided, the purpose is to make Normal Distribution Figure of the observation on normal probability paper be in straight line, will substantially be bent for Non-Gaussian Distribution.

3) data characteristic according to inspection carries out feature extraction from appropriate method.If being linear to use PCA between variable Carry out feature extraction, if be between variable it is nonlinear carry out feature extraction with KPCA, if between variable being being entered with ICA for non-Gaussian system Row feature extraction；

4) using the feature after extraction as the observation sequence O of HMM, HMM is trained, is obtained Its parameter lambda=(A, B, Π).

Wherein, A is hidden state transition probability matrix, and it describes the transition probability between each state in HMM model, See formula (7)~(8).

In formula,(it is abbreviated as a_i,j) represent in t, state is S_iUnder conditions of, it is S in t+1 moment states_jIt is general Rate.

B is observer state transition probability matrix, and it describes probability distribution of the Observable state under each hidden state, See formula (9)~(10).

In formula,(it is abbreviated as b_j,k) represent in t, hidden state is S_jUnder the conditions of, observer state is V_kProbability.

Π is initial state probabilities matrix.N is hidden state number, and M is Observable state number.

Step 2：Online classification

1) online acquisition measurement data y ∈ R^m；

2) data are carried out with specificity analysis, the suitable feature extracting method of selection, PCA, KPCA or ICA；

3) after obtaining corresponding observation sequence, calculate, and P (O | λ_i) (i=1,2 ... N) it represent in model library, give HMM parameter lambdas_iThe probability that observation sequence O occurs.Find maximum P (O | λ_j), illustrate that current failure j (j=1,2 ... N) occurs.

Above-described embodiment is used for illustrating the present invention, rather than limiting the invention, in spirit of the invention and In scope of the claims, any modifications and changes made to the present invention both fall within protection scope of the present invention.

Embodiment

Smelting iron and steel as one of most important basic industry in national economy, be weigh economic level for country and The important indicator of overall national strength.And blast furnace ironmaking is most important link in steel and iron industry production procedure, so to large blast furnace Damage is diagnosed to carry out studying significant with method for safe operation.

Blast furnace is a huge closed reaction vessel, and its internal smelting process is under high temperature, condition of high voltage, by one Serial complicated physical chemistry and heat transfer are reacted, and are a typical "black box" operations.Just because of the complexity inside blast furnace, So that the data of its collection have diversity, linear, non-linear, non-Gaussian system and dynamic etc..Therefore, it is proposed that side Diagnosis of the method to blast furnace failure has adaptability.The validity of the inventive method is illustrated with reference to No. 2 blast furnaces of Liu Gang.

Be found in the Liu Gang iron-smelters of 1958, be one have that 56 years equipment of brilliant history is advanced, equipment compared with Large-scale smelting enterprise high, major product is the pig iron, and byproduct has stove dirt, slag, blast furnace gas etc..It possesses 7 modernizations Blast furnace, blast furnace entirety dischargeable capacity is 11750 cubic metres, wherein No. 2 blast furnace dischargeable capacitys are 2000 cubic metres, it is current Guangxi Maximum blast furnace.After new blast furnace is gone into operation, iron-smelter will be provided with producing per year the integration capability of more than 10,000,000 tons of the pig iron.

Next implementation steps of the invention are set forth in reference to the detailed process：

Step one：Off-line modeling

8) Monitoring Data under 5 kinds of operating modes of hypothesis sensor collection, including a kind of nominal situation, and 4 kinds of failures, hanging, Collapse material, pipeline trip and stove cool.And every kind of Monitoring Data constitutes X=[x₁,x₂,…,x_n]^T∈R^n×m, wherein m represent monitoring become The number of amount, n represents number of samples, x_i∈R^m, i=1 ..., n represents i-th sample；

2) data characteristic analysis, including nonlinear metric and non-Gaussian system degree are carried out respectively to the 5 kinds of Monitoring Datas for gathering Amount；

2.1) nonlinear metric

2.1.1 raw data set X ∈ R) are given^n×m, l region is classified as, each area sample number is

2.1.2 the confidential interval of each element of correlation matrix) is calculated.The correlation matrix of raw data set can be represented such as Under：

According to Tables 1 and 2, the confidence limit of each element can be calculated using the average and variance of each variable, can be with table It is shown as following form：

Table 1：The calculating of average confidence limit

Table 2：The calculating of variance confidence limit

Pca model is equal to according to residual error and abandons characteristic value sum

Exact boundary can be tried to achieve by following formula：

Exact boundary is

By comparing residual error and exact boundary in regional, it may be determined that relationship between variables.If all of residual error all Fall within the exact boundary estimated, then illustrate between the data variable it is linear relationship, when at least one residual error falls in essence When beyond true border, it is non-linear relation to be judged as between variable.

