Generalized zero-shot fault diagnosis method for thermal power feedwater pump groups based on fault similarity

Abstract

A generalized zero-sample fault diagnosis method for thermal power generation water supply pump groups based on fault similarity belongs to the field of generalized zero-sample fault diagnosis for thermal power generation water supply pump groups. The present invention extracts fault-related features from historical production data by constructing a feature extraction model of a fault-related stacked denoising autoencoder; based on shallow expert knowledge, a similarity matrix between seen faults and seen faults, as well as a similarity matrix between seen faults and unseen faults are constructed. The similarity matrix between seen faults is calculated using seen fault data, and combined with shallow expert experience, a similarity matrix mapping relationship between data and expert knowledge is formed; based on an offline model, seen faults and unseen faults of the water supply pump group are diagnosed in real time. The present invention can diagnose the fault type of the water supply pump group in real time, overcomes the dependence on deep-level knowledge such as the cause, location and degree of the fault, and the defects of domain offset, and ensures the normal operation of the water supply pump group in thermal power generation.

Description

Technical Field

[0001] The invention belongs to the field of generalized zero-sample fault diagnosis of thermal power generation water supply pump groups, and provides an effective zero-sample fault diagnosis method. Background Art

[0002] Thermal power generation is a crucial component of the global energy mix. As an efficient and economical method of generating electricity, coal-fired power plants are widely used in large-scale power supply. However, the efficient operation of thermal power plants depends not only on the coordinated functioning of primary equipment such as boilers, steam turbines, and generators, but also on the stability and reliability of auxiliary equipment. As one of the core auxiliary equipment in thermal power plants, feedwater pump sets play an indispensable role in the entire power generation system.

[0003] The primary function of a feedwater pump unit is to transport condensate back to the boiler, ensuring smooth water circulation within the boiler and maintaining the system's thermal balance and operating efficiency. Its performance directly impacts the boiler's steam generation rate, the turbine's power output, and the economic and safety of the entire power plant. Failure of the feedwater pump unit can prevent condensate from reaching the boiler in a timely manner, leading to serious problems such as dry-burning and pressure loss. This can not only damage equipment but also cause power plant downtime, impacting power supply and economic profitability, and even potentially causing environmental pollution and safety incidents.

[0004] In actual production, fault data is typically stored in historical production databases. Fault types can be determined based on operator experience and workshop production records. These are known as seen faults. However, as equipment ages and many parameters change over time, new, foreseeable faults may occur, or even multiple faults may occur simultaneously. These undocumented faults are known as unseen faults. Because these faults are not pre-detected or recorded in historical databases, they often cannot be diagnosed in a timely manner upon occurrence. Therefore, rapidly identifying and diagnosing both seen and unseen faults constitutes a classic industrial zero-sample fault diagnosis problem. Solving this problem can significantly improve production efficiency, reduce unnecessary economic losses, and even prevent casualties caused by equipment failure. Against the backdrop of the continuous development of advanced sensing and data processing technologies, the introduction of intelligent diagnostic systems will help improve the operational reliability of feedwater pump units and the overall performance of coal-fired power plants.

[0005] In traditional zero-shot fault diagnosis, models are only required to be able to diagnose unseen faults. However, in practical industrial applications, models are often expected to effectively diagnose both seen and unseen faults. To address this issue, researchers have proposed generalized zero-shot fault diagnosis methods. Furthermore, in industrial production, in addition to single faults, compound faults are often encountered. Compound faults refer to the simultaneous occurrence of multiple faults in a system, resulting in more complex system behavior and posing significant challenges to fault diagnosis. The prediction of compound faults is typically based on past single faults, and their predictability provides the potential for generalized zero-shot fault diagnosis. Currently, research on generalized zero-shot compound fault diagnosis has primarily focused on bearing faults, where compound faults can be viewed as linear combinations of multiple single faults. Therefore, auxiliary information for compound faults can be obtained by linearly combining the auxiliary information from multiple single faults. However, existing methods rely on attributes such as fault location, severity, and cause. However, obtaining these attributes requires a significant amount of deep, specialized expert knowledge. This deep knowledge requires a profound understanding of the entire industrial production process, including complex workflows and equipment characteristics. Acquiring this deep knowledge is often time-consuming and labor-intensive. In contrast, operators with average experience are more likely to provide readily available superficial knowledge based on documentation, a basic understanding of process mechanisms, and interactions with experts. Based on this superficial knowledge, a fault similarity matrix is constructed to calculate the membership of each sample relative to both seen and unseen faults, enabling diagnosis of unseen faults. To this end, this paper proposes a generalized zero-sample composite fault diagnosis method for water supply pump groups based on fault similarity. Summary of the Invention

