Semi-supervised fault diagnosis system based on TCN network and deep clustering

Abstract

The present invention relates to the technical field of fault diagnosis of hydropower units, and discloses a semi-supervised fault diagnosis system based on TCN network and deep clustering. By designing system function modules, the peak-to-peak value of the vibration signal of the stability measurement point of the hydropower unit and the alarm signal of the unit over-alarm fault are obtained, a data set is constructed and annotated, and the training data is divided into labeled data and unlabeled data; based on the labeled data and unlabeled data after noise reduction, a semi-supervised model is trained through a TCN metric network, deep features in time and space are extracted, and pseudo labels are generated according to the k-means clustering algorithm, and according to the annotated true labels, clustering loss and distribution alignment loss are used as dual loss functions to optimize the model, and a fault diagnosis model is constructed to perform fault diagnosis on the hydropower unit. The present invention improves the accuracy and efficiency of fault diagnosis, reduces the cost and resource requirements in practical applications, and has a wide range of engineering application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower unit fault diagnosis, and in particular to a semi-supervised fault diagnosis system based on TCN network and deep clustering. Background Art

[0002] As a clean and renewable energy production facility, hydropower stations are playing an increasingly significant role. As the core equipment of hydropower stations, turbines are an important link in realizing the conversion of water energy into electrical energy. However, due to the complexity of the equipment and the ever-changing operating environment, turbines may face various potential faults, which may not only reduce the operating performance of the equipment and increase the operating cost, but also cause serious safety accidents. Therefore, it is particularly important to diagnose turbine faults in a timely and accurate manner. Existing deep learning-based methods perform well, but such methods are a kind of supervised learning method and there are still some issues worth considering. In order to train a deep learning model with good generalization ability, the unit operation data needs to be labeled with fault labels or normal labels in advance, which involves a lot of manpower investment and prior knowledge. As a result, in the actual training process, the fault diagnosis model often faces the problem of insufficient fully labeled training data, which makes the trained model perform unsatisfactorily in terms of generalization ability and robustness.

[0003] To address this problem, the prior art applies unsupervised clustering algorithms to vibration fault diagnosis of hydropower units. Unsupervised clustering algorithms can automatically cluster samples with similar attributes based on the attributes of the samples themselves. In the learning stage of diagnosis, it is not necessary to know the class labels of the samples in advance, and new faults can be re-clustered according to their attributes. Therefore, from the perspective of the practicality of the diagnostic method, it is more in line with engineering practice.

[0004] Therefore, it is necessary to consider how to use limited prior knowledge (small samples) to guide the clustering process and improve the accuracy of the fault diagnosis model. Based on the above considerations, the present invention proposes a new deep metric clustering framework. Specifically, the TCN network is used to extract the deep feature representation of the hydropower unit in time and space. Then, the pseudo-label module clusters the labeled data using the nearest neighbor selection to generate reliable pseudo-labels for the unlabeled data. Finally, the network fine-tuning module uses the true label and pseudo-label to minimize the double loss (softmax loss and distribution alignment loss) to fine-tune the lightweight TCN, iteratively co-train, that is, compare the Euclidean distance between the sample to be diagnosed and each cluster center, divide the sample into the category represented by the cluster center closest to it in the feature space, and update the cluster center so as to generate better pseudo-labels in subsequent training and further improve the diagnosis model. In this way, as the diagnosis work progresses, new samples will be continuously added to the diagnosis model, that is, the diagnosis model will continuously learn new "knowledge", thereby adjusting the cluster center, so that the diagnosis model is continuously improved.

[0005] However, the existing deep learning-based and unsupervised clustering-based methods perform well in hydropower unit fault diagnosis, but still have the following defects:

[0006] 1. Defects of deep learning methods: Deep learning methods belong to supervised learning and require a large amount of labeled data to train the model. Obtaining high-quality fault labels and normal labels requires a lot of expert knowledge and manpower investment, resulting in high labeling costs. However, in practical applications, the fault data of hydropower units is often scarce and unbalanced, resulting in insufficient training data. This makes the deep learning model perform poorly in terms of generalization ability and robustness, and is prone to overfitting.

