CN110647922A - Layered non-Gaussian process monitoring method based on public and special feature extraction - Google Patents

Abstract

The invention discloses a layered non-Gaussian process monitoring method based on public and special feature extraction. Dividing training data into a plurality of modes, and obtaining high-order public characteristics of the training data by applying a plurality of weight vectors, component vectors and other characteristic quantities in each mode; obtaining low-order public features of the training data by applying a plurality of weight vectors, component vectors and other feature quantities in the remaining low-order modes; constructing statistical limits and statistics in the public subspace according to the public characteristics, and carrying out fault detection; and constructing statistical limits and statistics in the remaining unique subspaces to detect faults. The method is superior to other traditional methods in multi-modal non-Gaussian process fault detection, can extract the special characteristics of the multi-modal and can also extract the common characteristics, and considers the mutual connection of the characteristics and the commonality among different modes, so that the multi-modal process monitoring is more effective.

Description

Layered non-Gaussian process monitoring method based on public and special feature extraction

Technical Field

The invention belongs to the field of industrial process system engineering, and relates to a multi-mode fault monitoring method applied to complex industrial processes such as chemical production and the like.

Background

The complexity of the complex industrial process is reflected in that the process of the modern industrial production generates a large amount of data, the process data generally have different process characteristics, and the distribution characteristics are unknown; the industrial process has a plurality of units, and the data has a mechanism relation; moreover, the traditional method mostly assumes that the chemical process is operated under a single and stable operation condition, but actually the plant-level data is multi-modal, so how to better monitor the multi-modal problem is very worthy of study.

In the past, a great deal of research has been conducted by researchers on advanced data classification and information fusion methods, so that the new measurement data can be accurately distributed to actual operation modes, and the monitoring results can be displayed more accurately. It is noted that there are some process monitoring methods that study global process monitoring methods by using different patterns of data simultaneously, such as support vector data description, k-nearest neighbor based methods, multi-block partial least squares. However, most of these methods related to multi-modal process monitoring can only acquire the unique features of each mode data, and ignore the common features among each mode data. In actual monitoring, it would be beneficial to monitor multimodal processes if the correlation of characteristics and commonalities between different modes could be obtained in advance. Some researchers have proposed a two-step model framework, the first step being the need to develop separate process monitoring models, and the second step being the examination of similarities between them to obtain models representing common features. There have also been some studies attempting to divide the data of each mode into common and unique parts by using the data of multiple modes simultaneously, and although this method has been further extended to the monitoring of the transition process between modes, this method can be used simply for the case of calculating both modes. Moreover, the common features and characteristic features between different modes are not clearly explained in the geometrical relationship, so that the method has no way to carry out deeper research on the multimode process. A two-step multi-set principal component analysis method has also been proposed to obtain common basis vectors for multiple sets of data. The scholars propose a multimode process monitoring method based on a least square subspace based on the obtained common base vector, although the multimode process monitoring method also successfully extends to the conversion process between monitoring modes, the common base vector does not span any subspace of subset data and only serves as an auxiliary vector to improve the correlation of a cross set to the maximum extent, and therefore, the common characteristics of multimode data cannot be accurately obtained by directly mapping the data to the subspace. Therefore, how to extract the relationship among the multi-modal data and apply different models to the closely-connected multi-modal data becomes an important problem for fault monitoring.

Disclosure of Invention

In order to overcome the problems that in the multi-modal monitoring process, the traditional multi-modal process fault monitoring is carried out, only the specific characteristics of each mode data are developed, and the mutual connection of the characteristics and the commonality among different modes is ignored, the invention aims to provide a layered non-Gaussian process monitoring method based on common and specific characteristic extraction, the data collected by a sensor in industrial production is utilized to carry out real-time analysis and processing to obtain the detection and monitoring of real-time fault states, and the technical problem that the multi-modal process monitoring effect is poor in the prior art is solved.

The invention seeks to simultaneously construct process monitoring models from all mode data, obtain a common subspace by capturing common characteristics behind different modes, obtain an independent subspace reflecting the characteristic characteristics of each mode, and then carry out fault detection method research on each subspace.

The purpose of the invention is realized by the following technical scheme:

firstly, the classified multi-modal data are classified to obtain public features through weight vector parameters and weight vector scores, and fault monitoring is carried out by utilizing the public features. And introducing a double-layer non-Gaussian monitoring method aiming at the residual characteristic features to monitor the faults.

