Industrial process unknown fault diagnosis method based on probability random forest

Abstract

The invention discloses an industrial process unknown fault diagnosis method based on a probability random forest, which belongs to the technical field of industrial process fault diagnosis, and realizes fault diagnosis by extracting potential fault features through a variation encoder, constructing a fault classifier based on the probability random forest and fusing a discriminator of an entropy and extreme value theory. According to the method, process variable data with the sampling frequency of 3 minutes and available fault labels are used for analysis, feature extraction is carried out through a variational encoder, and robust potential features with specific distribution are extracted; designing a fault classifier based on a probability random forest, classifying fault samples of known categories, and feeding classification loss of the classifier back to an encoder to adjust distribution of potential features so as to optimize extraction and classification performance of the model on fault features; a fault diagnosis method fusing entropy and extreme value theories is designed, unknown fault diagnosis in the industrial process is completed, and accurate recognition and efficient distinguishing of fault modes in the complex industrial process are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial process fault diagnosis, and particularly relates to an industrial process unknown fault diagnosis method based on probabilistic random forest. Background Art

[0002] Fault diagnosis in industrial processes is an important part of ensuring the efficient and safe operation of production lines. In industrial processes, abnormal faults in equipment, pipelines, and control systems often lead to a decrease in production efficiency, an increase in energy consumption, and even major safety accidents. With the rapid development of data acquisition and analysis technologies, data-driven fault diagnosis methods have gradually become the mainstream. In particular, the application of deep learning and machine learning algorithms can automatically identify potential faults through the learning and analysis of massive amounts of data. Existing fault diagnosis methods often have difficulty dealing with unknown category fault diagnosis problems and are easily affected by noise interference in a changing industrial environment, which affects the accuracy of diagnosis results. Therefore, developing an efficient diagnosis method that can handle both known and unknown faults is of great significance for improving the safety and stability of industrial production processes and also provides more accurate technical support for fault early warning and maintenance of industrial equipment.

[0003] Fault diagnosis in industrial processes has always been an important topic for ensuring production safety and stability. Its inducing factors are complex and diverse, including equipment aging, environmental changes, operation errors, and system load fluctuations. The operation mechanism of fault diagnosis is affected by multiple factors, such as fluctuations in production parameters, equipment operating states, and control system responses. These factors are intertwined, resulting in diverse fault manifestations and making it difficult to establish an accurate fault diagnosis model. In addition, with the complexity of industrial processes and the diversity of equipment, the types of faults are increasing day by day, and many of these faults belong to unknown categories, posing a huge challenge to industrial fault diagnosis. Existing fault diagnosis methods are difficult to effectively identify these unknown faults because they are usually trained only for known categories and will misclassify new types of faults as existing fault types when faced with new types of faults. Therefore, how to improve the model's diagnostic ability for unknown faults has become an urgent problem in industrial processes.

[0004] Therefore, it is necessary to establish an effective industrial process unknown fault diagnosis method to enhance the discrimination ability of fault features, thereby facilitating the accurate identification of unknown faults; extract robust latent features with a specific distribution through a variational encoder; design a fault classifier based on probabilistic random forest and replace the decoder to classify known category fault samples, and then feedback the classification loss of the classifier to the encoder to adjust the distribution of latent features, and further optimize the model's feature extraction and classification performance for fault samples; design a fault diagnosis method that combines entropy and extreme value theory to achieve accurate fault diagnosis and ensure the stable and safe operation of industrial processes. Summary of the Invention

[0005] The object of the present invention is to propose an unknown fault diagnosis method for industrial processes based on probabilistic random forests. Through a variational encoder, robust latent features with a specific distribution are extracted. A fault classifier based on probabilistic random forests is designed, and the decoder is replaced to classify known fault samples. Then, the classification loss of the classifier is fed back to the encoder to adjust the distribution of the latent features, thereby optimizing the feature extraction and classification performance of the model for fault samples. A fault diagnosis method integrating entropy and extreme value theory is designed to achieve the diagnosis of unknown faults in industrial processes, solve the problem of unknown fault diagnosis in complex industrial processes, improve the reliability and safety of industrial systems, and ensure the safe and stable operation of industrial production.

