A method for diagnosing minor faults in industrial processes based on probabilistic slow feature contrast learning

Abstract

The present invention provides a method for diagnosing minor faults in industrial processes based on probabilistic slow feature contrast learning, and belongs to the technical field of multivariable complex industrial process fault diagnosis. It solves the technical problem that traditional machine learning algorithms are difficult to effectively diagnose minor faults in complex nonlinear industrial processes. Its technical solution includes the following steps: S1, obtaining training data from historical data and preprocessing; S2, extracting probabilistic slow features by combining Jensen-Shannon (JS) divergence and slow feature analysis; S3, constructing a contrast learning network framework; S4, using probabilistic slow features as network input for training; S5, saving the trained network parameters; S6, obtaining online data and inputting it into the network to obtain fault diagnosis results. The beneficial effects of the present invention are: accurate extraction of minor fault features is achieved, and the accuracy of fault diagnosis is significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process fault diagnosis, and in particular to an industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning. Background Art

[0002] Whether the industrial process is in a healthy state directly determines the quality of the product, economic benefits and personal safety, so fault diagnosis has always been a hot topic. With the rapid development of science and technology, industrial processes are becoming automated, large-scale and complex. Along with the improvement of production efficiency, the impact of faults is becoming more and more destructive. Even small faults in the early stage will evolve into major faults that endanger the entire system over time. In this context, researchers have proposed many fault detection and diagnosis algorithms, such as some traditional multivariate statistical analysis methods: principal component analysis (PCA), slow feature analysis (SFA), etc.

[0003] As a new type of unsupervised data transformation and dimensionality reduction method, SFA can extract constant or slow-changing data components from rapidly changing time series data. In industrial processes, the powerful dynamic analysis capabilities of SFA make it easier to detect anomalies in the data. Patent CN116204775A discloses a mechanical equipment fault diagnosis method based on slow feature analysis, which extracts slow feature matrix diagrams from two-dimensional features and performs model training to improve the fault diagnosis capability. However, since the mathematical principles of multivariate statistical analysis methods are relatively simple, the extracted features are mostly shallow, the extracted features are of low quality, and the diagnostic accuracy is low, a single fault diagnosis strategy based on multivariate statistical methods has gradually been neglected.

[0004] On the other hand, with the surge in process data and the rapid development of deep learning, new research directions have been brought to the field of fault diagnosis, and intelligent fault diagnosis methods based on deep learning have been recognized and widely used. Researchers have successively proposed a series of intelligent diagnosis methods, such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), etc. In patent CN115165363B, a light bearing fault diagnosis method and system based on CNN is proposed, and the constructed light network has achieved good diagnostic results.

[0005] However, during the research process, it was found that although the introduction of deep learning has significantly improved the ability to extract fault features, and with its help, most fault types can be detected and diagnosed, the diagnosis effect of minor faults in multivariable complex industrial processes is still not ideal. These minor faults often occur in the early stages of the process, and have the characteristics of low amplitude, slow change, easy to be disturbed by process and measurement noise. Their early data features are very similar to normal data. Their effective detection and diagnosis has always been one of the difficult problems in the field of industrial process fault diagnosis. Therefore, how to deeply mine effective minor fault features from complex nonlinear industrial process data and realize timely monitoring and diagnosis is an important and challenging issue in current research. Summary of the invention

[0006] The present invention aims at the technical problem that traditional data-driven machine learning algorithms are difficult to effectively diagnose minor faults in complex nonlinear industrial processes, and provides a method for diagnosing minor faults in industrial processes based on probabilistic slow feature contrast learning; the method combines JS divergence with the SFA algorithm to extract probabilistic slow features of data, and then constructs positive and negative pairs with the extracted probabilistic slow features, which are input into a feature extractor network based on contrast learning for training, to achieve refined extraction of minor fault features, and finally the extracted features are input into a Softmax classifier to achieve fault diagnosis. The designed two-step feature extraction method not only expands the traditional slow feature extraction method to the probabilistic space, but also combines it with contrast learning for the first time, achieving efficient extraction of minor fault features, and greatly improving the accuracy of minor fault diagnosis in complex nonlinear industrial processes.