The specific method comprises the following steps:

1) acquiring data of an input variable x and an output variable y of the industrial production process as training data through a sensor under the condition that no fault is known in the industrial production process;

in industryAcquiring an input variable x of the industrial production process on line through a sensor under the condition that a fault needs to be detected in the production process_testAnd an output variable y_testThe data to be detected is used as the data to be detected;

2) classifying the training data by adopting a fault detection method of a layered non-Gaussian algorithm based on sample multi-modal classification to obtain classified training data;

3) aiming at the classified training data, each classification corresponds to one mode, an optimization objective function is established in each mode by using a plurality of weight vectors, component vectors and other characteristic quantities to obtain the classified training data and the common characteristics of the data to be detected, the step iii) in the fault monitoring method of the industrial production process based on the hierarchical non-Gaussian monitoring algorithm is adopted according to the common characteristics to extract the training data of the non-Gaussian part with the common characteristics in each mode, and the statistical limit of the training data of the non-Gaussian part with the common characteristics and the statistical quantity of the data to be detected are processed;

4) after training data of a non-Gaussian part with common characteristics are obtained, extracting the common characteristics of the remaining Gaussian part data, establishing an optimization objective function by using a plurality of weight vectors, component vectors and other characteristic quantities in each mode for the Gaussian part data to obtain the common characteristics of the Gaussian part data, and combining a partial least square method according to the common characteristics to obtain the statistical limit of the training data of the Gaussian part with the common characteristics and the statistical quantity of data to be detected;

5) after Gaussian part data with common characteristics and non-Gaussian part data with common characteristics are extracted, the rest data are data with characteristic characteristics, and according to the data with characteristic characteristics, the iv) step in the industrial production process fault monitoring method based on the hierarchical non-Gaussian monitoring algorithm is adopted to obtain the statistical limit of training data with characteristic characteristics and the statistical quantity of data to be detected;

6) and classifying the data to be detected in an online identification mode to obtain the monitored and classified data to be detected.

The input variable x is, for example, a process variable in an industrial process.

The output variable y is, for example, a quality variable in an industrial process.

In the step 3), the concrete steps are as follows:

a) forming all input variables X of the industrial process into an input variable data set X, wherein X belongs to R^n×M·mCarrying out standardization processing on each input variable X in an input variable data set X, wherein n is the number of sampling values in the input variable X, the input variable X is composed of a series of sampling values, M is the number of the input variables, and M is the classification number of training data, namely the mode number; outputting the variable data set as Y in the same way;

3, b) the classified input variable data is expressed as: x ═ X₍₁₎ X₍₂₎ ... X_(M)]The output variable data is expressed as: y ═ Y₍₁₎ Y₍₂₎ ... Y_(M)]；X₍₁₎ X₍₂₎ ... X_(M)Input variable data respectively expressed as first to Mth classifications, Y₍₁₎ Y₍₂₎ ... Y_(M)Output variable data respectively expressed as first to Mth classifications;

3, c) thus, the composition variable t of the input variable x for each modality in the data is obtained using the following formula_cAnd a component vector u of an output variable y of each modality_c：

Wherein p is_cAnd q is_cRespectively single input and output common weight vectors; α, β are input and output component weight vector parameters for combining M pieces of modal data, α ═ α₍₁₎ α₍₂₎ ... α_(M)]，α₍₁₎ α₍₂₎ ... α_(M)Expressed as input component weight vector parameters of first to mth classes, respectively, β ═ β₍₁₎ β₍₂₎ ... β_(M)]，β₍₁₎ β₍₂₎ ... β_(M)Output component weight vector parameters of a first classification to an Mth classification are respectively expressed, wherein i represents an ith mode;

establishing a first objective function as:

max{I(t_c,u_c)}

s.t.||p_c||＝||q_c||＝||α||＝||β||＝1

wherein, I (t)_c,u_c) Component variable t being input variable x_cAnd component vector u of output variable y_cThe mutual information value between, | | | | represents the modulus;

3, d) solving by adopting an iterative algorithm to obtain input and output public weight vectors p_c、q_cAnd input and output component weight vector parameters alpha and beta, and then processing to obtain non-Gaussian partial data with common features, which is expressed as

And is obtained by processing training data of non-Gaussian parts with common features

Statistical limit of statistics

3.e) pre-multiplying the input variable of the data to be detected by p_cObtaining a first fault estimation vector of the data to be detected, and obtaining the data to be detected by using the first fault estimation vector

Statistics are obtained.

In the step 3.d), the following iterative algorithm is used for solving and obtaining the input and output public weight vector p_c、q_cAnd input and output component weight vector parameters α, β:

(3.d.1) randomly initializing input and output component weight vector parameters alpha and beta;

(3.d.2) will loseThe input and output component weight vector parameters alpha and beta are substituted into the classified input variable data and the classified output variable data to obtain the weighted input data

And output data

(3.d.3) solving an input and output public weight vector by adopting an industrial production process fault monitoring method based on a layered non-Gaussian monitoring algorithm;

(3.d.4) inputting and outputting the public weight vector p obtained in the last step_cAnd q is_cIs brought into a first objective function;

(3, d.5) solving input and output component weight vector parameters alpha and beta by using a particle swarm algorithm;

(3.d.6) repeating the steps (2) - (6) until α, β and p_c，q_cConvergence, i.e. the four values are all less than the respective preset threshold.