[0007] 2. Defects of unsupervised clustering methods: Unsupervised clustering algorithms classify samples based on their own attributes, but they cannot fully utilize valuable prior knowledge, resulting in insufficient diagnostic accuracy and reliability. In the clustering process, unsupervised methods do not rely on any label information and cannot utilize some of the labeled data that has already been obtained. This completely unsupervised method is difficult to achieve the high diagnostic accuracy of supervised methods, and is also affected by initial conditions and parameter settings, resulting in instability and poor consistency in clustering results. In addition, unsupervised clustering methods cannot automatically and effectively identify and classify new fault types, affecting the adaptability and scalability of the diagnostic system. Summary of the invention

[0008] In order to solve the above problems, the purpose of the present invention is to provide a semi-supervised fault diagnosis system for hydropower units based on TCN network and k-means clustering, which is used for fault diagnosis of hydropower units and aims to make full use of prior knowledge, reduce annotation costs, improve diagnostic accuracy, and enhance the adaptability of new data.

[0009] In order to achieve the above technical objectives, the present application provides a semi-supervised fault diagnosis system based on TCN network and deep clustering, which is applied to the fault diagnosis of hydropower units, including:

[0010] The data processing module is used to obtain the peak-to-peak value of the vibration signal of the stability measurement point of the hydropower unit and the unit over-alarm fault alarm signal, build and annotate the data set, and divide the training data into labeled data and unlabeled data;

[0011] The model training module is used to perform semi-supervised model training based on the denoised labeled data and unlabeled data through the TCN metric network, extract deep features in time and space, and generate pseudo labels based on the k-means clustering algorithm;

[0012] The model optimization and fault diagnosis module is used to optimize the model based on pseudo labels and the labeled true labels, using clustering loss and distribution alignment loss as dual loss functions to build a fault diagnosis model for fault diagnosis of hydropower units.

[0013] Preferably, the data processing module is also used to perform noise reduction processing on the acquired signal of the hydropower unit through the ICEEMDAN algorithm.

[0014] Preferably, the data processing module is also used to extract valid alarm signal subsequences based on the data set after noise reduction processing as prior knowledge, and mark them as fault data. Similarly, the sequence signals when the unit is running stably are selected and marked as normal data.

[0015] Preferably, the model training module is further used to divide the labeled data into intra-class data and inter-class data, and to construct a distribution alignment loss by maximizing the distribution difference between inter-class data and minimizing the distribution difference between intra-class data.

[0016] Preferably, the loss function of the TCN metric network is expressed as:

[0017] ,

[0018] In the formula, and represents the subset of the kth fault class The samples in represents the subset of the kth normal class The samples in represents data sets from the same category, Represents data sets from different categories, M is the number of network layers, represents the trade-off parameter that controls the degree to which the k-th class samples are brought closer, It represents the trade-off parameter that controls the degree to which the k-th class sample is away from other class samples. represents the Frobenius norm of matrix A, λ is the regularization parameter, and They are the weight matrix and bias term of the deep metric network, and the weighted parameter and is a weighting parameter designed to balance the impact of imbalanced classes in the loss function.

[0019] Preferably, the model training module is further used to cluster the labeled features and the unlabeled features using a K-means clustering algorithm, generate pseudo labels for the unlabeled data through nearest neighbor selection (NNS), and assign the label of the feature center to the unlabeled sample with the smallest distance from the feature center.

[0020] Preferably, the model optimization and fault diagnosis module is also used to optimize the labeled samples and the unlabeled samples through the distribution loss CL and the distribution alignment loss DA, so as to prevent the labels generated by the pseudo-label module from being biased towards the majority class.

[0021] Preferably, the model optimization and fault diagnosis module is further used to obtain the distribution alignment loss DA according to the Kullback-Leibler divergence between the feature distributions of the unlabeled samples and the original samples.