[0006] The present invention aims to address the technical problems of low fault diagnosis efficiency and poor diagnostic accuracy in the prior art, and provides a generalized zero-sample fault diagnosis method for thermal power feedwater pump groups based on fault similarity. This method utilizes production data of historically observed faults to construct a fault-related stacked denoising autoencoder model, which can remove redundant information from the production data, improve resistance to noise, and enhance fault identification accuracy. Furthermore, a method for constructing a fault similarity matrix based on a combination of shallow expert experience and data-driven methods is proposed to effectively identify observed and unseen faults, and to implement real-time diagnosis of unseen faults, effectively improving diagnostic accuracy.

[0007] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0008] A generalized zero-shot fault diagnosis method for thermal power feedwater pump groups based on fault similarity involves offline establishment of a generalized zero-shot composite fault diagnosis model for feedwater pump groups and online diagnosis based on this model. Based on historical data, a fault-related stacked denoising autoencoder is constructed to extract fault-related features for preliminary classification. A fault similarity matrix is constructed using shallow expert experience, and the similarity matrix between observed faults is calculated in combination with historical production data, establishing a mapping relationship between data-driven and expert experience. A gating mechanism is used to initially distinguish between observed and unseen faults. Unseen faults are further distinguished using this mapping relationship and shallow expert experience. Through these steps, observed and unseen faults in feedwater pump groups are identified in real time, achieving accurate and real-time diagnosis of generalized zero-shot faults for feedwater pump groups.

[0009] The specific method is:

[0010] Step 1) Extract fault-related features from historical production data by building a fault-related stacked denoising autoencoder feature extraction model;

[0011] By utilizing fault-related stacked denoising autoencoders, valuable information is extracted and redundant information is eliminated, enhancing the model’s resistance to noise interference, accurately identifying and separating key feature information, and completing the classification of fault types. The model training is divided into two stages.

[0012] Phase 1: Layered pre-training; training multiple basic denoising autoencoders one by one;

[0013] The construction process of the first denoising autoencoder is divided into the following 5 steps:

[0014] Step 1: Build the input layer; define the number of input layer nodes to be consistent with the original data dimension, and receive training samples of historical production data;

[0015] Step 2: Add a corruption layer; perform random corruption on the input data, including randomly setting zeros or adding Gaussian noise to generate noisy input data;

[0016] Step 3: Construct the encoding layer. After the destruction layer, an encoding layer is constructed to map the noisy input into a low-dimensional feature representation. The encoding layer consists of multiple neural network layers, including linear layers, activation function layers, and possible regularization layers.

[0017] Step 4: Construct the decoding layer. The low-dimensional feature representation obtained in the previous step is input into the decoding layer, and the low-dimensional feature is restored to an output close to the original input data. The number of output nodes is consistent with the input layer. The structure of the encoding layer is opposite to that of the decoding layer. Its purpose is to use the low-dimensional features to reconstruct the original clean input data.