Step 4), removing the data of the non-gaussian part obtained in step 3) from the training data, and remaining the training data of the gaussian part, and then the specific steps are as follows:

a) Gaussian data of the classified input variables are represented as: x_G＝[X_G(1) X_G(2) ... X_G(M)]The gaussian data for the output variable is expressed as: y is_G＝[Y_G(1) Y_G(2) ... Y_G(M)]；X_G(1) X_G(2) ... X_G(M)Gaussian data, Y, respectively expressed as input variables of the first to Mth classes_G(1) Y_G(2) ... Y_G(M)Gaussian data respectively expressed as output variables of the first to mth classifications;

4.b) thus, the component variable t of the input variable x for each modality in the Gaussian partial data is obtained using the following formula_c,GAnd a component vector u of an output variable y of each modality_c,G：

Wherein p is_c,GAnd q is_c,GA single input, output common weight vector of the gaussian partial data, respectively; alpha is alpha_GAnd beta_GIs the input and output component weight vector parameter, alpha, of the combined M modal data of the Gaussian partial data_G＝[α_G(1) α_G(2) ... α_G(M)]，α_G(1) α_G(2) ... α_G(M)Expressed as the input component weight vector parameter, beta, of the first to Mth classes, respectively_G＝[β_G(1)β_G(2) ... β_G(M)]，β_G(1) β_G(2) ... β_G(M)Output component weight vector parameters of a first classification to an Mth classification are respectively expressed, wherein i represents an ith mode;

establishing a second objective function as:

max{I(t_c,G,u_c,G)}

s.t.||p_c,G||＝||q_c,G||＝||α_G||＝||β_G||＝1

wherein, I (t)_c,G,u_c,G) Component variable t of input variable x in Gaussian part data_c,GAnd component vector u of output variable y_c,GThe mutual information value between, | | | | represents the modulus;

4, c) solving and obtaining the input and output public weight vector p in the Gaussian part data by adopting an iterative algorithm_c,G、q_c,GAnd input and output component weight vector parameter alpha_G、β_GThen processing to obtain training data of a Gaussian part with common characteristics;

and T is obtained by processing the training data of Gaussian part with common characteristics_c ²Statistical limit of statistics

4, d) pre-multiplying the input variable of the data to be detected by p_c,GObtaining a second fault estimation vector of the data to be detected, and obtaining T of the data to be detected by using the second fault estimation vector in combination with a partial least square method_c ²Statistics are obtained.

In the step 4.c), the following iterative algorithm is used for solving and obtaining the input and output public weight vector p in the Gaussian partial data_c,G、q_c,GAnd input and output component weight vector parameter alpha_G、β_G：

(4.c.1) randomly initializing input and output component weight vector parameter alpha in Gaussian partial data_G、β_G；

(4.c.2) inputting and outputting weight vector parameter alpha of input and output components in Gaussian part data_G、β_GThe classified input variable data and the classified output variable data are brought in to obtain weighted input data

And output data

(4, c.3) solving input and output public weight vectors p in Gaussian part data by adopting industrial production process fault monitoring method based on hierarchical non-Gaussian monitoring algorithm_c,GAnd q is_c,G；

(4.c.4) inputting and outputting the public weight vector p in the Gaussian part data obtained in the last step_c,GAnd q is_c,GIs brought into a second objective function;

(4, c.5) solving the weight vector parameter alpha of the input and output components in the Gaussian part data by using a particle swarm algorithm_G、β_G；

(4, c.6) repeating the processes of steps (2) - (6) until alpha_G、β_GAnd p_c,G、q_c,GConvergence, i.e. the four values are all less than the respective preset threshold.

In the step 5), the concrete steps are as follows:

a) when the public weight vectorp_cAfter the determination, the training data of each modality with the characteristic features remaining is obtained by processing in the following way:

X_I(i)＝X_(i)-X_(i)*p_cp_c ^T

wherein, X_I(i)Input variable data set representing training data of the i-th modality with characteristic features, Y_I(i)An output variable data set representing training data of the ith modality with characteristic features, X_(i)Input variable data set, Y, representing training data for the ith modality_(i)An output variable data set representing training data for an ith modality;

5, b) aiming at the training data of each mode with the characteristic features, respectively obtaining a load matrix aiming at a non-Gaussian part and a Gaussian part of the training data by adopting step iii) to step iv) of the fault monitoring method of the industrial production process based on the layered non-Gaussian monitoring algorithm, and obtaining I by utilizing the load matrix to process_i ²Statistical limit corresponding to statistical quantity