[0007] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a method for diagnosing minor faults in industrial processes based on probabilistic slow feature contrast learning, comprising the following steps:

[0008] Step S1: Obtain normal operating condition data X from the historical database 0 , using its mean(X 0 ) and standard deviation std(X 0 ) for historical data X old Standardize to obtain the standardized training data X;

[0009] Step S2: Extract the main slow features S from the training data X using the slow feature analysis method. 1 ,S 2 ,...,S M ];

[0010] Step S3: Introduce JS divergence, combine the sliding window method to process the slow feature S, map it to the probability space, and obtain the probabilistic slow feature P of the training data X = [P 1 ,P2 ,...,P M ];

[0011] Step S4: Data set preparation: repeat steps S2 and S3 to obtain several probabilistic slow feature data samples of normal working conditions and various fault types, mark the fault labels, and use them as training data sets;

[0012] Step S5: construct comparative learning sample pairs using the training data set;

[0013] Step S6: using the training data set samples to train a contrastive learning feature extractor with 1D-CNN as the backbone network;

[0014] Step S7: training a Softmax classifier using features extracted by the trained feature extractor;

[0015] Step S8: After reaching the set training round, save the trained network parameters of the contrastive learning feature extractor and the Softmax classifier;

[0016] Step S9: Collect data from the industrial production process as test data X new , prepare an unlabeled test data set in the same manner as steps S1 to S4;

[0017] Step S10: input the prepared unlabeled test data set into the trained contrastive learning feature extractor and Softmax classifier in sequence to obtain a fault classification result.

[0018] Furthermore, in step S1, the historical data X old Using normal data X 0 Mean(X 0 ) and standard deviation std(X 0 ) is standardized as follows:

[0019] X=(Xold-mean(X 0 )) / std(X 0 )(1)

[0020] In the formula, X=[X 1 ,X 2 ,...,X n ] T ∈R n×m , n is the number of samples, and m is the number of variables.

[0021] Furthermore, in step S2, the specific steps of extracting the main slow features from the training data X using the slow feature analysis method are:

[0022] (1) Normal operating condition data X can be obtained from the historical database0 ;

[0023] (2) Find the whitening matrix: for normal data X 0 The covariance matrix of is subjected to singular value decomposition:

[0024] Get the whitening matrix Q = Λ -1 / 2 U T , X 0 The whitening transformation is:

[0025] (3) Find the transformation matrix: The first-order derivative of the whitened data B The covariance matrix of is subjected to singular value decomposition:

[0026] The transformation matrix W = PΛ -1 / 2 U T ;

[0027] (4) All slow features of training data X are calculated according to the following formula: S = WX, where

[0028] S=[S 1 ,S 2 ,...,S m ]∈R n×m ;

[0029] (5) Calculate normal operating data X 0 The slowness of each variable Δ(X 0j ), according to the formula Determine the number of main slow features M, where q = 0.1 represents the set

[0030] {Δ(X j )} 0.1 upper quantile;

[0031] (6) Obtain the main slow features S of the training data X = [S 1 ,S 2 ,...,S M ]∈R n×M , where M<m, represents the number of main slow features, and n is the number of samples.

[0032] Furthermore, in step S3, the specific steps of calculating the probability slow feature P of the training data X using the sliding window method using the JS divergence are:

[0033] The parameters of the sliding window algorithm are set as follows: the window width is set to H, the step size is set to 1, each sampling moment moves backward once, the number of data samples is n, and there are a total of N = n-H + 1 sampling moments;

[0034] The normal working condition data X in the historical database0 As the benchmark data, the main slow feature S of the benchmark data is obtained after the same processing as steps S1 and S2. B =[S B1 ,S B2 ,...,S BM ]∈R n×M ;

[0035] Calculate S B The mean and variance of B and the benchmark variance λ B ;

[0036] Calculate the mean μ of the main slow feature S of the training data X at the kth sampling time S (k) and variance λ S (k);

[0037] The probability slow feature P of the kth sampling moment of the training data X is calculated by formula (2). The expression of formula (2) is as follows:

[0038]

[0039] Calculated P = [P 1 ,P 2 ,...,P M ]∈R N×M , N is the number of sampling moments, and M is the number of main slow features.