2.2) non-Gaussian system measurement

Using normal probability paper principle, the non-Gaussian system of inspection data.Normal probability paper is whether a kind of inspection is totally The easy instrument relatively directly perceived of normal distribution.It is the ruled paper by being constituted perpendicular to some straight lines of transverse axis, the longitudinal axis.Transverse axis is By evenly divided, observation x, such as following formula are represented：

The longitudinal axis represents normal distribution cumulative probability value.The longitudinal axis is by non-evenly divided, the purpose is to make Normal Distribution Figure of the observation on normal probability paper be in straight line, will substantially be bent for Non-Gaussian Distribution.

3) data characteristic according to inspection carries out feature extraction from appropriate method.If being linear to use PCA between variable Carry out feature extraction, if be between variable it is nonlinear carry out feature extraction with KPCA, if between variable being being entered with ICA for non-Gaussian system Row feature extraction；

4) using the feature after extraction as the observation sequence O of HMM, HMM is trained, is obtained Its parameter lambda=(A, B, Π).

Wherein, A is hidden state transition probability matrix, and it describes the transition probability between each state in HMM model, See formula (7)~(8).

In formula,(it is abbreviated as a_i,j) represent in t, state is S_iUnder conditions of, it is S in t+1 moment states_jIt is general Rate.

B is observer state transition probability matrix, and it describes probability distribution of the Observable state under each hidden state, See formula (9)~(10).

In formula,(it is abbreviated as b_j,k) represent in t, hidden state is S_jUnder the conditions of, observer state is V_kProbability.

Π is initial state probabilities matrix.N is hidden state number, and M is Observable state number.M=5 in this example；

Step 2：Online classification

1) online acquisition measurement data y ∈ R^m；

2) data are carried out with specificity analysis, the suitable feature extracting method of selection, PCA, KPCA or ICA；

3) after obtaining corresponding observation sequence, calculate, and P (O | λ_i) (i=1,2 ... 5) it represent in model library, give HMM parameter lambdas_iThe probability that observation sequence O occurs.Find maximum P (O | λ_j), illustrate that (j=1,2 ... 5) current failure j occur.

Above-described embodiment is used for illustrating the present invention, rather than limiting the invention, in spirit of the invention and In scope of the claims, any modifications and changes made to the present invention both fall within protection scope of the present invention.

Claims (1)

1. a kind of industrial Fault Classification based on self-adaptive feature extraction, it is characterised in that as follows the step of the method：

Step one：Off-line modeling, to the off-line data of industrial process collection, carries out data characteristic analysis, for different data Characteristic, using the feature after extraction as the observation sequence of HMM (HMM), is instructed from suitable feature extracting method Practice HMM；

Step 2：Inline diagnosis, the data to online acquisition carry out data characteristic analysis, for different data characteristics from suitable The feature extracting method of conjunction, after obtaining corresponding observation sequence, the HMM model that selection is most matched, so that failure judgement type；

Off-line modeling process described in step one is as follows：

1) X=[x are constituted using the Monitoring Data of industrial process collection₁,x₂,…,x_n]^T∈R^n×m, wherein m represents monitoring variable Number, n represents number of samples, x_i∈R^m, i=1 ..., n represents i-th sample；

2) data for gathering are carried out with data characteristic analysis, including nonlinear metric and non-Gaussian system measurement；

2.1) nonlinear metric

2.1.1 raw data set X ∈ R) are given^n×m, l region is classified as, each area sample number is

2.1.2 the confidential interval of each element in the correlation matrix of raw data set, the correlation matrix table of raw data set) are calculated Show as follows：

The confidence limit of each element is calculated using the average and variance of each variable, is expressed as following form

Pca model is equal to according to residual error and abandons characteristic value sum

$\mathrm{\σ}=\underset{j=1}{\overset{m}{\Σ}}{\mathrm{\σ}}_j=\underset{k=a+1}{\overset{m}{\Σ}}{\mathrm{\λ}}_k---\left(3\right),$