C) removing the obtained data of the non-Gaussian part from the training data with the characteristic features, remaining the training data of the Gaussian part with the characteristic features, adopting step iii to step iv in the fault monitoring method of the industrial production process based on the hierarchical non-Gaussian monitoring algorithm to process the training data of the Gaussian part with the characteristic features to respectively obtain load matrixes of the Gaussian part and the Gaussian part of the training data, and obtaining T by utilizing the load matrix processing_i ²Statistical limits corresponding to the statistics and Q statisticsAnd a statistical limit Q_{_limit}；

5, d) pre-multiplying the data to be detected with the Gaussian part with the characteristic by the training data of the Gaussian part with the characteristic to obtain a load matrix, and pre-multiplying the input variable of the data to be detected with the Gaussian part by the load matrix P_GObtaining a fourth fault estimation vector of the data to be detected, and obtaining the data to be detected T by using the fourth fault estimation vector_i ²Statistics and Q statistics.

In the step 6), the concrete steps are as follows:

the data to be detected starts to operate, but the types of modal working conditions are unknown, one mode corresponds to one modal working condition, the industrial production process has modal working conditions under M modes in total, the data to be detected of each known mode is monitored to obtain each statistical limit, the data to be detected of each current unknown mode is monitored to obtain each statistical quantity, and the following judgment is carried out:

if the current unknown mode of the data to be detected

statistic/T_c ²The statistics being obtained while being higher than the corresponding training data of the first modality representing common features

Statistical limit of statistics

Or T_c ²Statistical limit of statisticsThe data to be detected in the current unknown mode is fault data;

if the current unknown mode of the data to be detected

statistic/T_c ²Obtained with statistics lower than corresponding training data of the first modality representing common features

Statistical limit of statistics

Or T_c ²Statistical limit of statistics

Then judge

statistic/T_i ²Whether or not the statistic/Q statistic is simultaneously higher than the corresponding representative characteristic

Statistical limit of statistics

T_i ²Statistical limit of statistics

And statistical limit Q of Q statistic_{_limit}：

If not, the data to be detected of the current unknown mode belongs to a first mode;

if the current mode is higher than the preset threshold, the training data of the second mode is called for judgment, and the data to be detected of the current unknown mode is judged

statistic/T_i ²Whether the statistic/Q statistic is simultaneously higher than that obtained by training data of a second modality representing characteristic features

Statistical limit of statistics

T_i ²Statistical limit of

And statistical limit of Q_{_limit}：

If not, the data to be detected in the current unknown mode belongs to a second mode;

if the current modal is higher than the preset threshold, training data of a third modal is called for judgment, and the current unknown modal to-be-detected data is judged

statistic/T_i ²Whether the statistic/Q statistic is simultaneously higher than the I obtained by the corresponding training data of the third modality representing the characteristic feature_i ²Statistical limit of statistics

T_i ²Statistical limit of statistics

And statistical limit Q of Q statistic_{_limit}：

If not, the data to be detected in the current unknown mode belongs to a third mode;

if the current modal is higher than the first modal, calling the training data of the fourth modal for judgment, and so on until the current unknown modal to be detected data I_i ²statistic/T_i ²The statistic/Q statistic being higher than that obtained by modal training data

Statistical limit of statistics

T_i ²Statistical limit of statistics

And statistical limit Q of Q statistic_{_limit}Consider the current unknown modality to be detected asAnd (4) fault data, and finishing monitoring and classification of the data to be detected.

And step 7) is also included after the step 6), specifically, the fault detection is carried out by utilizing the statistical limits and statistics in the step 3), the step 4) and the step 5).

The step 7), specifically, the statistical limit and the statistical quantity in the step 3), 4), 5) meet the data to be detected corresponding to any one of the following five conditions, namely, the industrial production process is considered to have a process fault:

for data of non-Gaussian part with common features in the data to be detected, the data is above the statistical limit

Data to be detected corresponding to the statistic;

for the data of Gaussian part with common features in the data to be detected, the data is positioned at T above the statistical limit_c ²Data to be detected corresponding to the statistic;

for data of non-Gaussian part with characteristic features in the data to be detected, the data is located above the statistical limitData to be detected corresponding to the statistic;

for the data of Gaussian part with characteristic features in the data to be detected, the data is positioned at T above the statistical limit_i ²And the statistic and the data to be detected corresponding to the Q statistic.

The invention has the beneficial effects that:

and extracting multi-modal common features and special features for tracking the running performance of the process and the quality of the process. The method seeks to simultaneously construct process monitoring models from all mode data, obtain a common subspace by capturing common features behind different modes, obtain an independent subspace reflecting the characteristic features of each mode, and then carry out fault detection method research on each subspace.

The method is superior to other traditional methods in fault detection of the non-Gaussian multi-modal process, can fully consider the highly complex coupling relation among variables, can extract the non-Gaussian part in the data with unknown distribution characteristics, and can effectively extract the common characteristics and the unique characteristics of the multi-modal data, so that the fault monitoring of the multi-modal process is more efficient and accurate.

