Complex chemical process fault diagnosis method based on ensemble learning
By integrating principal component analysis with Bayes combined with classifier, convolutional neural network classifier and graph neural network classifier, the problem of insufficient diagnostic accuracy of multiple fault types in complex chemical processes is solved, and high-precision fault diagnosis is achieved.
Patent Information
- Application Number
- CN202510430539.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-08
- Publication Date
- 2025-06-27
AI Technical Summary
The prior art is difficult to effectively diagnose various types of faults in complex chemical processes, resulting in insufficient diagnostic accuracy.
An integrated learning-based method is adopted, combining principal component analysis with Bayes combined with classifiers, convolutional neural network classifiers and graph neural network classifiers, and integrating these weak learners to improve fault recognition capabilities and overall diagnostic accuracy.
It realizes high-precision diagnosis of various types of faults in complex chemical processes, improving fault recognition capabilities and overall diagnostic accuracy.
Smart Images
Figure CN120215470A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of chemical process fault diagnosis, and particularly to a fault diagnosis method for complex chemical processes based on ensemble learning. Background Art
[0002] Large-scale chemical process systems are characterized by high coupling, complexity, and poor system robustness. Fault detection and diagnosis in chemical processes are of great significance for the reliability and safety of modern chemical systems. Currently, studying whether process variables in chemical processes deviate from normal states is the mainstream method for chemical fault diagnosis. Methods such as principal component analysis and dynamic principal component analysis are widely used. The combined model of principal component analysis and Bayesian can reduce the feature dimension and simplify the calculation; the convolutional neural network model can make full use of frequency domain information, reduce data noise, and improve accuracy; the graph neural network model with its unique topological structure can represent the causal relationship of chemical processes and make full use of the spatial information of chemical processes. However, all three methods are only sensitive to single or partial faults and cannot meet the accuracy requirements for multiple faults. Therefore, an ensemble learning framework is used to integrate three weak learners, namely the combined learner of principal component analysis and Bayesian, the convolutional neural network learner, and the graph neural network learner, to improve the fault recognition ability and overall fault diagnosis accuracy. The present invention aims to propose a fault diagnosis method for complex chemical processes based on ensemble learning. Summary of the Invention
[0003] Aiming at the problem that it is difficult to judge various fault types occurring in chemical processes, the present invention proposes a fault diagnosis method for complex chemical processes based on ensemble learning, which can give relatively high diagnosis accuracy in a timely manner for various fault types.
[0004] The present invention achieves the above object through the following technical solutions:
[0005] Sample the sensor data and perform analog-to-digital conversion, upload the sensor sampling data to the cloud server, and preprocess the data set and divide it into non-overlapping subsets: , where represents the category of the fault.
[0006] This ensemble learning algorithm selects three weak learners, namely the combined classifier of principal component analysis and Bayesian, the convolutional neural network classifier, and the graph neural network classifier.
[0007] Initialize the sample weights. For the initial weights of samples containing samples, they are:
[0008] ,
[0009] Train weak classifiers, and iterate the entire algorithm process times. The number of iterations is also the number of base learners. During the iteration process, perform fault category prediction, and record the prediction result as , where represents the th base learner, ranges from . Record the classification error probability as:
[0010] ϵ t = Σ i =1 N [ h t ( x i )≠ y i ] n
[0011] If , it means that the sample is misclassified at this time, and the weight does not change:
[0012]
[0013] Record the weight of the weak classifier as .
[0014] If the sample is correctly classified, the weight becomes:
[0015] D t +1 (i)= D t (i)∙ β t 1-[ h t x i ≠ y i ]
[0016] ranges from , and normalize the sample weights. The samples with adjusted weights are used for the training of the th weak classifier in the next iteration.
[0017] After weak classifiers are trained, combine each weak classifier to obtain a strong classifier, and record the classification result of the strong classifier as:
[0018]
[0019] Among them, , represents the fault category.
[0020] After deploying the trained classifier, the data is passed into the combined classifier of principal component analysis and Bayesian of the first weak learner for fault diagnosis. Preprocess the -dimensional sensor data. Assume the sensor data is , and the number of samples is .