[0040] Furthermore, in step S5, the specific steps of constructing comparative learning sample pairs using the training data set are:

[0041] The training data set can be expressed as D = {p i ,y i}, where p is the sample data, y∈{1,...,K} represents the fault label, and i∈I≡{1,2,...,N} represents the sample index;

[0042] In the dataset D, for any two different samples [p i ,p j ], if y i =y j , that is, if the fault labels are the same, they can constitute a positive sample pair, otherwise they constitute a negative sample pair;

[0043] The positive sample pair set can be obtained:

[0044] Negative sample pair set:

[0045] Total set of sample pairs: A = {P + ,P -}.

[0046] Furthermore, in step S6, the specific steps of using the training data set samples to train the contrastive learning feature extractor with 1D-CNN as the backbone network are as follows:

[0047] S601, adjusting the contrast loss temperature coefficient τ according to the slowness of the training data;

[0048] S602, using 1D-CNN as the backbone feature extraction network to extract sample p i Features of z i ;

[0049] S603, using the extracted feature z i Calculate the slow feature contrast loss L sf ;

[0050] S604: Use an adaptive moment estimation algorithm (Adam optimizer) to update network parameters and save the parameters.

[0051] Furthermore, in step S601, the specific method of adjusting the contrast loss temperature coefficient τ according to the slowness of the training data is:

[0052] Slowness is an important concept in slow feature analysis. The size of slowness is an important indicator to measure the speed of data change. Slow features that change faster over time are noise features, while slow features that change slower over time reflect the real changes of data.

[0053] The temperature coefficient τ is an important parameter of contrastive loss, with a value range of 0-1, which determines the discrimination degree of the model for samples in contrastive learning. The closer the temperature coefficient value is to 0, the more the model pays attention to difficult samples, but it also easily leads to slow model convergence and poor generalization ability. On the contrary, if the value is set too large and close to 1, it will lead to no difference in model learning;

[0054] According to the characteristic slowness calculation formula: The average slowness Δ(S) and the maximum slowness Δ(S) of all features of the training data can be calculated. max , minimum slowness Δ(S) min ; From the normal historical data, the standard slowness Δ(S) of normal characteristics can be calculated N ;

[0055] The temperature coefficient τ can be determined from the data slowness by the following formula:

[0056]

[0057] Compared with the temperature coefficient set empirically, the temperature coefficient calculated by this formula can help the feature extractor extract features more accurately and converge the model faster.

[0058] Furthermore, in step S602, the input signal in the feature extraction network with 1D-CNN as the backbone passes through the input layer, the first convolution layer, the first pooling layer, the second convolution layer, the second pooling layer, the flattening layer, and the output layer in sequence. The specific parameters and functions are as follows:

[0059] Input layer: accepts probability slow feature data as input signal p i ;

[0060] The first convolution layer: The convolution kernel size is 3, which is used to extract the features of the input signal;

[0061] First pooling layer: The pooling kernel size is 2, which is used to reduce the spatial size of each feature map;

[0062] The second convolution layer: The convolution kernel size is 3, which is used to further extract the signal features after one convolution process;

[0063] Second pooling layer: The pooling kernel size is 2, which is used to reduce the spatial size of each feature map;

[0064] Flattening layer: flatten the obtained multi-dimensional features into a one-dimensional feature vector;

[0065] Output layer: output the extracted features z i .