The method is superior to other traditional methods in multi-modal non-Gaussian process fault detection, can extract the special characteristics of the multi-modal and can also extract the common characteristics, and considers the mutual connection of the characteristics and the commonality among different modes, so that the multi-modal process monitoring is more effective.

Drawings

FIG. 1 is a diagram of the classification results of the multi-modal process of the present invention.

Fig. 2 is a diagram of the first failure detection result of the present invention.

Fig. 3 is a diagram of a second failure detection result of the present invention.

Detailed Description

The method of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

The specific implementation case adopted by the invention is a Tennessee-Emament (TE) process, which comprises five main units: a reactor, a condenser, a compressor, a separator, and a stripper.

The product stream from the reactor was cooled by a condenser and then sent to a vapor/liquid separator. The vapor from the separator is recycled to the reactor by means of a compressor. To prevent the accumulation of inert components and reaction by-products in the process, a portion of the recycle stream must be vented. The condensed components from the separator (stream 10) are pumped to the stripper. Stream 4 is used to strip the remaining reactants in stream 10 which are combined with the recycle stream via stream 5 for further reaction.

The Tennessee-Emersmann (TE) process has 41 measurement variables (including 22 process measurement variables and 19 component measurement variables) and 12 control variables, and in order to monitor the faults of the TE process, ensure the safety of the actual chemical production process and improve the economic benefit, 31 variables are selected from the 22 process measurement variables and the 12 control variables as input variables shown in Table 1 and 6 component measurement variables, namely quality variables, are selected from the 19 measurement variables as output variables when the process monitoring variables are determined, as shown in Table 2. The tennessee-eastman process also simulated 21 faults as shown in table 3.

TABLE 1

TABLE 2

TABLE 3

In order to demonstrate the superiority of the proposed processing method, the present invention is embodied as a multi-modal process case including different stable models, from which 3000 samples are collected in total, and a data matrix of 50 variables is used as a training data set, and the multi-modal process is divided into three modes, wherein the first mode includes 1000 samples, the second mode includes 1000 samples, and the third mode includes 1000 samples, as shown in fig. 1. In order to achieve the purpose of fault detection, two faults are designed, wherein the fault 1 is that a step fault is added to a sample from 300 to 1000 of a 22 th variable of a first modality, and the fault 2 is that a slope fault is added to a sample from 1500 to 2000 of the 22 th variable of a second modality, and the fault detection is respectively carried out as a data set to be detected, and the specific steps are as follows:

1) under the condition that no fault is known in the industrial production process, acquiring data of an input variable x and an output variable y in the industrial production process as training data through a sensor, and carrying out standardization processing to enable the mean value of the data to be 0 and the variance to be 1;

under the condition that the industrial production process needs to detect faults, the industrial production process is acquired on line through a sensorInput variable x of the program_testAnd an output variable y_testThe two fault data are used as data to be detected, and standardized processing is carried out, so that the mean value of the data is 0, and the variance is 1, and the data are used as the data to be detected;

2) classifying the training data by adopting a fault detection method of a layered non-Gaussian algorithm based on sample multi-modal classification to obtain classified training data;

3) aiming at the classified training data, each classification corresponds to one mode, an optimization objective function is established in each mode by using a plurality of weight vectors, component vectors and other characteristic quantities to obtain the classified training data and the common characteristics of the data to be detected, the step iii) in the fault monitoring method of the industrial production process based on the hierarchical non-Gaussian monitoring algorithm is adopted according to the common characteristics to extract the training data of the non-Gaussian part with the common characteristics in each mode, and the statistical limit of the training data of the non-Gaussian part with the common characteristics and the statistical quantity of the data to be detected are processed;

4) after training data of a non-Gaussian part with common characteristics are obtained, extracting the common characteristics of the remaining Gaussian part data, establishing an optimization objective function by using a plurality of weight vectors, component vectors and other characteristic quantities in each mode for the Gaussian part data to obtain the common characteristics of the Gaussian part data, and combining a partial least square method according to the common characteristics to obtain the statistical limit of the training data of the Gaussian part with the common characteristics and the statistical quantity of data to be detected;

5) after Gaussian part data with common characteristics and non-Gaussian part data with common characteristics are extracted, the rest data are data with characteristic characteristics, and according to the data with characteristic characteristics, the iv) step in the industrial production process fault monitoring method based on the hierarchical non-Gaussian monitoring algorithm is adopted to obtain the statistical limit of training data with characteristic characteristics and the statistical quantity of data to be detected;

6) and classifying the data to be detected in an online identification mode to obtain the monitored and classified data to be detected.

The invention can simultaneously construct the process monitoring model from all mode data, obtain the public subspace by capturing the public characteristics behind different modes, and obtain the independent subspace reflecting the characteristic characteristics of each mode, and then carry out the fault detection method research on each subspace, thereby improving the accuracy and effectiveness of fault monitoring.

