A chemical process fault classification method based on optimized stacked autoencoder network

Abstract

The present invention discloses a chemical process fault classification method based on an optimized stack autoencoder network, and belongs to the technical field of fault classification. The method uses labeled and unlabeled samples in the self-supervised learning training stage, and optimizes the stack autoencoder network in combination with the Fisher discriminant criterion, finds a mapping direction that is conducive to classification, reduces the intra-class distance of similar fault features layer by layer, and increases the inter-class distance of heterogeneous features. Through the training of the improved loss function constraint model, the neuron parameters updated by the stack autoencoder network during back propagation can minimize the reconstruction error, so that the stack autoencoder network extracts reconstruction features from a large number of unlabeled samples, and takes into account the use of label information, so that the stack autoencoder network extracts as many classification features as possible. Therefore, the fault classification method based on the optimized stack autoencoder network proposed in the present invention can learn effective classification feature information and improve the accuracy of fault classification.

Description

Technical Field

[0001] The invention relates to a chemical process fault classification method based on an optimized stack autoencoder network, and belongs to the technical field of fault classification. Background Art

[0002] In modern process industries, with the advancement and development of science and technology, chemical processes are becoming more and more complex. In order to ensure the safety and reliability of production processes, fault diagnosis has received more and more attention. In order to provide effective fault solutions, it is necessary to determine the fault category after diagnosing the fault in the chemical process. The complexity of chemical processes poses a huge challenge to fault classification.

[0003] Among the existing chemical process classification methods, the chemical process fault classification method based on stacked autoencoder network adopts self-supervised learning in the feature extraction stage to update the parameters in the deep neural network by minimizing the loss function. Its training goal is to minimize the reconstruction error loss function. However, since only the reconstruction features are considered, the information of the category label is not considered. Therefore, the features extracted by self-supervised learning may contain reconstruction information that is irrelevant to the classification results, which will affect the fault classification results and lead to inaccurate classification results. Summary of the invention

[0004] In order to solve the problems existing in the existing chemical process fault classification method based on stacked autoencoder network and further improve the accuracy of fault classification, the present invention provides a chemical process fault classification method based on optimized stacked autoencoder network. The traditional chemical process fault classification method of stacked autoencoder network does not use label information in the self-supervised learning training stage, resulting in the lack of label information constraint in the layer-by-layer mapping of the deep neural network and the poor identifiability of the extracted reconstruction feature information category. The present invention proposes a stacked autoencoder network fault classification method based on Fisher discriminant criterion optimization. In the self-supervised learning training stage, labeled and unlabeled samples are used to optimize the stacked autoencoder network in combination with the Fisher discriminant criterion to find a mapping direction that is conducive to classification, reduce the intra-class distance of similar fault features layer by layer, and increase the inter-class distance of heterogeneous features. Through the training of the improved loss function constraint model, the neuron parameters updated by the deep neural network during back propagation can minimize the reconstruction error, so that the stacked autoencoder network extracts reconstruction features from a large number of unlabeled samples, and considers the use of label information, so that the stacked autoencoder network extracts as many classification features as possible. Therefore, the fault classification method based on the optimized stacked autoencoder network proposed in the present invention can learn effective classification feature information and improve the accuracy of fault classification; the method includes:

[0005] Step 1: Collect fault variable parameters in the chemical process to form a training set X train={x1,x2,…,x p ,…,x m}, 1≤p≤m, m represents the number of samples; each sample contains J measurement variables, x p ={X 1 ,X 2 ,…,X k ,…,X J}; The measured variables include flow, pressure, temperature and liquid level;

[0006] Step 2: construct a stacked autoencoder network model and optimize it using the improved Fisher discriminant criterion to obtain an optimized stacked autoencoder network model;

[0007] Step 3, after adding labels to the training set collected in step 1, the optimized stacked autoencoder network model obtained in step 2 is trained to obtain a trained stacked autoencoder network model;

[0008] Step 4: Use the trained stacked autoencoder network model obtained in step 3 to implement fault classification for the chemical process.

[0009] Optionally, when the stacked autoencoder network model is optimized using the Fisher discriminant criterion in step 2, the error loss function of the stacked autoencoder network model and the loss function of the Fisher discriminant criterion are reconstructed, and the output of the stacked autoencoder network intermediate layer learning is constrained.