[0006] To achieve the above object, the present invention adopts the following technical solutions: An unknown fault diagnosis method for industrial processes based on probabilistic random forests extracts latent features through a variational encoder, constructs a fault classifier based on probabilistic random forests, and a discriminator integrating entropy and extreme value theory realizes fault diagnosis, specifically including the following steps: Step 1: Feature extraction based on a variational encoder to extract robust latent features with a specific distribution; Step 2: Design a fault classifier based on probabilistic random forests, replace the decoder, classify known fault samples, and then feed back the classification loss of the classifier to the encoder to adjust the distribution of the latent features, thereby optimizing the feature extraction and classification performance of the model for fault samples; Step 3: An unknown fault diagnosis strategy integrating entropy and extreme value theory to determine whether the current fault belongs to an unknown class.

[0007] As a further description of the above technical solution: The specific process of the above Step 1 is as follows: Step 1.1: Given a labeled data set , x j represents the process variable, y j represents its corresponding label, , N represents the number of samples, K represents the number of known class labels; Step 1.2: Extract fault latent features through a variational encoder. Input the original data into the variational encoder neural network, and map the original high-dimensional data x to the latent space by learning a probability distribution, x ={ x 1 ,…, x N}, the encoder will output two parameters: the mean μ and the variance σ . To make the forward propagation of the network differentiable, the reparameterization trick is used to generate the latent feature vector: (1); where ε is a random variable that follows the standard normal distribution N (0, 1), is the latent feature vector generated using the reparameterization trick, μ is the mean vector, σ is the standard deviation vector with the same dimension as the latent vector, which is calculated by two linear neurons; As a further description of the above technical solution: The specific process of step 2 is as follows: Step 2.1. Represent the fault latent features extracted from the original data by the variational encoder as the feature vector matrix , and pass it as the input to the random forest classifier; Step 2.2. The random forest classifier uses the bootstrap method to perform random sampling with replacement from the latent feature space to generate different training subsets for constructing the decision tree model in the random forest; The classification and regression tree algorithm measures the purity of the node through the Gini index to select the optimal feature and split point. The lower the Gini index value, the higher the proportion of samples of the same category in the child node. Given a node t and the sample probability of category in this node, the Gini index of node t is defined as follows: (2); where t represents the node, is the proportion of samples belonging to category in this node, C represents the number of categories contained in this node; Let s be the split point of node t , which distributes the samples in node t to p R and t R in a certain proportion p L , and distributes to t L in a certain proportion , that is (3); where Represents the Gini impurity of a node t , represents the Gini impurity of the left child node after splitting, represents the Gini impurity of the right child node after splitting, and are the number of samples in the left and right child nodes, is the total number of samples in the current node; then, through the following formula, the optimal feature and the best splitting point that can minimize the Gini impurity are obtained: (4); where is the best splitting point, is the feature corresponding to the best splitting point, is the amount of change in the Gini impurity after splitting at a node s through a certain feature t ; then, during the construction of the decision tree, the Classification and Regression Tree (CART) algorithm recursively calls the above function to repeatedly calculate the optimal feature and splitting point of each node, so as to ensure that each split can minimize the Gini impurity to the greatest extent until the set stopping condition is reached, so that the finally constructed decision tree can achieve higher classification accuracy at the leaf nodes; the classification task is performed through a random forest composed of decision trees and the number of votes for each sample belonging to each category is calculated ; Step 2.3. Use the SoftMax activation function to convert these number of votes into the probability distribution of categories. The SoftMax activation function takes as the input and, after calculation (5); where K represents the total number of known fault categories, is the number of votes for the sample belonging to the th category, is the probability that the sample belongs to the th category; the number of votes for the classifier to output the sample belonging to each category uses the normalized SoftMax activation function to convert the number of votes into probability values , and the range of the output probability is from 0 to 1, and the sum of the probabilities is equal to 1; Step 2.4. Calculate the classification loss of the random forest classifier, and feedback the classification loss of the classifier to the encoder to adjust the distribution of latent features, thereby optimizing the feature extraction and classification performance of the model for fault samples. The goal of the model is to learn how to accurately map the input data to its corresponding class label through optimization, by continuously optimizing the loss function LGradually update the latent space and continuously adjust the distribution of latent features, so that the classifier can more accurately predict the class labels of the data, and then learn the required latent features. The model loss function is defined as: (6); Wherein, represents the classification error of the known classes, and cross-entropy loss is used for classification; is KL the weight hyperparameter of the divergence, which measures the difference between the approximate posterior and ; As a further description of the above technical solution: The specific process of step 3 is as follows: Step 3.1: For the vote count vector output by the classifier, after using the SoftMax activation function to convert the vote count into a probability value , calculate its Shannon entropy as a measure of uncertainty, which can be specifically expressed as: (7); Wherein, represents the probability that the sample is classified as the th class. The larger H(X) is, the greater the uncertainty. Calculate the Shannon entropy of each fault through the probability vector . For the known-class fault samples, the optimized model enables it to output a more certain classification result, that is, the probability of a certain class in the probability vector is close to 1, and the probabilities of other classes are very low. At this time, the Shannon entropy value of the model for the known-class samples should be very small, indicating that the model has a high certainty in classifying these samples; Step 3.2: Denote the latent features of the training dataset obtained after the encoder is trained as z i,j , representing the th correctly classified sample of the th class in the training set, and calculate the class-average latent feature : (8); Wherein, is the number of samples, z i,j is the latent feature vector of the th sample of the th class, is the class-average latent feature of the th class. By calculating the given class-average latent feature z i,j ​The Euclidean distance between them is used as the distribution value of the Weibull model, denoted as d i,j : (9); where d i,j represents the th sample of the z i,j th class's latent vector and the Euclidean distance between the average latent feature of the th class; the tail size is the expected number of unknown samples in the 5% specified training set; according to the extreme value theory (EVT), the tail data of the th class is fitted to the Weibull distribution, and the maximum likelihood estimation (MLE) method is constructed by using the cumulative distribution function CDF of the Weibull distribution to estimate the three parameters of the location, scale, and shape of the EVT model is the number of samples, is the location parameter of the EVT model, is the scale parameter of the EVT model, is the shape parameter of the EVT model; for each sample classified as the th class , we use the Weibull model of this class (11); where , , is 's location, scale, and shape parameters, represents the Euclidean distance between the latent feature of the validation test failure and the average latent feature of the th class , Step 3.3; at the same time, calculate the Shannon entropy of the validation set failure and the Euclidean distance between the latent feature of the validation set failure and the relevant average feature of the th class , and calculate the rejection probability of the validation set failure. A hyperparameter Indicates the percentage of samples in the validation set failure that are accepted as known classes, and this boundary entropy value and probability are used as the entropy threshold and the probability rejection threshold; Step 3.4: Calculate the Shannon entropy of the test set failure and the rejection probability of the test set failure , where represents the entropy threshold, and represents the probability rejection threshold; The fault diagnosis evaluation logic is: (12); When the evaluation metrics and are both less than or equal to the threshold, the fault belongs to the known class fault; otherwise, it belongs to the unknown class fault.