The above embodiments are disclosed to illustrate the present invention, not to limit the present invention, and any modifications and changes made within the spirit of the present invention and the scope of the claims fall within the scope of the present invention.

Claims (9)

1. A layered non-Gaussian process monitoring method based on public and unique feature extraction is characterized by comprising the following steps:

1) acquiring data of an input variable x and an output variable y of the industrial production process as training data through a sensor under the condition that no fault is known in the industrial production process;

under the condition that the industrial production process needs to detect faults, acquiring an input variable x of the industrial production process on line through a sensor_testAnd an output variable y_testThe data to be detected is used as the data to be detected;

2) classifying the training data to obtain classified training data;

3) aiming at the classified training data, each classification corresponds to one mode, an optimization objective function is established in each mode by using a plurality of weight vectors, component vectors and other characteristic quantities to obtain the classified training data and the common characteristics of the data to be detected, the training data of the non-Gaussian part with the common characteristics in each mode are extracted according to the common characteristics, and the statistical limit of the training data of the non-Gaussian part with the common characteristics and the statistical quantity of the data to be detected are processed;

4) after training data of a non-Gaussian part with common characteristics are obtained, extracting the common characteristics of the remaining Gaussian part data, establishing an optimization objective function by using a plurality of weight vectors, component vectors and other characteristic quantities in each mode for the Gaussian part data to obtain the common characteristics of the Gaussian part data, and combining a partial least square method according to the common characteristics to obtain the statistical limit of the training data of the Gaussian part with the common characteristics and the statistical quantity of data to be detected;

5) after Gaussian part data with common characteristics and non-Gaussian part data with common characteristics are extracted, the rest data are data with characteristic characteristics, and statistical limits of training data with characteristic characteristics and statistics of data to be detected are obtained according to the data with characteristic characteristics;

6) and classifying the data to be detected in an online identification mode to obtain the monitored and classified data to be detected.

2. The hierarchical non-gaussian process monitoring method based on common and unique feature extraction as claimed in claim 1, wherein:

in the step 3), the concrete steps are as follows:

a) forming all input variables X of the industrial process into an input variable data set X, wherein X belongs to R^n×M·mStandardizing each input variable X in an input variable data set X, wherein n is the number of sampling values in the input variable X, M is the number of input variables, and M is the classification number of training data, namely a modal number;

3, b) the classified input variable data is expressed as: x ═ X₍₁₎ X₍₂₎ ... X_(M)]The output variable data is expressed as: y ═ Y₍₁₎ Y₍₂₎ ... Y_(M)]；X₍₁₎ X₍₂₎ ... X_(M)Input variable data respectively expressed as first to Mth classifications, Y₍₁₎ Y₍₂₎ ... Y_(M)Output variable data respectively expressed as first to Mth classifications;

3, c) thus, the composition variable t of the input variable x for each modality in the data is obtained using the following formula_cAnd a component vector u of an output variable y of each modality_c：

Wherein p is_cAnd q is_cRespectively single input and output common weight vectors; α, β are input and output component weight vector parameters for combining M pieces of modal data, α ═ α₍₁₎ α₍₂₎ ... α_(M)]，α₍₁₎ α₍₂₎ ... α_(M)Expressed as input component weight vector parameters of first to mth classes, respectively, β ═ β₍₁₎ β₍₂₎ ... β_(M)]，β₍₁₎ β₍₂₎ ... β_(M)Output component weight vector parameters of a first classification to an Mth classification are respectively expressed, wherein i represents an ith mode;

establishing a first objective function as:

max{I(t_c,u_c)}

s.t.||p_c||＝||q_c||＝||α||＝||β||＝1

wherein, I (t)_c,u_c) Component variable t being input variable x_cAnd component vector u of output variable y_cThe mutual information value between, | | | | represents the modulus;

3, d) solving by adopting an iterative algorithm to obtain input and output public weight vectors p_c、q_cAnd inputting and outputting the component weight vector parameters alpha and beta, then processing to obtain non-Gaussian part data with common characteristics, and processing the non-Gaussian part data with common characteristics to obtain the non-Gaussian part data

Statistical limit of statistics

3.e) pre-multiplying the input variable of the data to be detected by p_cObtaining a first fault estimation vector of the data to be detectedObtaining data to be detected using the first fault estimate vector

Statistics are obtained.