[0010] Optionally, constructing the stacked autoencoder network model in step 2 includes:

[0011] Using training set data X train , train the first autoencoder network and obtain the weight matrix and the bias vector And the hidden layer output

[0012] Next, we use h1 as input to train the second autoencoder network model, which is the same as training the first autoencoder network, and obtain the weight matrix and the bias vector And the hidden layer output

[0013] Repeat the above process until the stacked autoencoder network is built and the parameters are initialized;

[0014] Finally, the stack of n-layer autoencoder networks is built. The relationship between the input layer and the middle layer of the autoencoder network is expressed as:

[0015]

[0016] in is the total weight matrix of the encoder, is the entire bias vector of the encoder; x is the input layer data, h is the middle layer of the stacked autoencoder and the encoder output, f n and f n-1 They represent the activation functions of the nth layer and the n-1th layer in the stacked autoencoder respectively;

[0017] Finally, the relationship between the decoder output and the intermediate layer of the n-layer stacked autoencoder network is expressed as:

[0018]

[0019] in is the total weight matrix of the decoder, is the total bias vector of the decoder; Output data for the decoder, that is, reconstructed data.

[0020] Optionally, the reconstruction error function of the n-layer stacked autoencoder network model constructed in step 2 is:

[0021]

[0022] P(x) represents the theoretical probability density distribution function of the input data:

[0023]

[0024] where δ(·) is the Dirac Delta function, x k ∈R J×1 is the kth sampled variable, R J×1 represents the real number space of J×1 dimension;

[0025] According to the selectivity of the Dirac Delta function:

[0026]

[0027] Where t0 represents the intermediate derived variable;

[0028] The following loss function is derived:

[0029]

[0030] Further simplified to:

[0031]

[0032] The error loss function shown in the above formula (8) is used as the first condition for constraining the output of the intermediate layer learning of the stacked autoencoder network.

[0033] Optionally, in step 2, the stacked autoencoder network is optimized using an improved Fisher discriminant criterion, including:

[0034] The training set X consisting of the fault variable parameters in the chemical process collected in step 1 train The m samples in are divided into c categories according to the fault category, X = {X1, X2, ..., X c}There are c categories, the i-th category X i Contains m i samples, m1+m2+…+m i +…m c =m, 1≤i≤c;

[0035] Calculate the intra-class distance S of the sample w :

[0036]

[0037] in

[0038] Calculate the inter-class distance S of the sample b :

[0039]

[0040] in

[0041] The partial loss function of the Fisher discriminant criterion is defined as:

[0042]

[0043] The partial loss function shown in the above formula (11) is used as the second condition for constraining the output of the intermediate layer learning of the stacked autoencoder network;

[0044] The loss function of the optimized stacked autoencoder network model is defined as:

[0045] L{W e ,b e ,W d ,b d =L f {W e ,b e ,W d ,b d}+L r {W e ,b e ,W d ,b d} (12).

[0046] Optionally, the step three includes:

[0047] After self-supervised learning training of the constructed n-layer stacked autoencoder network model according to the loss function of the stacked autoencoder network model optimized by the above definition, a fully connected layer and a Softmax classification layer are added to the encoder of the stacked autoencoder network to form a classifier, and the classifier is trained using labeled samples and label information;

[0048] For a given input X i , the classification layer output is:

[0049]

[0050] where θ1,θ2,...,θ c is the parameter of the classification layer model, denoted as θ = [θ1, θ2, ..., θ c ] T , It is the normalization of the probability distribution of classification output; the training of the classifier part uses the multi-classification cross entropy loss function:

[0051]

[0052] Taking minimizing the loss function as the optimization goal, the classifier is trained with supervised learning as a whole to obtain the final fault classification result.

[0053] Optionally, the step 1 further includes standardizing the samples to obtain standardized training set data:

[0054]

[0055] Among them, mean(X train ) is the mean of the training set, std(X train ) is the standard deviation of the training set.

[0056] Optionally, the step 4 includes:

[0057] For specific chemical processes, collect corresponding measurement variables;

[0058] The collected measurement variables are input into the trained classifier to obtain the corresponding fault classification results.

[0059] The present application also provides the application of the above fault classification method in a chemical production process.