[0021] Calculate the covariance matrix of the training set samples:
[0021]
[0023] Furthermore, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors:
[0024] Y = U Λ U T = P, P ¯ Λ [P, P ¯ ]
[0025] where is a diagonal matrix.
[0026] Sort the eigenvalues from largest to smallest, and select the first eigenvectors to form the loading matrix , and the remaining -dimensional matrix is used as the residual loading matrix .
[0027] Denote the principal subspace as , the residual subspace as , and the score matrix as .
[0028] Calculate the statistic and control limits of the test set samples. If the system is operating normally, the sample value is less than the control limit. If the system fails, the sample value is greater than the control limit. The formula for is:
[0029] The formula for the control limit is:
[0030] .
[0031] Calculate the statistic:
[0032]
[0033] The control limit is:
[0034] Q α = Θ 1 [ C α h 0 2 Θ 2 Θ 1 +1+ Θ 2 h 0 ( h 0 -1) Θ 1 2 ] 1 h 0
[0035] where The confidence limit of the standard normal distribution.
[0036] If the system is operating normally, then The statistic should be less than The control limit. If the system fails, then The statistic should be greater than The control limit.
[0037] When using both The statistic and The control limit and The statistic and The control limit, only And The statistic exceeding the control limit simultaneously represents that the system has a fault.
[0038] When the sample statistic exceeds the control limit, that is, when the system fails, the data reduced in dimension by the principal component analysis method is input into the Bayesian network classifier to calculate the posterior probability corresponding to each category, and the category with the highest probability is selected as the prediction result.
[0039] The data is passed into the second convolutional neural network weak learner for chemical process fault diagnosis, and the time series data is transformed into frequency domain data through the fast Fourier transform.
[0040] Using a convolutional kernel of size Perform a convolution operation on the frequency domain data, where Is the dimension of the sensor data.
[0041] Calculate the output feature map:
[0042]
[0043] Where Is the input, Is the activation function, Is the bias, Is the convolutional kernel, and the activation function is The activation function.
[0044] Group the feature map in channels and perform a normalization operation on each group.
[0045] Change the dimension of the feature map from [N, C, H, W] To [N, G, C / / G, H, W] , and the dimension after normalization is [C / / G, H, W] ;
[0046] When , the group normalization is transformed into layer normalization, and the dimension becomes [C, H, W] ;
[0047] When happens, the group normalization operation becomes instance normalization, and the dimension becomes [H, W] .
[0048] Perform max pooling on the feature map. The role of the pooling operation is to reduce the dimension of the feature map for faster operation speed.
[0049] Finally, pass the feature map into the fully connected layer to integrate and normalize the features, obtain the probabilities of various faults, and get the fault probability distribution according to the fully connected layer and the activation function.
[0050] The data is passed into the third graph neural network weak learner for chemical fault diagnosis.
[0051] Use the K-nearest neighbor graph construction algorithm to model multi-channel sensor data, construct a spatio-temporal relationship causal graph, with time series as the embedding of nodes and chemical reaction relationships as edges.
[0052] The K-nearest neighbor graph construction algorithm means connecting a node to its nearest nodes, and the edges between nodes represent node similarity.
[0053] Input the fault-free state data graph and the fault data graph into training, and use graph convolution operations to extract the corresponding features in the normal state and fault state of the chemical process.
[0054] The convolution uses the graph convolution algorithm, and its propagation method is:
[0055]
[0056] Among them, , A is the adjacency matrix of the graph network, I is the identity matrix, is the degree matrix of, and the formula is
[0057]
[0058] H is the feature of each layer. For the input layer, H is the input X, is the non-linear activation function.
[0059] After the graph convolution operation, it enters the graph pooling layer, aiming to reduce the number of nodes, reduce the computational complexity, and achieve feature dimension reduction and hierarchical learning.