[0018] Step 5: Train the parameters of the autoencoder; use the original uncorrupted data as the reconstruction target, define the mean square error as the loss function, select the Adam optimizer and set the learning rate and batch size hyperparameters, and use the backpropagation algorithm to optimize the parameters of the encoder and decoder;

[0019] The construction process of the second denoising autoencoder is similar to that of the first one and is divided into the following five steps:

[0020] Step 1: Construct the input layer; fix the parameters of the first denoising autoencoder and use the low-dimensional features output by its encoding layer as the input data of the second denoising autoencoder;

[0021] Step 2: Add a corruption layer; inject noise into the input features, including random zero or Gaussian noise, to generate noisy features;

[0022] Step 3: Construct the encoding layer. The construction process is the same as the encoding layer of the first autoencoder, which further compresses the noisy features to obtain lower-dimensional features that contain deeper information.

[0023] Step 4: Construct the decoding layer: The construction process is consistent with the decoding layer of the first denoising autoencoder, reconstructing the lower-dimensional features into the undestructed features of the input layer;

[0024] Step 5: Train the autoencoder parameters. Fix all parameters of the first denoising autoencoder and optimize only the parameters of the second denoising autoencoder. The loss function, optimizer, and hyperparameter settings are the same as those of the first denoising autoencoder.

[0025] The construction process of the third denoising autoencoder is similar to that of the above denoising autoencoder. The output of the encoding layer of the second denoising autoencoder is used as the input data of the third denoising autoencoder. During training, the parameters of the second denoising autoencoder are fixed, and only the relevant parameters of the current denoising autoencoder are optimized.

[0026] The second stage is global fine-tuning. After completing the layered pre-training, the stacked denoising autoencoder is globally optimized end-to-end. The input layer, noise layer, and encoding layers of multiple denoising autoencoders obtained by layered pre-training are stacked in sequence to form a feature extraction network. ; The initialization parameters inherit the parameters of the layered pre-training stage and will have N tr Historical production data of samples X tr Input stacking network, the network encodes the input data layer by layer and generates a high-level feature representation H; this high-level feature representation is used in two parts:

[0027] The first part: reconstruct the original data; build a decoding layer, which consists of multiple neural network layers, including linear layers and activation function layers, and input high-level feature representations into the decoding layer; the decoding layer decodes the high-level features and outputs reconstructed data with the same dimension as the original input data ; Calculate the mean square error between the reconstructed data and the original data as the reconstruction loss. The reconstruction loss expression is as follows:

[0028] ;in and They are X tr and The nth sample of

[0029] The second part: fault classification; establish a SoftMax regression model as a classifier to distinguish between seen and unseen faults and determine the category of a specific seen or unseen fault, map the high-level feature representation to the probability distribution of the fault category and obtain the fault category judged by the classifier as , using cross entropy as the classification loss function, the expression of the cross entropy loss function is as follows:

[0030] ; Where p represents the category, 1{ =p} is the indicator function, that is, if When , its value is 1, otherwise it is 0;

[0031] Combining the above two parts of loss, the overall loss function of the model is defined as follows: ; Among them, α1 and α2 are weight parameters, and the back propagation algorithm is used to optimize the parameters of the entire network.

[0032] Step 2) Based on shallow expert knowledge, a similarity matrix between seen faults and seen faults is constructed, as well as a similarity matrix between seen faults and unseen faults. The similarity matrix between seen faults is calculated using seen fault data, and combined with shallow expert experience, a similarity matrix mapping relationship is formed between the data and expert knowledge.