3. The hierarchical non-gaussian process monitoring method based on common denominator feature extraction as claimed in claim 2, wherein:

in the step 3.d), the following iterative algorithm is used for solving and obtaining the input and output public weight vector p_c、q_cAnd input and output component weight vector parameters α, β:

(3.d.1) randomly initializing input and output component weight vector parameters alpha and beta;

(3.d.2) substituting the input and output component weight vector parameters alpha, beta into the classified input variable data and the classified output variable data to obtain weighted input data

And output data

(3.d.3) solving an input and output public weight vector;

(3.d.4) inputting and outputting the public weight vector p obtained in the last step_cAnd q is_cIs brought into a first objective function;

(3, d.5) solving input and output component weight vector parameters alpha and beta by using a particle swarm algorithm;

(3.d.6) repeating the steps (2) - (6) until α, β and p_c，q_cAnd (6) converging.

4. The hierarchical non-gaussian process monitoring method based on common denominator feature extraction as claimed in claim 1, wherein:

step 4), removing the data of the non-gaussian part obtained in step 3) from the training data, and remaining the training data of the gaussian part, and then the specific steps are as follows:

a) Gaussian data of the classified input variables are represented as: x_G＝[X_G(1) X_G(2) ... X_G(M)]The gaussian data for the output variable is expressed as: y is_G＝[Y_G(1) Y_G(2) ... Y_G(M)]；X_G(1) X_G(2) ... X_G(M)Gaussian data, Y, respectively expressed as input variables of the first to Mth classes_G(1) Y_G(2) ... Y_G(M)Gaussian data respectively expressed as output variables of the first to mth classifications;

4.b) thus, the component variable t of the input variable x for each modality in the Gaussian partial data is obtained using the following formula_c,GAnd a component vector u of an output variable y of each modality_c,G：

Wherein p is_c,GAnd q is_c,GA single input, output common weight vector of the gaussian partial data, respectively; alpha is alpha_GAnd beta_GIs the input and output component weight vector parameter, alpha, of the combined M modal data of the Gaussian partial data_G＝[α_G(1) α_G(2) ... α_G(M)]，α_G(1)α_G(2) ... α_G(M)Expressed as the input component weight vector parameter, beta, of the first to Mth classes, respectively_G＝[β_G(1) β_G(2) ... β_G(M)]，β_G(1) β_G(2) ... β_G(M)Output component weight vector parameters of a first classification to an Mth classification are respectively expressed, wherein i represents an ith mode;

establishing a second objective function as:

max{I(t_c,G,u_c,G)}

s.t.||p_c,G||＝||q_c,G||＝||α_G||＝||β_G||＝1

wherein, I (t)_c,G,u_c,G) Component variable t of input variable x in Gaussian part data_c,GAnd component vector u of output variable y_c,GThe mutual information value between, | | | | represents the modulus;

4, c) solving and obtaining the input and output public weight vector p in the Gaussian part data by adopting an iterative algorithm_c,G、q_c,GAnd input and output component weight vector parameter alpha_G、β_GThen processing to obtain training data of a Gaussian part with common characteristics;

and T is obtained by processing the training data of Gaussian part with common characteristics_c ²Statistical limit of statistics

4, d) pre-multiplying the input variable of the data to be detected by p_c,GObtaining a second fault estimation vector of the data to be detected, and obtaining T of the data to be detected by using the second fault estimation vector in combination with a partial least square method_c ²Statistics are obtained.

5. The hierarchical non-Gaussian process monitoring method based on common denominator feature extraction as claimed in claim 4, characterized in that:

in the step 4.c), the following iterative algorithm is used for solving and obtaining the input and output public weight vector p in the Gaussian partial data_c,G、q_c,GAnd input and output component weight vector parameter alpha_G、β_G：

(4.c.1) randomly initializing input and output component weight vector parameter alpha in Gaussian partial data_G、β_G；

(4.c.2) inputting and outputting weight vector parameter alpha of input and output components in Gaussian part data_G、β_GThe classified input variable data and the classified output variable data are brought in to obtain weighted input data

And output data

(4, c.3) solving input and output public weight vectors p in Gaussian part data by adopting industrial production process fault monitoring method based on hierarchical non-Gaussian monitoring algorithm_c,GAnd q is_c,G；

(4.c.4) inputting and outputting the public weight vector p in the Gaussian part data obtained in the last step_c,GAnd q is_c,GIs brought into a second objective function;

(4, c.5) solving the weight vector parameter alpha of the input and output components in the Gaussian part data by using a particle swarm algorithm_G、β_G；

(4, c.6) repeating the processes of steps (2) - (6) until alpha_G、β_GAnd p_c,G、q_c,GAnd (6) converging.