[0060] The graph pooling selects the edge contraction pooling mechanism, calculates a score for each edge, and performs iterative contraction according to the score;
[0061] The fractional value of an edge is calculated through a linear transformation of node features, including the following steps:
[0062] Step 1: Calculate the fractional value of the edge connected to node and node as:
[0063]
[0064] Step 2: Normalize the fractional values of the edges adjacent to the nodes so that the fractional values are controlled within [0,1] for easier comparison of fractional values. The normalized edge fractional values are:
[0065]
[0066] Step 3: Sort the edge fractional values and sequentially contract the nodes with higher rankings. Repeat this operation until 50% of the nodes in the aggregated graph are reached.
[0067] After the graph pooling operation, the feature representations of each node have been obtained. At this time, a readout layer is required to represent the features of the global graph. In the present invention, an averaging operation is used. Let the graph nodes be represented as:
[0068]
[0069] where represents the number of nodes, represents the feature representation of the th node, and the readout layer function is:
[0070]
[0071] represents the feature representation of the entire graph, is the sigmoid activation function.
[0072] Finally, the result of the readout layer is passed into the fully connected layer to integrate and normalize the features, and the fault probability distribution is obtained according to the activation function.
[0073] The final fault classification result is calculated according to the combination of the strong classifier , and the fault type is output through the human-machine interface. Description of the Drawings
[0074] Figure 1 Schematic diagram of the integrated learning structure of the complex chemical process implemented in the present invention.
[0075] Figure 2 Schematic diagram of the process of the classifier combining principal component analysis and Bayesian implemented in the present invention.
[0076] Figure 3 Schematic diagram of the convolutional neural network classifier process implemented by the present invention.
[0077] Figure 4 Schematic diagram of the graph neural network classifier process implemented by the present invention.
[0078] Figure 5 Schematic diagram of the integrated learning process of the complex chemical process implemented by the present invention. Detailed implementation manners
[0079] The present invention will be further described in detail below with reference to the accompanying drawings. The present invention proposes a fault diagnosis method for complex chemical processes based on integrated learning, which can perform online diagnosis and alarm when a fault occurs in a complex chemical process.
[0080] The present invention realizes the above object through the following technical solutions:
[0081] Upload the sensor sampling data to the cloud server, and preprocess the data set and divide it into non-overlapping subsets: , where represents the category of the fault.
[0082] This integrated learning algorithm selects three weak learners, namely the classifier combined with principal component analysis and Bayesian, the convolutional neural network classifier, and the graph neural network classifier.
[0083] Initialize the sample weights. For the initial weights of the samples containing samples, they are: ,
[0084] Train the weak classifier. The entire algorithm process iterates times. The number of iterations is also the number of base learners. During the iteration process, fault category prediction is performed, and the prediction result is denoted as . Denote the classification error probability as:
[0085] ϵ t = Σ i =1 N [ h t ( x i )≠ y i ] n
[0086] If , it means that the sample is misclassified at this time, and the weights do not change:
[0087] . Denote the weight of the weak classifier as .
[0088] If the sample is correctly classified, the weight becomes:
[0089] D t +1 (i)= D t (i)∙ β t 1-[ h t x i ≠ y i ]
[0090] ranges from , and the sample weights are normalized. The samples with adjusted weights are used for the training of the th weak classifier in the next iteration.
[0091] After the training of weak classifiers is completed, the weak classifiers are combined to obtain a strong classifier. Denote the classification result of the strong classifier as:
[0092]
[0093] where , represents the fault category.
[0094] After deploying the trained classifier, the data is sent to the first weak learner combined classifier of principal component analysis and Bayesian for fault diagnosis. The -dimensional sensor data is preprocessed. The sensor data is:
[0095]
[0096] where the number of samples is , and the sample dimension, i.e., the number of sensors, is 52.
[0097] Calculate the covariance matrix of the training set samples:
[0098]
[0099] Furthermore, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors:
[0100] Y = U Λ U T = P, P ¯ Λ [P, P ¯ ]
[0101] where is a diagonal matrix.
[0102] Sort the eigenvalues from largest to smallest and select the top The load matrix consists of eigenvectors , and the remaining -dimensional matrix serves as the residual load matrix .
[0103] Denote the principal component subspace as , the residual subspace as , and the score matrix as .