[0060] The beneficial effects of the present invention are:

[0061] The present application discloses a fault classification method based on an optimized stacked autoencoder network, which uses the Fisher discriminant criterion to optimize the stacked autoencoder model to obtain a powerful deep feature learning ability, better learn the potential distribution law of the data, is suitable for processing chemical process data, and has a good fault classification effect;

[0062] This application integrates the feature learning method of finding the best projection direction in the Fisher discriminant criterion into the self-supervised learning training of the stacked autoencoder network, and uses the sample label information to learn the best projection direction in the layer-by-layer nonlinear mapping of the stacked autoencoder network. Self-supervised learning training under the constraint of the loss function optimized by the Fisher discriminant criterion increases the inter-class distance of fault features of different categories and reduces the intra-class distance of features of the same category. Because the constraints of minimizing the reconstruction features and maximizing the classification features are designed at the same time in the self-supervised learning training of the optimized stacked autoencoder network, the trained stacked autoencoder network can extract more effective feature information to improve the accuracy of the final fault classification;

[0063] This application can effectively perform automatic feature extraction and fault classification on chemical process signals. The constructed deep autoencoder network can automatically learn low-level features and gradually form more abstract high-level representations, and finally directly output the chemical process fault category, thus realizing end-to-end chemical process fault classification to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0065] Figure 1 It is a flow chart of a chemical process fault classification method based on an optimized stack autoencoder network provided in this application.

[0066] Figure 2 It is a structural diagram of a stacked autoencoder network constructed in a chemical process fault classification method based on an optimized stacked autoencoder network provided in the present application.

[0067] Figure 3 This is a feature visualization diagram obtained by using a traditional stacked autoencoder network for fault classification.

[0068] Figure 4 This is a feature visualization diagram obtained by using the optimized stacked autoencoder network provided in this application to perform fault classification. DETAILED DESCRIPTION

[0069] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0070] Embodiment 1:

[0071] This embodiment provides a chemical process fault classification method based on an optimized stacked autoencoder network. Figure 1 , the method comprising:

[0072] Step 1: Use the measuring instruments installed in chemical production equipment to collect chemical process samples to form a training set X train ={x1,x2,…,x p ,…,x m}, 1≤p≤m, m represents the number of samples; each sample contains J measurement variables such as flow, pressure, temperature, and liquid level, that is, x p ={X 1 ,X 2 ,…,X k ,…,X J}, 1≤k≤J; the measured variables may be different for different chemical production.

[0073] Step 2: Standardize the samples, as shown in formula (1);

[0074]

[0075] Among them, mean(X train ) is the mean of the training set, std(X train ) is the standard deviation of the training set;

[0076] Step 3: Build a stacked autoencoder network model;

[0077] The stacked autoencoder network structure is composed of multiple single-hidden-layer autoencoder networks stacked together, and the parameters of the neural network are initialized layer by layer through self-supervised learning training.

[0078] First, use the training set data X train , train the first autoencoder network and obtain the weight matrix and the bias vector And the hidden layer output σ represents the activation function Rule.

[0079] Next, we use h1 as input to train the second autoencoder network model, which is the same as training the first autoencoder network, and obtain the weight matrix and the bias vector And the hidden layer output

[0080] Repeat the above process until the stacked autoencoder network is built and the parameters are initialized.

[0081] Finally, the stack of n-layer autoencoder networks is built. The relationship between the input layer and the middle layer of the autoencoder network can be expressed as:

[0082]

[0083] in is the total weight matrix of the encoder, is the entire bias vector of the encoder; x is the input layer data, h is the middle layer of the stacked autoencoder and the encoder output, f n and f n-1 They represent the activation functions of the nth layer and the n-1th layer in the stacked autoencoder respectively.

[0084] The relationship between the decoder output and the intermediate layer of the n-layer stacked autoencoder network can be expressed as:

[0085]

[0086] in is the total weight matrix of the decoder, is the total bias vector of the decoder; Output data for the decoder, that is, reconstructed data.

[0087] Ideally, the goal of self-supervised training of a stacked autoencoder network is to minimize the reconstruction error:

[0088]

[0089] P(x) represents the theoretical probability density distribution function of the input data. In actual industrial processes, the input data is a collection of finite sampling samples.

[0090]

[0091] where δ(·) is the Dirac Delta function, x k ∈R J×1 is the kth sampled variable, R J×1 represents a J×1-dimensional real number space, and m represents the number of sampling samples.

[0092] According to the selectivity of the Dirac Delta function:

[0093]

[0094] Where t0 represents the intermediate derived variable.

[0095] The following loss function is derived:

[0096]

[0097] In the subsequent back propagation calculation of the deep neural network, for the convenience of derivation, the loss function shown in formula (7) is usually rewritten as:

[0098]

[0099] Minimizing this loss function is the goal of self-supervised learning training of traditional stacked autoencoder networks.