[0033] Step 1: Based on the shallow expert experience and historical production data, a similarity matrix between the observed faults and the observed faults of the water supply pump group is constructed:

[0034] ; and the similarity matrix between seen and unseen faults

[0035] ;

[0036] Step 2: Using the historical data of observed faults, the feature center of each observed fault category is calculated to further obtain the similarity matrix between the observed fault categories. First, based on the observed fault samples in the historical production data of the water supply pump group, the feature center of each observed fault category is calculated. The feature center of the pth category of observed faults is defined as:

[0037] ;in ∈H tr is the nth sample of the pth type of seen fault, is the number of samples of the p-th type of fault, and then the similarity between the feature centers of any two seen fault categories is calculated, which is defined as:

[0038] Finally, we get the similarity matrix between the observed fault categories calculated based on the observed fault data:

[0039] ;

[0040] Step 3: Combined with the similarity matrix obtained from shallow expert experience, the similarity matrix mapping relationship between data and shallow expert knowledge is obtained:

[0041] ; Through mapping relationships, a similarity diagnosis basis combining data-driven and expert knowledge is realized, which is used to distinguish and diagnose the seen faults and unseen faults of the water supply pump group in the online stage.

[0042] Step 3) Based on the offline model, diagnose the water supply pump group's observed and unseen faults in real time.

[0043] 3.1) Input the test sample data from online monitoring into a pre-trained feature extractor to obtain the feature representation corresponding to the test sample. Input the obtained test sample features into a pre-trained classifier to obtain the posterior probability distribution of the test sample belonging to each class of observed faults.

[0044] 3.2) Compare the maximum probability value of the posterior probability distribution obtained in 3.1) with the set gating threshold to distinguish whether the test sample belongs to a seen fault or a not-seen fault. When the maximum posterior probability value is greater than the gating threshold, the test sample is determined to be a seen fault. When the maximum posterior probability value is less than or equal to the gating threshold, the test sample is determined to be a not-seen fault.

[0045] 3.3) For the test sample identified as a seen fault in step 3.2), its fault category is determined to be the seen fault category with the highest posterior probability:

[0046] ; where g lpis the posterior probability that the lth test sample belongs to the pth type of seen fault;

[0047] 3.4) For the test samples determined as having no faults in step 3.2), calculate the similarity between the sample features and the feature centers of each seen fault category obtained during the offline training phase:

[0048] ;in is the feature of the lth test sample, is the feature center of the p-th type of seen fault; the similarity ω between the current test sample and each seen fault is obtained l =[ ,……, ], using the similarity matrix mapping relationship between offline constructed data and shallow expert knowledge, the mapped The similarity w of the qth type of unseen fault obtained by shallow expert knowledge in the offline stage relative to each seen fault q ∈W Q,P , q=1,2,……,Q for comparison to determine the unseen fault category to which the current sample belongs:

[0049] ;

[0050] 3.5) Based on the diagnostic results of steps 3.3) and 3.4), accurate classification and diagnosis of the observed faults and unobserved compound faults of the water supply pump group during the online phase are achieved.

[0051] The beneficial effects of the present invention are:

[0052] (1) In traditional autoencoder training, the extracted features may not always be beneficial to a specific task, and the extracted features may be redundant. Therefore, the present invention constructs a fault-related stacked denoising autoencoder to extract fault-related features, significantly enhancing the ability to identify fault features in production data. At the same time, it effectively overcomes the problem of low diagnostic accuracy of traditional models due to data noise interference;

[0053] (2) Traditional zero-sample fault diagnosis methods often rely on attributes such as the location, extent, and cause of the fault, and have limited ability to identify unseen faults. Therefore, the present invention uses shallow expert knowledge and historical data to construct a fault similarity matrix, and establishes a similarity mapping relationship with expert experience in a data-driven manner, enabling the diagnostic model to efficiently and accurately achieve real-time diagnosis of unseen faults. The present invention can not only classify single faults, but also accurately classify unseen compound faults, better meeting practical needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is the working flow diagram of the water supply pump group;

[0055] Figure 2 This is the structure diagram of the generalized zero-sample fault diagnosis model for the water pump group;

[0056] Figure 3 Schematic diagram of seen and unseen faults diagnosed based on posterior probability samples;

[0057] Figure 4 The accuracy graph of each indicator under different numbers of training samples. DETAILED DESCRIPTION

[0058] The following specific embodiment has been put into practical use in a thermal power plant water pump workshop and has achieved remarkable results. Figure 1 As shown, the structure of the generalized zero-sample fault diagnosis model for the water supply pump group is as follows: Figure 2 shown.