[0008] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are: 1. In the present invention, based on the publicly available Tennessee Eastman process fault dataset, the process variable data with a sampling frequency of 3 minutes and the corresponding fault labels are used for analysis. Feature extraction is carried out through a variational autoencoder to extract robust latent features with a specific distribution; A fault classifier based on a probabilistic random forest is designed to replace the decoder, classify the fault samples of known classes, and feedback the classification loss of the classifier to the variational autoencoder to adjust the distribution of the latent features, thereby optimizing the feature extraction and classification performance of the model for fault samples; A fault diagnosis method integrating entropy and extreme value theory is designed to realize the diagnosis of unknown faults in industrial processes and accurately identify and efficiently distinguish fault patterns in complex industrial processes.

[0009] 2. In the present invention, data analysis and feature extraction are carried out by combining a variational autoencoder and a probabilistic random forest classifier, and known faults are classified by the classifier; At the same time, an unknown fault diagnosis method integrating entropy and extreme value theory is designed to effectively diagnose unknown faults in industrial systems and ensure the stability and safety of industrial production processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a schematic flow chart of the method for diagnosing unknown faults in industrial processes based on probabilistic random forests proposed by the present invention; Figure 2 is a fault diagnosis result diagram of the Weibull probability model discriminator based on EVT in the experiment of the method for diagnosing unknown faults in industrial processes based on probabilistic random forests proposed by the present invention; Figure 3 is a fault diagnosis result diagram of the discriminator based on entropy in the experiment of the method for diagnosing unknown faults in industrial processes based on probabilistic random forests proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0011] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0012] Please refer to the attached Figure 1 - attached Figure 3 , the present invention provides a technical solution: an unknown fault diagnosis method for industrial processes based on probabilistic random forests, which extracts latent features through a variational encoder, designs a fault classifier based on probabilistic random forests, and then designs a discriminator that fuses entropy and extreme value theory for unknown fault diagnosis. The specific steps are as follows: Step 1: Feature extraction based on a variational encoder to extract robust latent features with a specific distribution; Step 2: Design a fault classifier based on probabilistic random forests, replace the decoder, classify the fault samples of known categories, and then feedback the classification loss of the classifier to the encoder to adjust the distribution of the latent features, thereby optimizing the feature extraction and classification performance of the model for fault samples; Step 3: An unknown fault diagnosis strategy that fuses entropy and extreme value theory to determine whether the current fault belongs to an unknown category.