[0100] In order to make the features extracted from the stacked autoencoder network contain both reconstruction information and classification information, the present application constructs a new loss function in combination with the Fisher discriminant criterion, and trains the stacked autoencoder network together with the reconstruction error.

[0101] Step 4: Use Fisher discriminant criterion to optimize stacked autoencoder network feature learning;

[0102] The application of Fisher discriminant criterion in layer-by-layer nonlinear mapping of deep neural networks is to find the nonlinear projection direction that minimizes the intra-class distance and maximizes the inter-class distance.

[0103] Suppose there is a labeled modeling dataset X = {x1, x2, ..., x m} consists of m samples, where m samples are divided into c categories, X = {X1, X2, ..., X c}There are c categories, the i-th category X i Contains m i samples, m1+m2+…+m i +…m c =m.

[0104] Calculate the intra-class distance S of the sample w :

[0105]

[0106] in

[0107] Calculate the inter-class distance S of the sample b :

[0108]

[0109] in

[0110] The partial loss function of the Fisher discriminant criterion is defined as:

[0111]

[0112] The loss function of the reconstruction function is defined as:

[0113] L{W e ,b e ,W d ,b d =L f {W e ,b e ,W d ,b d}+L r {W e ,b e ,W d ,b d} (12)

[0114] In the training of the stacked autoencoder network, the present application simultaneously constrains the output of the stacked autoencoder network intermediate layer learning by reconstructing the error loss function and the loss function of the Fisher discriminant criterion.

[0115] After the back-propagation algorithm updates all weight matrices and bias vectors in the stacked autoencoder network, it is ensured that the learned features can both maximize the reconstruction of the original input and obtain effective classification features that minimize the intra-class distance and maximize the inter-class distance.

[0116] Step 5: Combine the feature information extracted by the optimized stacked autoencoder network model with the Softmax classifier to ultimately achieve fault classification of the chemical process.

[0117] After building the optimized stacked autoencoder network model and completing the self-supervised learning training, a fully connected layer and a Softmax classification layer are added to the encoder of the stacked autoencoder network to form a classifier, and the classifier is trained using labeled samples and label information.

[0118] For a given input X i , the classification layer output is:

[0119]

[0120] where θ1,θ2,...,θ c is the parameter of the classification layer model, denoted as θ = [θ1, θ2, ..., θ c ] T , It is to normalize the probability distribution of classification output, so that the sum of the probabilities of all output categories is 1. The training of the classifier part uses the multi-classification cross entropy loss function:

[0121]

[0122] Taking minimizing the loss function as the optimization goal, the classifier is trained with supervised learning as a whole to obtain the final fault classification result.

[0123] Embodiment 2:

[0124] This embodiment provides a specific implementation process of a chemical process fault classification method based on an optimized stacked autoencoder network, see Figure 1 The specific implementation process of the method includes two parts: offline modeling and online classification:

[0125] Offline Modeling:

[0126] 1) Data preprocessing: Get the training set X train , for standardization;

[0127] 2) Parameter initialization: Design the stacked autoencoder network structure in the chemical process fault classification method based on the optimized stacked autoencoder network proposed in this application, including the number of hidden layers, the number of neurons in each layer, and randomly initialize the connection parameters between layers, i.e., the stacked autoencoder network weight matrix, the weight matrix of the fully connected layer, and the bias vector;

[0128] 3) Feature extraction: Using all historical samples and corresponding label information, the stacked autoencoder network obtained in the above steps is pre-trained using the back-propagation algorithm to minimize the loss function, update the connection parameters, and extract feature information;

[0129] 4) Classifier training: The extracted feature information is fed into the classifier to minimize the multi-classification cross entropy loss function, complete the training of the classifier parameters and fine-tune the connection parameters of the encoder part in the stacked autoencoder network.

[0130] Online Classification:

[0131] 1) Obtain online samples and perform standardization;

[0132] 2) Feature extraction: extracting features from each online sample using the chemical process fault classification method based on the optimized stacked autoencoder network described in Example 1;

[0133] 3) Fault classification: The feature information extracted from each online sample is sent to the classifier to identify the fault type based on the feature information.

[0134] Combination Figure 2As shown, the chemical process fault classification method based on the optimized stack autoencoder network proposed in this application is mainly divided into two parts. The first part is to realize feature extraction based on the data collected in real time in the chemical process. This part improves the stack autoencoder network model based on the Fisher discriminant criterion and obtains classification feature information through self-supervised learning training; the second part is to train the FSAE classifier through supervised learning to realize fault classification.