[0066] Furthermore, in step S603, the extracted feature z is used i Calculate the slow feature contrast loss L sf The specific steps are:

[0067] Assuming that the total number of samples in a training batch is N, the contrast loss can be calculated as L sf

[0068]

[0069] Among them, the function dis(·) is used to calculate the Euclidean distance between two features, and τ is the temperature coefficient;

[0070] Training this loss can maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs, making samples with the same label more similar and amplifying the differences between samples with different labels.

[0071] Furthermore, in step S7, the specific steps of training the Softmax classifier using the features extracted by the trained feature extractor are:

[0072] The feature extraction operation up to step S6 can be defined as a function F(), then z i =F(x i ).

[0073] First, according to the dimension of the input data, one or more fully connected layers are used to transform z i Converted to a vector of the same length as the number of categories, the output of the Softmax classifier can be expressed as:

[0074]

[0075] Among them, θ is the fully connected layer parameter to be optimized;

[0076] The Adam optimizer is used to optimize the parameter θ, and the loss function is as follows:

[0077]

[0078] Among them, y is the true label of the sample, and the parameters are retained after optimization.

[0079] Furthermore, in steps S9 and S10, online test data is obtained from a variety of sensors installed on industrial production equipment. After being processed, the data is input into the trained comparative learning model saved in step S8 to output the type of minor faults.

[0080] In the above method, steps S1 to S3 are the probabilistic slow feature extraction stage, steps S4 to S8 are the slow feature comparison learning stage, and steps S9 to S10 are the online data fault diagnosis stage.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] 1. The industrial process minor fault diagnosis method provided by the present invention introduces JS divergence and combines JS divergence with slow feature analysis to realize the extraction of probabilistic slow features.

[0083] 2. The present invention proposes a method for determining the temperature coefficient of a key hyperparameter in contrastive learning according to data slowness, which helps the contrastive learning model to converge faster and obtain better feature extraction capabilities.

[0084] 3. The present invention adopts a two-step feature extraction method, innovatively integrates slow feature analysis with contrastive learning, and uses contrastive learning to construct a feature extractor to achieve refined extraction of minor fault features.

[0085] 4. Compared with the existing deep learning method for minor fault diagnosis, the constructed feature extraction network only contains two convolutional layers, which occupies less memory and computing resources, which helps the model to be applied in more scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0087] Figure 1 The figure is a general flow chart of the method of the present invention for diagnosing minor faults.

[0088] Figure 2 This is a process flow chart of the TE chemical process used in Example 1 of the present invention.

[0089] Figure 3 This is a comparison chart of the fault diagnosis accuracy of various methods in Example 1 of the present invention.

[0090] Figure 4 This is the confusion matrix of the diagnostic effect of the PRSFCL method of the present invention on TE minor faults in Example 1 of the present invention.

[0091] Figure 5 (a) is the t-SNE distribution diagram of the original data of the TE chemical process of Example 1 of the present invention.

[0092] Figure 5 (b) is the t-SNE distribution diagram of the data after the PRSFCL method of the present invention is used in Example 1 of the present invention.

[0093] Figure 6 This is a flow chart of the RP-1043 process operation used in Example 2 of the present invention.

[0094] Figure 7 This is the confusion matrix of the diagnostic effect of the RP-1043 data set using the PRSFCL method of the present invention in Example 2 of the present invention.

[0095] Figure 8 This is a comparison chart of the fault diagnosis accuracy of each method in Example 2 of the present invention.

[0096] Fig. 9 (a) is the t-SNE distribution diagram of the original data of the RP-1043 process of Example 2 of the present invention.

[0097] Fig. 9 (b) is the t-SNE distribution diagram of the data after the PRSFCL method of the present invention is used in Example 2 of the present invention. DETAILED DESCRIPTION

[0098] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0099] Example 1

[0100] See also Figures 1 to 5 This embodiment is based on the Tennessee-Eastman process data set, referred to as the TE process, which is a typical nonlinear industrial process derived from the actual industrial process of Eastman Chemical Company in the United States. The Ricker Process Control Laboratory of the University of Washington rewrote the TE chemical process using Matlab so that it can be simulated on Matlab. Its data has the characteristics of nonlinearity, time-varying and strong coupling, so it is widely used in the research and verification of fault detection and fault diagnosis methods.