[0104] Calculate the statistic and control limit for the test set samples. If the system is operating normally, the sample value is less than the control limit. If the system fails, the sample value is greater than the control limit, and the calculation formula for
[0105] The calculation formula for the control limit is:
[0106] .
[0107] Calculate the statistic:
[0108]
[0109] The control limit is:
[0110] Q α = Θ 1 [ C α h 0 2 Θ 2 Θ 1 +1+ Θ 2 h 0 ( h 0 -1) Θ 1 2 ] 1 h 0
[0111] where is the confidence limit of the standard normal distribution.
[0112] If the system is operating normally, then the statistic should be less than the control limit. If the system fails, then the statistic should be greater than the control limit.
[0113] When using both the statistic and the control limit and the statistic and the control limit, only when and the statistics exceed the control limit simultaneously does it represent that the system has failed.
[0114] When the sample statistic exceeds the control limit, that is, when the system fails, the data reduced in dimension by the principal component analysis method is input into the Bayesian network classifier to calculate the posterior probability corresponding to each category, and the category with the highest probability is selected as the prediction result.
[0115] The data is passed into the second convolutional neural network weak learner for chemical process fault diagnosis, and the time series data is transformed into frequency domain data through the fast Fourier transform.
[0116] Use a convolutional kernel of size to perform a convolution operation on the frequency domain data.
[0117] Calculate the output feature map:
[0118]
[0119] where is the input, is the activation function, is the bias, is the convolutional kernel, and the activation function is activation function.
[0120] Group the feature map by channels and perform a normalization operation on each group.
[0121] Convert the dimension of the feature map from [N, C, H, W] to [N, G, C / / G, H, W] , and the dimension after normalization is [C / / G, H, W] ;
[0122] When , the group normalization is converted to layer normalization, and the dimension becomes [C, H, W] ;
[0123] When , the group normalization operation becomes instance normalization, and the dimension becomes [H, W] .
[0124] Perform max pooling on the feature map. The role of the pooling operation is to reduce the dimension and compress the feature map to speed up the operation.
[0125] Finally, pass the feature map into the fully connected layer, integrate and normalize the features, give the probabilities of each fault, and obtain the fault probability distribution according to the fully connected layer plus the activation function.
[0126] The data is passed into the third graph neural network weak learner for chemical process fault diagnosis.
[0127] Model the multi-channel sensor data using the K-nearest neighbor graph construction algorithm to construct a spatio-temporal relationship causal graph, with time series as the embedding of nodes and chemical reaction relationships as edges.
[0128] The K-nearest neighbor graph construction algorithm means connecting a node to its nearest nodes, and the edges between nodes represent node similarity.
[0129] Input the fault-free state data graph and the fault data graph into training, and use graph convolution operations to extract the corresponding features in the normal state and fault state of the chemical process.
[0130] The convolution uses the graph convolution algorithm, and its propagation method is:
[0131]
[0132] Among them, , A is the adjacency matrix of the graph network, I is the identity matrix, is the degree matrix of, and the formula is
[0133]
[0134] H is the feature of each layer. For the input layer, H is the input X, is the non-linear activation function.
[0135] After the graph convolution operation, it enters the graph pooling layer, aiming to reduce the number of nodes, reduce the computational complexity, and achieve feature dimensionality reduction and hierarchical learning.
[0136] The graph pooling selects the edge contraction pooling mechanism, calculates a score for each edge, and performs iterative contraction according to the score;
[0137] The score of the edge is calculated through a linear transformation of the node features, including the following steps:
[0138] Step 1, calculate the score of the edge connected to node and node as:
[0139]
[0140] Step 2, normalize the scores of the adjacent edges of the node to control the scores within [0,1] for easy score comparison. The normalized edge score is:
[0141]
[0142] Step 3: Sort the edge scores, and sequentially contract the nodes with higher rankings. Repeat this operation until 50% of the nodes in the aggregated graph are reached.