[0013] The specific process of Step 1 is as follows: Step 1.1: Given a labeled dataset , x j represents the process variable, y j represents its corresponding label, , N 800 represents the number of samples, K 6 represents the number of known class labels; Step 1.2: Extract fault latent features through a variational encoder, input the original data into the variational encoder neural network, and map the original high-dimensional data x to the latent space by learning a probability distribution, x ={ x 1 ,…, x N}, the encoder will output two parameters: the mean μ and the variance σ . In order to make the forward propagation of the network differentiable, the reparameterization trick is used to generate the latent feature vector: (1); Among them, ε is a random variable subject to the standard normal distribution N (0, 1), is the latent feature vector generated using the reparameterization trick, μ is the mean vector, σ is the standard deviation vector with the same dimension as the latent vector, calculated by two linear neurons; The specific process of step 2 is as follows: Step 2.1: Represent the fault latent features extracted from the original data by the variational autoencoder as the feature vector matrix and pass it as input to the random forest classifier; Step 2.2: The random forest classifier uses the bootstrap method to perform random sampling with replacement from the latent feature space to generate different training subsets for constructing the decision tree model in the random forest; the classification and regression tree algorithm measures the purity of the node through the Gini index to select the optimal feature and split point. The lower the Gini index value, the higher the proportion of samples of the same category in the child node. Given a node t and the probability of the sample of category in this node, the Gini index of node t is defined as follows: (2); Among them, t represents the node, is the proportion of samples belonging to category in this node, C represents the number of categories included in this node; let s be the split point of node t , which divides the samples in node t into a certain proportion p R and assigns them to t R , and into a certain proportion p L and assigns them to t L , that is, , therefore, the decrease in the Gini impurity index is defined as follows: (3); Among them, represents the Gini impurity of node t , represents the Gini impurity of the left child node after splitting, represents the Gini impurity of the right child node after splitting, and is the number of samples of the left and right child nodes, is the total number of samples of the current node; then, through the following formula, the optimal feature and the best splitting point that can minimize the Gini impurity are obtained: (4); where, is the best splitting point, is the feature corresponding to the best splitting point, is the change in Gini impurity after splitting at node s through a certain feature t ; then, during the construction of the decision tree, the Classification and Regression Tree (CART) algorithm recursively calls the above function to repeatedly calculate the optimal feature and splitting point of each node, so as to ensure that each split can minimize the Gini impurity to the greatest extent until the set stopping condition is reached, so that the finally constructed decision tree can achieve higher classification accuracy at the leaf nodes; the classification task is performed through the random forest composed of decision trees and the number of votes of the sample belonging to each category is calculated ; Step 2.3, Use the SoftMax activation function to convert these number of votes into the probability distribution of the category. The SoftMax activation function takes as the input, and after calculation (5); where, K 6 represents the total number of known fault categories, is the number of votes of the sample belonging to the th category, is the probability that the sample belongs to the th category; the number of votes of the classifier output for the sample belonging to each category uses the normalized SoftMax activation function to convert the number of votes into probability values , and the range of the output probability is from 0 to 1, and the sum of the probabilities is equal to 1; Step 2.4, Calculate the classification loss of the random forest classifier, and feedback the classification loss of the classifier to the encoder to adjust the distribution of the latent features, thereby optimizing the feature extraction and classification performance of the model for fault samples. The goal of the model is to learn how to accurately map the input data to its corresponding class label through optimization. By continuously optimizing the loss function L gradually update the latent space, continuously adjust the distribution of the latent features, so that