Exact boundary can be tried to achieve by following formula：

$\begin{array}{c}{\mathrm{\λ}}_{k_{MAX}}=\underset{{\mathrm{\ΔS}}_{{\mathrm{zz}}_{MAX}}}{{\mathrm{argmax\λ}}_k}(S_{ZZ}+{\mathrm{\ΔS}}_{{\mathrm{ZZ}}_{MAX}})\\ {\mathrm{\λ}}_{k_{MIN}}=\underset{{\mathrm{\ΔS}}_{{\mathrm{zz}}_{MIN}}}{{\mathrm{argmax\λ}}_k}(S_{ZZ}+{\mathrm{\ΔS}}_{{\mathrm{ZZ}}_{MIN}})\end{array}---\left(4\right),$

Exact boundary is

${\mathrm{\σ}}_{MAX}=\underset{k=a+1}{\overset{m}{\Σ}}{\mathrm{\λ}}_{k_{MAX}},{\mathrm{\σ}}_{MIN}=\underset{k=a+1}{\overset{m}{\Σ}}{\mathrm{\λ}}_{k_{MIN}}---\left(5\right),$

By comparing residual error and exact boundary in regional, it may be determined that relationship between variables, all fall if all of residual error Within the exact boundary of estimation, then illustrate between the data variable it is linear relationship, when at least one residual error falls on accurate side When beyond boundary, it is non-linear relation to be judged as between variable；

2.2) non-Gaussian system measurement

Using normal probability paper principle, the non-Gaussian system of inspection data, normal probability paper be by perpendicular to transverse axis, the longitudinal axis it is some The ruled paper that bar straight line is constituted, transverse axis is, by evenly divided, to represent observation x, such as following formula：

$\mathrm{\φ}\left(X\right)=\underset{-\mathrm{\∞}}{\overset{+\mathrm{\∞}}{\∫}}\frac{1}{\sqrt{2\mathrm{\π}}}{e}^{-\frac{t^2}{2}}dt---\left(6\right)$

The longitudinal axis represents normal distribution cumulative probability value, and the longitudinal axis is by non-evenly divided, the purpose is to make the sight of Normal Distribution Figure of the measured value on normal probability paper is in straight line, will substantially be bent for Non-Gaussian Distribution；

3) data characteristic according to inspection carries out feature extraction from appropriate method, if being linear to be carried out with PCA between variable Feature extraction, if be between variable it is nonlinear carry out feature extraction with KPCA, if being that non-Gaussian system with ICA carries out spy between variable Levy extraction；

4) using the feature after extraction as the observation sequence O of HMM, HMM is trained, obtains its ginseng Number λ=(A, B, Π)；

Wherein, A is hidden state transition probability matrix, describes the transition probability between each state in HMM model, sees below formula (7)~(8),

$\begin{array}{c}{a}_{S_i,S_j}=a_{i,j}=P\[q_{t+1}=S_j|q_t=S_i\],1\≤i,j\≤N\\ \underset{j=1}{\overset{N}{\Σ}}{a}_{i,j}=1,1\≤i\≤N\end{array}---\left(8\right)$

In formula,It is abbreviated as a_i,j, representing in t, state is S_iUnder conditions of, it is S in t+1 moment states_jProbability；

B is observer state transition probability matrix, describes probability distribution of the Observable state under each hidden state, sees below formula (9)~(10),

$\begin{array}{c}{b}_{S_j,V_k}=b_{j,k}=P\[O_t=V_k|q_t=S_j\],1\≤j\≤N,1\≤k\≤M\\ \underset{k=1}{\overset{M}{\Σ}}{b}_{j,k}=1,1\≤j\≤N\end{array}---\left(10\right)$

In formula,It is abbreviated as b_j,k, representing in t, hidden state is S_jUnder the conditions of, observer state is V_kProbability；

Π is initial state probabilities matrix；

Online classification process described in step 2 is as follows：

1) online acquisition measurement data y ∈ R^m；

2) data are carried out with specificity analysis, the suitable feature extracting method of selection, PCA, KPCA or ICA；

3) after obtaining corresponding observation sequence, the HMM model most matched with it using forwards algorithms selection, so that failure judgement class Type；

The data of described industrial failure have complex data characteristic, including linear, non-linear, non-gaussian；

Described industrial failure is blast furnace ironmaking process failure.