[0022] Preferably, the model optimization and fault diagnosis module is also used to optimize the model by using CL loss and DA loss to refine network parameters after optimizing the deep metric network in the deep metric learning module based on pseudo labels and true labels.

[0023] The present invention discloses the following technical effects:

[0024] The present invention reduces the reliance on a large amount of manual labeling by utilizing a small amount of labeled data (prior knowledge) and a large amount of unlabeled data, thereby reducing the labeling cost. Combined with the semi-supervised learning method, the existing partial labeled data is used to improve the model performance. Prior knowledge is introduced into the clustering process, and the cluster centers are continuously updated and optimized through pseudo-label generation and collaborative training, so that the diagnostic model can learn new "knowledge" and improve diagnostic performance. In this way, prior knowledge can be utilized and the changes in data can be automatically adapted.

[0025] The present invention combines deep learning and unsupervised clustering methods, uses the TCN algorithm to fully extract the deep features of data in time and space, and performs feature extraction and classification on the data; uses a clustering algorithm to cluster the deep features, and uses prior knowledge to continuously improve the accuracy of clustering, thereby improving the accuracy and efficiency of hydropower unit fault diagnosis.

[0026] The present invention not only improves the accuracy and efficiency of fault diagnosis of hydropower units, but also reduces the cost and resource requirements in practical applications, and has broad engineering application prospects. This collaborative learning framework will provide strong technical support for the intelligent operation and maintenance of hydropower units, and help ensure the safe and stable operation of hydropower stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 is a flow chart of the ICEEMDAN algorithm of the present invention;

[0029] Figure 2 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0030] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0031] like Figure 1-2 As shown, the present invention provides a semi-supervised fault diagnosis technology based on TCN network and deep clustering, which is applied to the fault diagnosis of hydropower units, including the following contents:

[0032] 1. Data preparation:

[0033] The peak-to-peak signal of the stability measurement point and the alarm signal of the unit over-alarm fault obtained from the online monitoring system are subjected to noise reduction processing by the ICEEMDA algorithm to filter out noise interference. Secondly, the effective alarm signal subsequence is extracted and this part of the data is used as prior knowledge, that is, labeled data (fault). Similarly, the sequence signal of the unit during stable operation is selected and marked as normal. and are labeled data and unlabeled data respectively. , every element in X belongs to R d , and x is d-dimensional data, which includes fault data and normal data. In addition, the labeled test samples are recorded as .

[0034] Among them, the denoising process of the ICEEMDA algorithm is as follows:

[0035] ICEEMDAN is a signal processing method improved on the basis of empirical mode decomposition (EMD). Different from the traditional EMD processing method, ICEEMDAN introduces adaptive noise and complete integration strategy to improve the stability and accuracy of decomposition. Unlike EMD, which directly adds Gaussian white noise during the decomposition process, ICEEMDAN selects the kth IMF component after the white noise is decomposed by EMD to add.

[0036] EMD is an adaptive, nonlinear signal decomposition method, mainly used to process nonlinear and non-stationary signals. Based on the local characteristic time scale of the signal itself, this method decomposes the complex signal into a series of simple, physically meaningful intrinsic mode functions (IMFs).

[0037] The flowchart of ICEEMDAN algorithm is as follows Figure 1 shown.

[0038] definition: is the original acquisition signal; Indicates the envelope calculation of the input signal to obtain the local mean of the input signal; represents the modal component IMF of the kth stage obtained by EMD decomposition; is the i-th Gaussian white noise that follows the N(0,1) distribution (mean 0, unit variance) ; is the residual of the kth stage; the coefficient represents the signal-to-noise ratio of the kth stage. When k=0, , when k ≥ 1, (in For the amplitude parameters set in advance, is the standard deviation operator); Represents an operator that computes an average.

[0039] The principle steps of ICEEMDAN decomposition algorithm are as follows:

[0040] In the original signal Here, white noise is added according to formula (1).