[0143] After the graph pooling operation, the feature representations of each node have been obtained. At this time, a readout layer is needed to represent the features of the global graph. In the present invention, an average operation is used. Let the graph nodes be represented as:
[0144]
[0145] where denotes there are nodes, represents the -th node's feature representation, and the readout layer function is:
[0146]
[0147] represents the overall graph feature representation, is the sigmoid activation function.
[0148] Finally, the result of the readout layer is passed into the fully connected layer to integrate and normalize the features, and the fault probability distribution is obtained according to the activation function.
[0149] The final fault classification result is calculated according to the combination of the strong classifier
Claims
1. A complex chemical process fault diagnosis method based on ensemble learning, which can obtain sensor data online, determine whether the data is abnormal and determine the fault type, and the method includes the following steps: Step 1: Use sensors to sample control variables and process variables in complex chemical processes, perform analog-to-digital conversion on the sampled signals, and upload them to a cloud server for storage; Step 2, read the data collected in step 1 and perform data cleaning; Step 3, read the data in step 2, divide the data set into a training set and a test set in proportion, and train an integrated learning chemical fault diagnosis model; Step 4: deploy the integrated learning chemical fault diagnosis model trained in step 3 in the cloud server, repeat steps 1 and 2, obtain real-time data of complex chemical processes and input them into the model, and output the fault type; Step 5: Display the fault type diagnosed by the model in step 4 through the human-computer interaction interface.
2. According to claim 1, a complex chemical process fault diagnosis method based on ensemble learning can determine whether the data is abnormal and pre-process the data, characterized in that: Step 2 includes the following steps: Step 2.1: Fill the missing data using the mean method; Step 2.2: Convert the abnormal data type into numerical data.
3. A complex chemical process fault diagnosis method based on ensemble learning according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Divide the data set S evenly into in Indicates the category of faults, and then divided into training set and test set in a ratio of 7:3; Step 3.2: Read the dataset S in step 2, train three weak classifiers in sequence, namely: principal component analysis and Bayesian combined classifier, convolutional neural network classifier and graph neural network classifier, initialize the sample weights, and train the The initial weight of the samples is: Algorithm process iteration The number of iterations is the number of base learners. Fault category prediction is performed during the iteration process, and the prediction result is recorded as ,in Indicates A base learner, The range is , the probability of classification error is ,like , which means that the sample is classified incorrectly at this time, otherwise it is classified correctly; Step 3.3: Use the ensemble learning algorithm to adaptively learn and modify the sample weights. If the sample is misclassified, the weight remains unchanged: If the sample is classified correctly, the weight becomes: in The range is ,sample Will be used to train Weak classifiers; Step 3.5: After completing the training of all weak classifiers, record the weight of the weak classifier as , the ensemble learning classification result is: in, , Represents the fault category, is the number of iterations.
4. A complex chemical process fault diagnosis method based on ensemble learning according to claim 1, characterized in that: The convolutional neural network weak learner inference process in step 3.2 is: Step 1: Convert the time series data into frequency domain data through fast Fourier transform; Step2: Use the size The convolution kernel performs convolution operation on the frequency domain data, where For the sensor data dimension, calculate the output feature map; Step 3: Group the feature maps at the channel and perform normalization on each group; Step 4: Perform maximum pooling on the feature map and perform dimensionality reduction compression to speed up the calculation; Step 5: Pass the feature map into the fully connected layer, integrate and normalize the features, and obtain the fault probability distribution based on the fully connected layer and activation function.
5. The complex chemical process fault diagnosis method based on ensemble learning according to claim 1 is characterized in that: The inference process of the graph neural network classifier in step 3.2 is: Step 1: Use the K-nearest neighbor graphing algorithm to model multi-channel sensor data and construct a spatiotemporal relationship causal graph, with time series as node embeddings and chemical reaction relationships as edges; Step 2: Use graph convolution operations to extract the corresponding features of the chemical process under normal and fault conditions; Step 3: After the graph pooling operation, the feature representation of each node is obtained, and the features of the global graph are represented through the readout layer; Step 4: Pass the readout layer results to the fully connected layer, integrate and normalize the features, and obtain the fault probability distribution based on the activation function.