[0059] In the offline modeling stage, a feature extraction model of a fault-related stacked denoising autoencoder is constructed to remove redundant information from the training data, extract fault-related features from historical production data, and use shallow expert knowledge to construct a fault similarity matrix. The similarity matrix between the seen faults calculated by combining the available seen fault data is analyzed and the relationship between the similarity matrix provided by the expert is analyzed. In the online diagnosis stage, a gating mechanism is used to divide the test samples into seen faults and unseen faults, and the fault similarity matrix is combined to diagnose which fault category the test samples belong to, completing real-time diagnosis of seen and unseen faults. The seen and unseen faults that the samples are diagnosed as are as follows: Figure 3 shown.

[0060] The offline data classification and recognition algorithm of the present invention is used to model and diagnose the collected data online. In order to prove the effectiveness of the proposed method, the following indicators are defined: G S =N Ts / N S is the accuracy rate of the seen samples being diagnosed as seen faults, G U =N Tu / N u is the accuracy rate of unseen samples being diagnosed as unseen faults, where N s and N u are the number of seen test samples and the number of unseen test samples respectively; N Ts is the number of seen samples diagnosed as seen faults, N Tu is the number of unseen samples diagnosed as unseen faults.

[0061] The accuracy of seen samples and unseen samples that are correctly diagnosed as the corresponding fault type is defined as follows:

[0062]

[0063] Acc S and Acc U Respectively represent the accuracy of seen samples and unseen samples that are correctly diagnosed as the corresponding fault type, N Cs and N Cu Represented as the number of correctly diagnosed seen samples and unseen samples respectively. Two comprehensive indicators are used to evaluate the performance of the model:

[0064]

[0065] Among them, G H Acc is the comprehensive accuracy of gated diagnosis for seen faults and unseen faults. H Indicates the accuracy of further diagnosis of specific fault types based on gated diagnosis of seen faults and unseen faults. Figure 4 For a fixed number of 1800 test samples, the accuracy of various metrics for different numbers of training samples is shown, demonstrating the impact of varying numbers of training samples on model performance. The total number of training samples used was 800, 1600, 2400, and 3200, respectively. It can be seen that the performance metrics of the diagnostic model generally improve with an increase in the number of training samples. This suggests that more training samples provide richer information about observed faults, thereby improving the performance of the proposed model.

[0066] To compare the fault diagnosis performance of the model, the model of this embodiment is compared with several common zero-shot fault diagnosis methods, including direct attribute prediction (DAP), attribute label embedding (ALE), and SSHTN. DAP and ALE are traditional zero-shot learning methods that exclude seen classes during the test phase, while SSHTN includes both seen and unseen classes during the test phase. This embodiment compares the fault diagnosis performance of these three models using the various evaluation metrics proposed above.

[0067]

[0068] As can be seen from the table above, the model proposed in this invention shows performance that surpasses other existing methods in all evaluation indicators. Among them, the DAP and ALE methods are mainly used to solve the zero-shot learning problem, which may lead to domain shift problems. SSHTN solves the generalized zero-shot learning problem, which not only alleviates the domain shift problem but also improves accuracy. However, the above methods all rely on deep knowledge-based attribute information. The FSGZSCFD model proposed in this invention not only effectively alleviates the domain shift problem, but also uses shallow knowledge to obtain the fault similarity matrix for generalized zero-shot fault diagnosis. It can still maintain high prediction accuracy and reliability when information may be incomplete or difficult to obtain.