the classifier can more accurately predict the class label of the data, and then learn the required latent features. The model loss function is defined as: (6); where, Represents the classification error of the known class, and cross-entropy loss is used for classification; is 0.5 KL the weight hyperparameter of the divergence, which measures the approximate posterior and the difference between them; The specific process of step 3 is as follows: Step 3.1. For the vote count vector output by the classifier, use the SoftMax activation function to convert the vote count into a probability value After that, calculate its Shannon entropy as a measure of uncertainty, which can be specifically expressed as: (7); Among them, represents the probability that the sample is classified into the th class, H(X) The larger it is, the greater the uncertainty; Calculate the Shannon entropy of each fault through the probability vector For known-class fault samples, the optimized model enables it to output a more definite classification result, that is, the probability of a certain class in the probability vector is close to 1, while the probabilities of other classes are very low. At this time, the Shannon entropy value of the model for known-class samples should be very small, indicating that the model has a high certainty in classifying these samples; Step 3.2. Denote the latent features of the training dataset obtained after the encoder is trained as z i,j , representing the th class in the training set and the th correctly classified sample, and calculate the class-average latent feature : (8); Among them, 800 is the number of samples, z i,j is the latent feature vector of the th sample of the th class, is the class-average latent feature of the th class; By calculating the Euclidean distance between the given class-average latent feature z i,j of the fault sample and the latent feature d i,j as the distribution value of the Weibull model, which is expressed as (9); Among them, d i,j represents the th class and the Latent vector of a sample z i,j With the Average latent feature of the class Euclidean distance between; Tail size Specify the expected number of unknown samples in the training set as 5%; According to the extreme value theory (EVT), fit the tail data of the class to the Weibull distribution, and construct a maximum likelihood estimation (MLE) method using the cumulative distribution function CDF of the Weibull distribution to estimate the three parameters of the location, scale, and shape of the EVT model : (10); Wherein, 800 is the number of samples, Is the location parameter of the EVT model, Is the scale parameter of the EVT model, Is the shape parameter of the EVT model; For each sample classified as the class , we use the Weibull model of this class Calculate its CDF probability as the rejection probability, that is, the probability of belonging to an unknown class: (11); Wherein, , , Is The location, scale, and shape parameters of, Indicates the Euclidean distance between the latent feature of the verification test failure and the Average latent feature of the class between, Indicates the rejection probability of the verification test failure; Step 3.3; At the same time, calculate the Shannon entropy of the verification set failure and the Euclidean distance between the latent feature of the verification set failure and the Relevant average feature of the class between, and calculate the rejection probability of the verification set failure to determine the hyperparameter Is 93%, and use this boundary entropy value and probability as the entropy threshold and probability rejection threshold; Step 3.4, calculate the Shannon entropy of the test set failure and the rejection probability of the test set failure, Indicates the entropy threshold, Indicates the probability rejection threshold; The fault diagnosis evaluation logic is: (12); When the evaluation indicators and are both less than or equal to the threshold value, the fault belongs to the known type of fault; otherwise, it belongs to the unknown type of fault.

[0014] To prove the feasibility and superiority of the present invention, experiments are conducted using 10 types of faults in the publicly available Tennessee fault dataset. Each type of fault has 800 samples, and each sample consists of 52 observed variables, including 22 continuous process measurements, 19 components, and 11 manipulated variables. During the experiment, 5 groups of experiments are set up. For each group, 6 types of faults are used for training, with the number of training samples being 6×800, 2 types of faults are used for validation, with the number of validation samples being 2×800, and 8 types of faults are used for testing, with the number of testing samples being 8×800. The diagnostic effect of unknown faults in the industrial process is as Figure 2 - Figure 3 shown.