[0101] The TE chemical process consists of five main units: reactor, condenser, compressor, separator and stripper. After the main gas raw material enters the process, through the operation of these parts and the chemical reactions therein, two main liquid products are produced from the four reactants, as well as an inert product and a by-product, a total of eight components: A, B, C, D, E, F, G, H. The process flow chart is as follows Figure 1 As shown, the various reactions in the reactor are:

[0102] A+C+D→G

[0103] A+C+E→H

[0104] A+E→F

[0105] 3D→2F

[0106] The TE process includes 12 operational variables and 41 measurement variables. In this example, 11 operational variables and 22 measurement variables commonly used in research are selected to construct a data set, a total of 33 variables. The selected variables are shown in Table 1.

[0107] The TE process is configured with 21 fault types, which are divided into step, unknown and other fault types. Among the many faults, fault 3, fault 9 and fault 15 are particularly special. They are minor faults that are difficult to detect and are also the research focus of this paper. The fault causes and types of these three minor faults are shown in Table 2.

[0108] In the construction of the data set, this example selects 800 samples from each fault category. At the same time, in order to test the model's ability to distinguish between minor faults and normal data, 800 normal samples are also selected as normal categories, that is, 3 minor faults and 1 group of normal classes, a total of 3200 samples constitute the data set. 70% of the samples are randomly selected as the training data set, and 30% as the test data set.

[0109] Table 1 List of variables

[0110]

[0111] Table 2 Causes and types of TE minor faults

[0112] Fault number Cause Fault type Fault 3 D Feed temperature change Step micro fault Fault 9 D Feed temperature change Random minor glitches Fault 15 Condenser cooling water valve stuck Sticky micro-fault

[0113] After the fault is detected, in order to evaluate the fault diagnosis effect of different methods, the diagnostic accuracy (ACC) performance index is used to compare the effects of different methods. It is defined as the proportion of the number of samples correctly predicted by the model to the total number of samples. The calculation method is as follows:

[0114]

[0115] Among them, TP is the number of samples predicted as positive, TN is the number of samples predicted as negative, FP is the number of samples predicted as positive, and FN is the number of samples predicted as negative.

[0116] In this embodiment, two methods, deep convolutional neural network method DCNN, deep learning slow feature analysis method DSFA and the probabilistic slow feature comparative learning method PRSFCL of the present invention are used for comparison. In terms of parameter setting, the width H of the sliding window used for slow feature analysis in this example is 40, and the remaining comparative learning network structure and parameters are shown in Table 3 below. In addition, all methods used for comparison use cross-validation methods to obtain the best results.

[0117] Table 3 Comparison of learning model composition and network parameter settings

[0118]

[0119] After testing, Table 4 shows the diagnostic accuracy results of three methods, DCNN, DSFA and the PRSFCL method of the present invention, for three minor faults in the TE process. As can be seen from the above table, the effect of DCNN is not ideal, and the accuracy of 70% is unacceptable in industrial production. Although DSFA has improved to some extent, it has only barely reached 90% accuracy. The PRSFCL method provided by the present invention can achieve an average accuracy of 98.3%. By further analyzing the accuracy comparison chart ( Figure 3 ), it can be found that the PRSFCL method achieved effective convergence in about 50 rounds, which is significantly better than the convergence speed of DSFA, and it has always maintained a high accuracy without overfitting.