Claims

1. A generalized zero-sample fault diagnosis method for thermal power feedwater pump groups based on fault similarity, characterized by: The steps are: Step 1) Extract fault-related features from historical production data by building a fault-related stacked denoising autoencoder feature extraction model; Step 2) Based on shallow expert knowledge, a similarity matrix between seen faults and seen faults, as well as a similarity matrix between seen faults and unseen faults, is constructed; Use the observed fault data to calculate the similarity matrix between the observed faults, and combine it with shallow expert experience to form a similarity matrix mapping relationship between data and expert knowledge; The specific method of constructing the fault similarity matrix module is: Step 1: Based on the shallow expert experience and historical production data, a similarity matrix between the observed faults and the observed faults of the water supply pump group is constructed: and the similarity matrix between seen and unseen faults ； Step 2: Using the historical data of observed faults, the feature center of each observed fault category is calculated to further obtain the similarity matrix between the observed fault categories. First, based on the observed fault samples in the historical production data of the water supply pump group, the feature center of each observed fault category is calculated. The feature center of the pth category of observed faults is defined as: ;in ∈H tr is the nth sample of the pth type of seen fault, is the number of samples of the p-th type of fault, and then the similarity between the feature centers of any two seen fault categories is calculated, which is defined as: Finally, we get the similarity matrix between the observed fault categories calculated based on the observed fault data: ； Step 3: Combined with the similarity matrix obtained from shallow expert experience, the similarity matrix mapping relationship between data and shallow expert knowledge is obtained: A similarity diagnosis basis combining data-driven and expert knowledge is realized through mapping relationships, which is used to distinguish and diagnose the observed and unseen faults of the water supply pump group in the online stage; Step 3) Based on the offline model, diagnose the water supply pump group's observed and unseen faults in real time.

2. The generalized zero-sample fault diagnosis method for a thermal power generation water pump group based on fault similarity according to claim 1 is characterized by: In step 1), the specific solution is: using a fault-related stacked denoising autoencoder to extract valuable information and eliminate redundant information, enhance the model's resistance to noise interference, identify and separate key feature information, and complete fault type differentiation; model training is divided into two stages.