[0015] Figure 2 The figure shows the fault diagnosis result graph of the Weibull probability model discriminator based on EVT. X-axis: task group number; Y-axis: values of evaluation indicators of each classification model, with the unit being percentage. The blue bars represent the accuracy values, the orange bars represent the known accuracy values, the gray bars represent the precision values, the yellow bars represent the recall values, and the light blue bars represent the F1 score values. Figure 3 The figure shows the fault diagnosis result graph of the discriminator based on entropy. X-axis: task group number; Y-axis: values of evaluation indicators of each classification model, with the unit being percentage. From Figure 2 and Figure 3 it can be seen that the Precision of the VEC method is basically above 0.91, and the f1-score is above 0.93, indicating that the proposed variational autoencoder classifier fault diagnosis method can correctly detect unknown faults.

[0016] As described above, the above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. An unknown fault diagnosis method for industrial processes based on probabilistic random forests, characterized in that: The potential fault features are extracted by variational encoder, a fault classifier is constructed based on probabilistic random forest, and the discriminator of entropy and extreme value theory is integrated to realize fault diagnosis, which specifically includes the following steps: Step 1, based on variational encoder, robust potential features with specific distribution are extracted; Step 2: Design a fault classifier based on probabilistic random forest to diagnose known faults, and feed the loss function of the classifier back to the variational encoder to adjust the distribution of its potential features, thereby optimizing the model's feature extraction and classification performance for fault samples; Step 3: Establish an unknown fault diagnosis strategy that integrates entropy and extreme value theory to determine whether the current fault belongs to the unknown class.

2. The method for industrial process unknown fault diagnosis based on probabilistic random forest according to claim 1 is characterized in that: The specific process of step 1 is as follows: Step 1.1: Given a labeled dataset , x j represents the process variable, y j Indicates the corresponding label, , N represents the number of samples, K Indicates the number of known class labels; Step 1.2: Extract potential fault features through the variational encoder, input the original data into the variational encoder neural network, and transform the original high-dimensional data into x Mapped into the latent space, x ={ x 1,…, x N }, the encoder will output two parameters: mean μ and variance σ , in order to make the forward propagation of the network differentiable, the reparameterization technique is used to generate the potential feature vector: (1)； in, ε is a standard normal distribution N A random variable with the value (0,1). is the latent feature vector generated using the reparameterization technique, μ is the mean vector, σ is a standard deviation vector of the same dimension as the latent vector, computed by two linear neurons.

3. The method for industrial process unknown fault diagnosis based on probabilistic random forest according to claim 1 is characterized in that: The specific process of step 2 is as follows: Step 2.1: Extract the fault potential feature vector matrix from the original data using the variational encoder , passed as input to the random forest classifier; Step 2.2: The random forest classifier uses the bootstrap method to perform random sampling with replacement from the potential feature space to generate different training subsets for building the decision tree model in the random forest. The classification and regression tree algorithm uses the Gini index to measure the purity of the node to select the optimal features and split points. The lower the Gini index value, the higher the proportion of samples of the same category in the child node. For a given node t and the categories in this node The sample probability of ,node t The Gini index is defined as follows: （2）； in, t Represents a node, Is the node belonging to the category The proportion of samples C Indicates the number of categories contained in the node; s For the festival t The split point of t The samples in a certain proportion p R Assigned to t R , according to a certain proportion p L Assigned to t L ,Right now , therefore, the decrease in the Gini impurity index is defined as follows: （3）； in, Representation Node t The Gini impurity of Represents the Gini impurity of the left child node after the split, Represents the Gini impurity of the right child node after the split, and is the number of samples of the left and right child nodes, is the total number of samples at the current node; then, the optimal features and optimal segmentation points that can minimize the Gini impurity are obtained through the following formula: （4）； in, is the best split point, is the feature corresponding to the optimal split point, Through a feature s At the node t The amount of change in Gini impurity after the division; then, in the process of decision tree construction, the classification and regression tree algorithm recursively calls the above function to repeatedly calculate the optimal features and split points of each node, so as to ensure that each division can minimize the Gini impurity until the set stop condition is reached, so that the final decision tree can achieve higher classification accuracy at the leaf node; the classification task is performed through the random forest composed of decision trees and the number of votes for the sample belonging to each category is calculated ; Step 2.3: Use the SoftMax activation function to convert these votes Converted into the probability distribution of categories, the SoftMax activation function converts As input, calculated （5）； in, K Represents the total number of known fault categories, The sample belongs to The number of votes for the category, The sample belongs to The probability of the class; the number of votes that the classifier outputs for the sample belongs to each category Use the normalized SoftMax activation function to convert the number of votes into probability values , the output probability ranges from 0 to 1, and the sum of the probabilities is equal to 1; Step 2.4: Calculate the classification loss of the random forest classifier and feed the classification loss of the classifier back to the encoder to adjust the distribution of potential features, thereby optimizing the model's feature extraction and classification performance for fault samples. The goal of the model is to learn how to accurately map input data to its corresponding category label by optimizing the learning function. L Gradually update the latent space and continuously adjust the distribution of latent features so that the classifier can more accurately predict the category label of the data and then learn the required latent features. The model loss function is defined as: （6）； in, Represents the classification error of known classes, using cross entropy loss for classification; yes KL The weighted hyperparameter of the divergence, which measures the approximate posterior and The difference between.