[0135] The data set of this application is derived from 4 types of fault samples in the TE chemical process, including 3 types of step faults and 1 type of random variable fault. The specific fault types are shown in Table 1. There are 480 fault samples of each type in the training set and 800 fault samples of each type in the test set.

[0136] The TE process is a simulation of an actual chemical process. It is an open and challenging chemical model simulation platform developed by Eastman Chemical Company of the United States - Tennessee Eastman (TE) simulation platform, which can realize the simulation of any complex industrial process. The TE process mainly involves five main units: reactor, condenser, compressor, separator and stripping tower; when simulating the actual chemical process, it can generate parameter data in the actual chemical process to verify the fault classification method of this application; for the specific introduction of the TE process, please refer to the introduction of Baidu entry, and this application will not be further introduced in detail; the specific fault type information given in Table 1 below in this application is the simulation data of the actual chemical process generated by the TE process;

[0137] Table 1 TE process fault type table

[0138]

[0139] In order to verify that the chemical process fault classification method based on the optimized stacked autoencoder network proposed in the present invention can effectively utilize label information to improve the feature extraction ability of the stacked autoencoder network in the feature extraction stage, feature visualization is used to verify the effect of feature learning, and compared with the traditional stacked autoencoder network (Stacked Autoencoder, SAE) method.

[0140] like Figure 3 As shown, the features extracted by the traditional stacked autoencoder network model are visualized. Among them, fault 4 and fault 12 are mixed together, and the interval between classes is small and almost indistinguishable.

[0141] The fault feature information extracted by the stacked autoencoder network model based on the Fisher discriminant criterion optimization proposed in the present invention is as follows: Figure 4As shown, the classification results can be clearly displayed in the visualization, and the intervals between classes are relatively obvious. This is because the fault classification method proposed in this application minimizes the intra-class distance and maximizes the inter-class distance during training, making it easier for the extracted effective features to find the class center.

[0142] In order to compare the advantages and disadvantages of the present invention with the current mainstream intelligent fault diagnosis algorithms, representative support vector machine (SVM), Fisher discriminant method (Fisher) and stacked autoencoder network (SAE) were selected, and each method was trained with the same training set and tested with the same test set. The evaluation index of the test results is the classification accuracy, and the test results are shown in Table 2.

[0143] Table 2 Comparison of TE process fault classification accuracy

[0144]

[0145] Fault 4 is a change in the reactor cooling water inlet temperature, a step type fault, and fault 12 is a change in the condenser cooling water inlet temperature, a random variable fault. The classification accuracy of the SVM method for these two types of faults is poor, 78.0% and 60.5% respectively, and it fails to distinguish the two types of faults well. The diagnostic accuracy of the Fisher discriminant method for faults 4 and 12 has improved to 95.6% and 86.4% respectively. Compared with the SAE method and the method proposed in this application, since the traditional SAE method only considers the information reconstruction of unlabeled samples and the corresponding reconstruction error in the feature extraction stage, the classification accuracy of these two types of faults is relatively low, with a classification accuracy of 84.6% for fault 4 and only 34.1% for fault 12.

[0146] The chemical process fault classification method based on the optimized stack autoencoder network proposed in this application takes into account the label information in the feature extraction stage, and optimizes the traditional stack autoencoder network in combination with the Fisher discriminant criterion. The Fisher discriminant criterion is used to consider the projection direction of the classification feature, find the feature mapping direction that is conducive to classification, add classification features, and remove the features in the reconstruction information that are irrelevant to the classification results, thereby improving the performance of fault classification. Through the training of the improved reconstruction error loss function and the Fisher discriminant loss function constraint model, the neuron parameters updated by the stack autoencoder network during back propagation can not only minimize the reconstruction error, so that the stack autoencoder network can extract reconstruction features from a large number of unlabeled samples, but also take into account the use of label information, so that the stack autoencoder network can extract as many classification features as possible. Therefore, on fault 4, the fault classification accuracy rate increased by 15.2%, and on fault 12, it increased by 64.7%, demonstrating the effectiveness of the optimized stack autoencoder network fault classification proposed by the present invention.