3. The generalized zero-sample fault diagnosis method for a thermal power generation water pump group based on fault similarity according to claim 2 is characterized by: The two stages of model training are specifically: Phase 1: Layered pre-training; training multiple basic denoising autoencoders one by one; The construction process of the first denoising autoencoder is divided into the following 5 steps: Step 1: Build the input layer; define the number of input layer nodes to be consistent with the original data dimension, and receive training samples of historical production data; Step 2: Add a corruption layer; perform random corruption on the input data, including randomly setting zeros or adding Gaussian noise to generate noisy input data; Step 3: Construct the encoding layer; An encoding layer is constructed after the destruction layer to map the noisy input into a low-dimensional feature representation; The encoding layer consists of multiple neural network layers, including linear layers, activation function layers, and regularization layers; Step 4: Construct the decoding layer; input the low-dimensional feature representation obtained in the previous step into the decoding layer, and restore the low-dimensional features to an output close to the original input data. The number of output nodes is consistent with the input layer; the structure of the encoding layer is opposite to that of the decoding layer, and the low-dimensional features are used to reconstruct the original clean input data; Step 5: Train the parameters of the autoencoder; use the original uncorrupted data as the reconstruction target, define the mean square error as the loss function, select the Adam optimizer and set the learning rate and batch size hyperparameters, and use the backpropagation algorithm to optimize the parameters of the encoder and decoder; The construction process of the second denoising autoencoder is similar to that of the first one and is divided into the following five steps: Step 1: Construct the input layer; fix the parameters of the first denoising autoencoder and use the low-dimensional features output by its encoding layer as the input data of the second denoising autoencoder; Step 2: Add a corruption layer; inject noise into the input features, including random zero or Gaussian noise, to generate noisy features; Step 3: Construct the encoding layer. The construction process is the same as the encoding layer of the first autoencoder, which further compresses the noisy features to obtain lower-dimensional features that contain deeper information. Step 4: Construct the decoding layer: The construction process is consistent with the decoding layer of the first denoising autoencoder, reconstructing the lower-dimensional features into the undestructed features of the input layer; Step 5: Train the autoencoder parameters. Fix all parameters of the first denoising autoencoder and optimize only the parameters of the second denoising autoencoder. The loss function, optimizer, and hyperparameter settings are the same as those of the first denoising autoencoder. The construction process of the third denoising autoencoder is similar to that of the above denoising autoencoder. The output of the encoding layer of the second denoising autoencoder is used as the input data of the third denoising autoencoder. During training, the parameters of the second denoising autoencoder are fixed, and only the relevant parameters of the current denoising autoencoder are optimized. The second stage is global fine-tuning. After completing the layer-wise pre-training, the stacked denoising autoencoders are globally optimized end-to-end. The input layer, noise layer and encoding layers of multiple denoising autoencoders obtained by layered pre-training are stacked in sequence to form a feature extraction network ; The initialization parameters inherit the parameters of the layered pre-training stage and will have N tr Historical production data of samples X tr Input stacking network, the network encodes the input data layer by layer to generate high-level feature representation H; This high-level feature representation is used in two parts: The first part: reconstructs the original data; builds the decoding layer, which consists of multiple neural network layers, including linear layers and activation function layers, and inputs high-level feature representations into the decoding layer; The decoding layer decodes the high-level features and outputs reconstructed data with the same dimension as the original input data. ; Calculate the mean square error between the reconstructed data and the original data as the reconstruction loss. The reconstruction loss expression is as follows: ;in and They are X tr and The nth sample of The second part: fault classification; establish a SoftMax regression model as a classifier to distinguish between seen and unseen faults and determine the category of a specific seen or unseen fault, map the high-level feature representation to the probability distribution of the fault category and obtain the fault category judged by the classifier as , using cross entropy as the classification loss function, the expression of the cross entropy loss function is as follows: ; Where p represents the category, 1{ =p} is the indicator function, that is, if When , its value is 1, otherwise it is 0; Combining the above two parts of loss, the overall loss function of the model is defined as follows: , where α1 and α2 are weight parameters, and the back propagation algorithm is used to optimize the parameters of the entire network.

4. The generalized zero-sample fault diagnosis method for a thermal power generation water pump group based on fault similarity according to claim 1 is characterized by: In the step 3), the specific method is: 3.1) Input the test sample data from online monitoring into a pre-trained feature extractor to obtain the feature representation corresponding to the test sample. Input the obtained test sample features into a pre-trained classifier to obtain the posterior probability distribution of the test sample belonging to each class of observed faults. 3.2) Compare the maximum probability value of the posterior probability distribution obtained in 3.1) with the set gating threshold to distinguish whether the test sample belongs to a seen fault or a not-seen fault. When the maximum posterior probability value is greater than the gating threshold, the test sample is determined to be a seen fault. When the maximum posterior probability value is less than or equal to the gating threshold, the test sample is determined to be a not-seen fault. 3.3) For the test sample identified as a seen fault in step 3.2), its fault category is determined to be the seen fault category with the highest posterior probability: ; where g lp is the posterior probability that the lth test sample belongs to the pth type of seen fault; 3.4) For the test samples determined as having no faults in step 3.2), calculate the similarity between the sample features and the feature centers of each seen fault category obtained during the offline training phase: ;in is the feature of the lth test sample, is the characteristic center of the p-th type of observed fault; Get the similarity ω between the current test sample and each seen fault l =[ ,……, ], using the similarity matrix mapping relationship between offline constructed data and shallow expert knowledge, the mapped The similarity w of the qth type of unseen fault obtained by shallow expert knowledge in the offline stage relative to each seen fault q ∈W Q,P , q=1,2,……,Q for comparison to determine the unseen fault category to which the current sample belongs: ； 3.5) Based on the diagnostic results of steps 3.3) and 3.4), accurate classification and diagnosis of the observed faults and unobserved compound faults of the water supply pump group during the online phase are achieved.

Images