6. The hierarchical non-gaussian process monitoring method based on common and unique feature extraction as claimed in claim 1, wherein:

in the step 5), the concrete steps are as follows:

a) when the common weight vector p_cAfter the determination, the training data of each modality with the characteristic features remaining is obtained by processing in the following way:

X_I(i)＝X_(i)-X_(i)*p_cp_c ^T

wherein, X_I(i)Input variable data set representing training data of the i-th modality with characteristic features, Y_I(i)An output variable data set representing training data of the ith modality with characteristic features, X_(i)Input variable data set, Y, representing training data for the ith modality_(i)Representing the ith modalityAn output variable data set of training data;

5, b) aiming at the training data of each mode with the characteristic features, processing the training data by adopting an industrial production process fault monitoring method based on a layered non-Gaussian monitoring algorithm to respectively obtain load matrixes aiming at the non-Gaussian part and the Gaussian part of the training data, and processing the load matrixes to obtain the load matrixes

Statistical limit corresponding to statistical quantity

C) removing the obtained data of the non-Gaussian part from the training data with the characteristic features, remaining the training data of the Gaussian part with the characteristic features, processing the training data of the Gaussian part with the characteristic features by adopting an industrial production process fault monitoring method based on a hierarchical non-Gaussian monitoring algorithm to respectively obtain load matrixes of the Gaussian part and the Gaussian part of the training data, and processing the load matrixes to obtain T_i ²Statistical limits corresponding to the statistics and Q statistics

And a statistical limit Q_{_limit}；

5, d) pre-multiplying the data to be detected with the Gaussian part with the characteristic by the training data of the Gaussian part with the characteristic to obtain a load matrix, and pre-multiplying the input variable of the data to be detected with the Gaussian part by the load matrix P_GObtaining a fourth fault estimation vector of the data to be detected, and obtaining the data to be detected T by using the fourth fault estimation vector_i ²Statistics and Q statistics.

7. The hierarchical non-gaussian process monitoring method based on common and unique feature extraction as claimed in claim 1, wherein:

in the step 6), the concrete steps are as follows:

the data to be detected starts to operate, modal conditions under M modals are totally generated in the industrial production process, the data to be detected of each known modality are monitored to obtain each statistical limit, the data to be detected of each current unknown modality are monitored to obtain each statistic, and the following judgment is carried out:

if the current unknown mode of the data to be detected

Statistic-

The statistics being obtained while being higher than the corresponding training data of the first modality

Statistical limit of statistics

Or

Statistical limit of statistics

The data to be detected in the current unknown mode is fault data;

if the current unknown mode of the data to be detected

Statistic-

Obtained with training data having statistics lower than the corresponding first modality

Statistical limit of statistics

Or

Statistical limit of statistics

Then judge

statistic/T_i ²Whether statistic/Q statistic is simultaneously higher than corresponding

Statistical limit of statisticsT_i ²Statistical limit of statistics

And statistical limit Q of Q statistic_{_limit}：

If not, the data to be detected of the current unknown mode belongs to a first mode;

if the current mode is higher than the preset threshold, the training data of the second mode is called for judgment, and the data to be detected of the current unknown mode is judged

statistic/T_i ²Whether the statistic/Q statistic is simultaneously higher than the training data of the second modality

Statistical limit of statistics

T_i ²Statistical limit of

And statistical limit of Q_{_limit}：

If not, the data to be detected in the current unknown mode belongs to a second mode;

if the current modal is higher than the preset threshold, training data of a third modal is called for judgment, and the current unknown modal to-be-detected data is judged

statistic/T_i ²Whether the statistic/Q statistic is simultaneously higher than the corresponding training data of the third modality

Statistical limit of statistics

T_i ²Statistical limit of statistics

And statistical limit Q of Q statistic_{_limit}：

If not, the data to be detected in the current unknown mode belongs to a third mode;

if the current modal is higher than the preset threshold, the training data of the fourth modal is called for judgment, and the like is repeated until the current unknown modal to-be-detected data

statistic/T_i ²The statistic/Q statistic being higher than that obtained by modal training data

Statistical limit of statistics

T_i ²Statistical limit of statistics

And statistical limit Q of Q statistic_{_limit}And considering the data to be detected in the current unknown mode as fault data.

8. The hierarchical non-gaussian process monitoring method based on common and unique feature extraction as claimed in claim 1, wherein:

and step 7) is also included after the step 6), specifically, the fault detection is carried out by utilizing the statistical limits and statistics in the step 3), the step 4) and the step 5).

9. The hierarchical non-gaussian process monitoring method based on common and unique feature extraction as claimed in claim 1, wherein:

the step 7), specifically, the statistical limit and the statistical quantity in the step 3), 4), 5) meet the data to be detected corresponding to any one of the following five conditions, namely, the industrial production process is considered to have a process fault:

for data of non-Gaussian part with common features in the data to be detected, the data is above the statistical limit

Data to be detected corresponding to the statistic;

for the data of Gaussian part with common features in the data to be detected, the data is above the statistical limit

Data to be detected corresponding to the statistic;

for data of non-Gaussian part with characteristic features in the data to be detected, the data is located above the statistical limit

Data to be detected corresponding to the statistic;

for the data of Gaussian part with characteristic features in the data to be detected, the data is positioned at T above the statistical limit_i ²And the statistic and the data to be detected corresponding to the Q statistic.

Images