4. The method for diagnosing unknown faults in industrial processes based on probabilistic random forests according to claim 1 is characterized in that: The specific process of step 3 is as follows: Step 3.1: For the vote vector output by the classifier , use the SoftMax activation function to convert the number of votes into a probability value After that, the Shannon entropy is calculated as a measure of uncertainty, which can be specifically expressed as: （7）； in, Indicates that the sample is classified as The probability of the class, H(X) The larger the probability vector The Shannon entropy of each fault is calculated. For known fault samples, the optimized model can output a more certain classification result, that is, the probability vector The probability of a certain category is close to 1, while the probability of other categories is very low. At this time, the Shannon entropy value of the model for known class samples should be very small, indicating that the model has a high degree of certainty in the classification of these samples. Step 3.2: After the encoder is trained, the potential features of the training data set are recorded as z i,j , which means the first Class correctly classified samples and calculate the class average latent features : （8）； in, is the sample size, z i,j yes Type of fault The potential feature vector of samples, It is The class average latent feature of the class; by calculating the average latent feature of a given class of fault samples With potential characteristics z i,j The Euclidean distance between is taken as the distribution value of the Weibull model and is expressed as d i,j : （9）； in, d i,j Indicates Class The latent vector of the sample z i,j With The average latent feature of the class The Euclidean distance between The expected number of unknown samples in the training set is specified as 5%; according to the extreme value theory (EVT), The tail data of the class is fitted to the Weibull distribution, and the maximum likelihood estimation (MLE) method is constructed by using the cumulative distribution function CDF of the Weibull distribution to estimate the location, scale and shape parameters of the EVT model. : （10）； in, is the sample size, is the location parameter of the EVT model, is the scale parameter of the EVT model, is the shape parameter of the EVT model; for each classification Sample of class , we use this type of Weibull model Calculate its CDF probability as the rejection probability, that is, the probability of belonging to the unknown class: （11）； in, , , for The location, scale and shape parameters of The potential characteristics that indicate a proof test failure are similar to the The average latent feature of the class The Euclidean distance between represents the rejection probability of verification test failure; Step 3.3: Simultaneously calculate the Shannon entropy of the validation set failure and the potential characteristics of the validation set failure and the The average feature of the class The Euclidean distance between them is used to calculate the rejection probability of the validation set failure, and a hyperparameter is predefined. Used to determine the entropy threshold and probability rejection threshold, Indicates the percentage of samples in the validation set that are accepted as known classes, and this boundary entropy value and probability as entropy threshold and probability rejection threshold; Step 3.4: Calculate the Shannon entropy of the test set failure and the rejection probability of test set failure , represents the entropy threshold, represents the probability rejection threshold; the fault diagnosis evaluation logic is: （12）； When the evaluation index and When both are less than or equal to the threshold, the fault belongs to a known fault; otherwise, it belongs to an unknown fault.

Images