[0120] The accuracy of minor fault diagnosis can be analyzed by Figure 4 The confusion matrix is ​​refined to each fault category. It can be found that the PRSFCL method achieves an efficient balance in the diagnosis ability of all minor fault categories, and can achieve a diagnostic accuracy of almost 99% for all fault types. The contrastive learning model trained by this method has excellent feature extraction and resolution capabilities. Figure 5(a) The original data distribution and Figure 5 (b) shows the data distribution after being processed by the contrast learning feature extractor. It can be found that the model has achieved efficient distinction of the original difficult-to-distinguish data. All this is due to the two-step feature extraction method adopted by the PRSFCL method. First, the data probability information is used to realize the preliminary mining of minor fault information. After contrast learning, the minor fault characteristics are further amplified, which significantly improves the accuracy of minor fault diagnosis. In summary, the minor fault diagnosis effect of the PRSFCL method of the present invention is significantly better than that of the other two methods.

[0121] Table 4 TE minor fault diagnosis accuracy results

[0122]

[0123]

[0124] Example 2

[0125] See also Figures 6 to 8 This embodiment uses chiller operation data from the American Society of Heating, Ventilation, Refrigeration and Air-Conditioning Engineers (ASHRAE) research project RP-1043. The project aims to obtain normal and fault data from a representative chiller. The researchers used a 90-ton centrifugal chiller as a test bench, using R134a refrigerant and a constant ambient temperature of 720F. The experimental setup includes three water-water heat exchangers, a steam-water heat exchanger, an electrically controlled three-way valve, six electrically controlled two-way valves, three water pumps, and two vortex flowmeters. It includes five water flow routes: condenser water, evaporator water and hot water loops, city water supply and steam supply lines. Its operation flow chart is shown in the following figure. Figure 6 shown.

[0126] In real industrial processes, operating conditions often change dynamically. In order to verify the accuracy of the proposed PRSFCL method for minor fault diagnosis under variable conditions, this study introduces the operating data of a chiller process. Since chiller processes usually require frequent switching of operating conditions, they are suitable for testing the proposed method.

[0127] RP-1043 consists of seven typical chiller faults. Each fault can be divided into four fault levels from low to high, among which the lowest level fault is a minor fault. The specific fault types and fault levels are shown in Table 5. This example dataset contains 7 types of fault data. The sensor obtains 64 measurements at intervals of 10 seconds. Considering the actual cost, only 16 of them are selected as measurement variables. In addition, various operating conditions are simulated by switching operating parameters in real time, covering 27 operating conditions included in the ASHRERP-1043 process. In the construction of the dataset, this example selects 800 samples from each fault category, including 7 minor faults, and a total of 5600 samples constitute the dataset. 70% of the samples are randomly selected as training datasets and 30% as test datasets.

[0128] In this embodiment, the deep convolutional neural network method (DCNN), the deep learning slow feature analysis method (DSFA) and the probability slow feature comparison learning method (PRSFCL) of the present invention are also used for comparison. In terms of parameter setting, the width H of the sliding window used for slow feature analysis in this example is 35, and the other parameters are the same as those in Example 1.

[0129] Table 5 RP-1043 dataset fault types and descriptions

[0130]

[0131]

[0132] After testing, Table 6 shows the diagnostic accuracy results of DCNN, DSFA and the PRSFCL method of the present invention for 7 types of faults in the RP-1043 data set. Figure 7 It can be seen that the PRSFCL method has a significantly higher diagnostic accuracy than the other two methods, reaching an average accuracy of 99.3%. Figure 7 ), it can be more intuitively found that the PRSFCL method is also better than the other two methods in terms of convergence speed while maintaining a higher accuracy. Fig. 9 The t-SNE distribution diagram intuitively demonstrates the excellent feature extraction and resolution capabilities of the model in this example. This shows that even when faced with multiple working conditions and multiple changes in the process, the proposed PRSFCL method can still have excellent fault diagnosis capabilities. In summary, the PRSFCL method of the present invention has a significantly better effect on minor fault diagnosis than the other two methods.