[0147] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0148] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A chemical process fault classification method based on an optimized stacked autoencoder network, characterized in that: The method comprises: Step 1: Collect fault variable parameters in the chemical process to form a training set X train ={x1,x2,…,x p ,…,x m }, 1≤p≤m, m represents the number of samples; each sample contains J measurement variables, x p ={X 1 ,X 2 ,…,X k ,…,X J }; The measured variables include flow, pressure, temperature and liquid level; Step 2: construct a stacked autoencoder network model and optimize it using the improved Fisher discriminant criterion to obtain an optimized stacked autoencoder network model; Step 3, after adding labels to the training set collected in step 1, the optimized stacked autoencoder network model obtained in step 2 is trained to obtain a trained stacked autoencoder network model; Step 4: Using the trained stacked autoencoder network model obtained in step 3 to implement fault classification for the chemical process; When the stacked autoencoder network model is optimized using the Fisher discriminant criterion in the step 2, the error loss function of the stacked autoencoder network model and the loss function of the Fisher discriminant criterion are reconstructed, and the output of the stacked autoencoder network intermediate layer learning is constrained at the same time; The step 2 of constructing the stacked autoencoder network model includes: Using training set data X train , train the first autoencoder network and obtain the weight matrix and the bias vector And the hidden layer output Next, we use h1 as input to train the second autoencoder network model, which is the same as training the first autoencoder network, and obtain the weight matrix and the bias vector And the hidden layer output Repeat the above process until the stacked autoencoder network is built and the parameters are initialized; Finally, the stack of n-layer autoencoder networks is built. The relationship between the input layer and the middle layer of the autoencoder network is expressed as: in is the total weight matrix of the encoder, is the entire bias vector of the encoder; x is the input layer data, h is the middle layer of the stacked autoencoder and the encoder output, f n and f n-1 They represent the activation functions of the nth layer and the n-1th layer in the stacked autoencoder respectively; Finally, the relationship between the decoder output and the intermediate layer of the n-layer stacked autoencoder network is expressed as: in is the total weight matrix of the decoder, is the total bias vector of the decoder; Output data for the decoder, i.e., reconstructed data; In the step 2, the stacked autoencoder network is optimized using the improved Fisher discriminant criterion, including: The training set X consisting of the fault variable parameters in the chemical process collected in step 1 train The m samples in are divided into c categories according to the fault category, X = {X1, X2, ..., X c }There are c categories, the i-th category X i Contains m i samples, m1+m2+…+m i +…m c =m, 1≤i≤c; Calculate the intra-class distance S of the sample w : in Calculate the inter-class distance S of the sample b : in The partial loss function of the Fisher discriminant criterion is defined as: The partial loss function shown in the above formula (11) is used as the second condition for constraining the output of the intermediate layer learning of the stacked autoencoder network; The loss function of the optimized stacked autoencoder network model is defined as: L{W e ,b e ,W d ,b d }=L f {W e ,b e ,W d ,b d }+L r {W e ,b e ,W d ,b d }(12) Where L r {W e ,b e ,W d ,b d } is the reconstruction error function of the n-layer stacked autoencoder network model.

2. The method according to claim 1, characterized in that The reconstruction error function of the n-layer stacked autoencoder network model constructed in step 2 is: P(x) represents the theoretical probability density distribution function of the input data: where δ(·) is the Dirac Delta function, x k ∈R J×1 is the kth sampled variable, R J×1 represents the real number space of J×1 dimension; According to the selectivity of the Dirac Delta function: Where t0 represents the intermediate derived variable; The following loss function is derived: Further simplified to: The error loss function shown in the above formula (8) is used as the first condition for constraining the output of the intermediate layer learning of the stacked autoencoder network.

3. The method according to claim 2, characterized in that The step three comprises: After self-supervised learning training of the constructed n-layer stacked autoencoder network model according to the loss function of the stacked autoencoder network model optimized by the above definition, a fully connected layer and a Softmax classification layer are added to the encoder of the stacked autoencoder network to form a classifier, and the classifier is trained using labeled samples and label information; For a given input X i , the classification layer output is: where θ1,θ2,...,θ c is the parameter of the classification layer model, denoted as θ = [θ1, θ2, ..., θ c ] T , It is the normalization of the probability distribution of classification output; the training of the classifier part uses the multi-classification cross entropy loss function: Taking minimizing the loss function as the optimization goal, the classifier is trained with supervised learning as a whole to obtain the final fault classification result.

4. The method according to claim 3, characterized in that The step 1 also includes standardizing the samples to obtain the standardized training set data: Among them, mean(X train ) is the mean of the training set, std(X train ) is the standard deviation of the training set.

5. The method according to claim 4, characterized in that The fourth step comprises: For specific chemical processes, collect corresponding measurement variables; The collected measurement variables are input into the trained classifier to obtain the corresponding fault classification results.

6. Application of the fault classification method described in any one of claims 1 to 5 in a chemical production process.

Images