[0133] Table 6 RP-1043 dataset fault diagnosis accuracy results

[0134]

[0135] 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 method for diagnosing minor faults in industrial processes based on probabilistic slow feature contrast learning, characterized in that: The following steps are involved: Step S1: Get the normal operating condition data X0 in the historical database, and use its mean (mean(X0) and standard deviation std(X0) to analyze the historical data X0. old Standardize to obtain the standardized training data X; Step S2: Extract slow features S = [S1, S2, ..., S M ]; Step S3: Introduce Jensen-Shannon divergence JS, combine the sliding window method to process the slow feature S, map it to the probability space, and obtain the probabilistic slow feature P of the training data X = [P1, P2, ..., P M ]; Step S4: Data set preparation: repeat steps S2 and S3 to obtain several probabilistic slow feature data samples of normal working conditions and various fault types, mark the fault labels, and use them as training data sets; Step S5: construct comparative learning sample pairs using the training data set; Step S6: using the training data set samples to train a contrastive learning feature extractor with a one-dimensional convolutional neural network 1D-CNN as the backbone network; In step S6, the specific steps of using the training data set samples to train the contrastive learning feature extractor with 1D-CNN as the backbone network are as follows: S601, adaptively adjusting the contrast loss temperature coefficient τ according to the slowness of training data; S602, using 1D-CNN as the backbone feature extraction network to extract sample p i Features of z i ; S603, using the extracted feature z i Calculate the slow feature contrast loss L sf ; S604, using the adaptive moment estimation algorithm Adam optimizer to update the network parameters and save the parameters; In step S601, the specific method of adjusting the contrast loss temperature coefficient τ according to the slowness of the training data is: Slowness is an important concept in slow feature analysis. The size of slowness is an important indicator to measure the speed of data change. Slow features that change faster over time are noise features, while slow features that change slower over time reflect the real changes of data. The temperature coefficient τ is an important parameter of contrastive loss. Its value range is 0-1. It determines the discrimination degree of the model for samples in contrastive learning. The closer the temperature coefficient value is to 0, the more the model pays attention to difficult samples, but it also easily leads to slow model convergence and poor generalization ability. On the contrary, if the value is set too large and close to 1, it will lead to no emphasis on model learning. According to the characteristic slowness calculation formula: Calculate the average slowness Δ(S) and maximum slowness Δ(S) of all features of the training data max , minimum slowness Δ(S) min ; From the normal historical data, calculate the standard slowness Δ(S) of normal characteristics N ; The temperature coefficient τ is determined by the data slowness through the following formula: The temperature coefficient calculated by this formula helps the feature extractor to extract features more accurately and converge the model faster than the temperature coefficient set by experience; Step S7: training a Softmax classifier using features extracted by the trained feature extractor; Step S8: After reaching the set training round, save the trained network parameters of the contrastive learning feature extractor and the Softmax classifier; Step S9: Collect data from the industrial production process as test data X new , prepare an unlabeled test data set in the same manner as steps S1 to S4; Step S10: input the prepared unlabeled test data set into the trained contrastive learning feature extractor and Softmax classifier in sequence to obtain a fault classification result.

2. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: In step S1, the historical data X old The expression for standardization using the mean(X0) and standard deviation std(X0) of normal data X0 is: X=(X old -mean(X0)) / std(X0) (1) Where X=[X1,X2,...,X n ] T ∈R n×m , n is the number of samples, and m is the number of variables.

3. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: In step S2, the specific steps of extracting slow features from the training data X using the slow feature analysis method are: (1) Obtain normal operating condition data X0 from the historical database; (2) Calculate the whitening matrix: Perform singular value decomposition on the covariance matrix of the normal data X0: Get the whitening matrix Q = Λ -1 / 2 U T , the whitening transformation of X0 is: (3) Find the transformation matrix: The first-order derivative of the whitened data B The covariance matrix of is subjected to singular value decomposition: The transformation matrix W = PΛ -1 / 2 U T ; (4) All slow features of training data X are calculated according to the following formula: S = WX, where S = [S1, S2, ..., S m ]∈R n×m ; (5) Calculate the slowness Δ(X 0j ), according to the formula Determine the number of slow features M, where q = 0.1 represents the set {Δ(X j )} 0.1 upper quantile; (6) Obtain the slow features S of the training data X = [S1, S2, ..., S M ]∈R n×M , where M<m, represents the number of main slow features, and n is the number of samples.

4. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: In step S3, the specific steps of calculating the probability slow feature P of the training data X using the sliding window method using the JS divergence are: The parameters of the sliding window algorithm are set as follows: the window width is set to H, the step size is set to 1, each sampling moment moves backward once, the number of data samples is n, and there are a total of N = n-H + 1 sampling moments; The normal operating condition data X0 in the historical database is used as the benchmark data, and the benchmark data slow feature S is obtained after the same processing as steps S1 and S2. B =[S B1 ,S B2 ,...,S BM ]∈R n×M ; Calculate S B The mean and variance of B and the benchmark variance λ B ; Calculate the mean μ of the main slow feature S of the training data X at the kth sampling time S (k) and variance λ S (k); The probability slow feature P of the kth sampling moment of the training data X is calculated by formula (2). The expression of formula (2) is as follows: Calculate P = [P1, P2, ..., P M ]∈R N×M , N is the number of sampling moments, and M is the number of main slow features.

5. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: In step S5, the specific steps of constructing comparative learning sample pairs using the training data set are: The training data set is represented as D = {p i ,y i }, where p is the sample data, y∈{1,...,K} represents the fault label, and i∈I≡{1,2,...,N} represents the sample index; In the dataset D, for any two different samples [p i ,p j ], if y i =y j , that is, if the fault labels are the same, they constitute a positive sample pair, otherwise they constitute a negative sample pair; Get the positive sample pair set: Negative sample pair set: Total set of sample pairs: A = {P + ,P - }.

6. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: In step S6, in step S602, the input signal in the feature extraction network with 1D-CNN as the backbone passes through the input layer, the first convolution layer, the first pooling layer, the second convolution layer, the second pooling layer, the flattening layer, and the output layer in sequence. The specific parameters and functions are as follows: Input layer: accepts probability slow feature data as input signal p i ; The first convolution layer: The convolution kernel size is 3, which is used to extract the features of the input signal; First pooling layer: The pooling kernel size is 2, which is used to reduce the spatial size of each feature map; The second convolution layer: The convolution kernel size is 3, which is used to further extract the signal features after one convolution process; Second pooling layer: The pooling kernel size is 2, which is used to reduce the spatial size of each feature map; Flattening layer: flatten the obtained multi-dimensional features into a one-dimensional feature vector; Output layer: output the extracted features z i ; In step S603, the extracted feature z is used i Calculate the slow feature contrast loss L sf The specific steps are: Assuming the total number of samples in a training batch is N, the contrast loss is calculated as L sf Among them, the function dis(·) is used to calculate the Euclidean distance between two features, and τ is the temperature coefficient; The loss is trained to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs, making samples with the same label more similar and amplifying the differences between samples with different labels.

7. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: In step S7, the specific steps of training the Softmax classifier using the features extracted by the trained feature extractor are: Define the feature extraction operation up to step S6 as function F(·), then z i =F(x i ); First, one or more fully connected layers are used to transform z into i Converted to a vector of the same length as the number of categories, the output of the Softmax classifier is expressed as: Among them, θ is the fully connected layer parameter to be optimized; The Adam optimizer is used to optimize the parameter θ, and the loss function is as follows: Among them, y is the true label of the sample, and the parameters are retained after optimization.

8. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: In steps S9 and S10, online test data is obtained from a variety of sensors installed on industrial production equipment. After being processed, the data is input into the trained comparative learning model saved in step S8 to output minor fault diagnosis results.

9. The industrial process minor fault diagnosis method based on probabilistic slow feature contrast learning according to claim 1 is characterized in that: Steps S1 to S3 are the probabilistic slow feature extraction stage, steps S4 to S8 are the slow feature comparison learning training stage, and steps S9 to S10 are the online data fault diagnosis stage.

Images