[0041] (1)

[0042] The signal obtained According to the EMD method, the i local means in formula (1) are calculated respectively: , and get the residual of stage k=1 :

[0043] (2)

[0044] Calculate the component of the original signal at stage k=1, recorded as :

[0045] (3)

[0046] In the residual of stage 1, add white noise and calculate the signal The local mean of , according to formula (4), the residual of the k=2th stage is obtained :

[0047] (4)

[0048] Calculate the k=2 stage component of the original signal, denoted as :

[0049] (5)

[0050] when When , calculate the residual of the kth stage :

[0051] (6)

[0052] Calculate the component of the kth stage of the original signal, recorded as :

[0053] (7)

[0054] Return to step (6) and calculate the next k.

[0055] The original signal after ICEEMDAN decomposition is expressed as:

[0056] (8)

[0057] 2. Deep metric learning: In actual production, the number of fault data is seriously unbalanced compared to the number of normal samples. Therefore, this task belongs to class imbalance semi-supervised learning. , defined as the input space, is the output space, is the joint space of labeled data.

[0058] The training set is , where D L represents the set of label data, D U represents a collection of unlabeled data, , ,N and Q represent the number of elements in the two data sets respectively.

[0059] In descending order of cardinality The classes are sorted, that is and , imbalance ratio , label ratio Then, the labeled data is divided into and between classes , respectively:

[0060] (9)

[0061] Where k represents the kth class, Represents a class other than itself, represents data sets from the same category, Represents datasets from different classes.

[0062] Next, a multi-layer TCN network is used as a deep metric network to calculate the nonlinear representation of the input data. The output of the mth layer is:

[0063] (10)

[0064] , (11)

[0065] in, is the input vector, m is the number of network layers, It is m-1 The output of the layer, It represents the bias term of the mth layer. is the number of neurons in the mth layer. is the activation function, is based on Equi-parameterized mapping function.

[0066] Two input data and The pairwise relationship between them can be measured by the L2 distance between the nonlinear representations at the mth layer, as follows:

[0067] (12)

[0068] Next, we design a loss function for the output of the training pair at the top layer of the deep metric network. and ,in Bring similar pairs closer together, and The dissimilar pairs are further separated, and the specific formula is as follows:

[0069] (13)

[0070] in and represents the subset of the kth fault class The samples in represents the subset of the kth normal class The samples in and is the output of the top layer of the deep metric network. is the margin enforced between similar and dissimilar pairs.

[0071] Combining the above formulas, we can get the final metric learning loss function as follows:

[0072] (14)

[0073] in, represents the trade-off parameter that controls the degree to which the k-th class samples are brought closer, , represents the trade-off parameter that controls the degree to which the k-th class sample is away from other class samples, represents the Frobenius norm of matrix A, λ is the regularization parameter, and are the weight matrix and bias term of the deep metric network respectively. In addition, the weighted parameters are designed and is a weighting parameter designed to balance the impact of imbalanced classes in the loss function.

[0074] 3. Pseudo label generation:

[0075] Metric-based learning loss function For deep metric networks After feature space optimization, the feature space of unlabeled data can also be improved. Instead of using the model’s predictions as pseudo labels for unlabeled data, the pseudo-labeling module optimizes the unbiased assignment of pseudo labels based on the improved feature representation to further improve the performance of the minority class.

[0076] In the deep metric learning module, the present invention computes the label feature representation from the top layer of the deep metric network and unlabeled feature representation . Use clustering techniques to represent the markers The samples are grouped into clusters. Different clustering techniques can be used for this purpose. K-means is used in the present invention because it is easy to implement and has excellent computational efficiency on large-scale data sets. The feature center vector in is expressed as .

[0077] Then, based on the assumption that two input samples close to each other should have the same label, pseudo labels are generated for the unlabeled data through nearest neighbor selection (NNS). The label is assigned to the qth unlabeled sample with the smallest distance to the feature center , which can be expressed as:

[0078] (15)

[0079] in is the distance metric function defined in Eq. is with The nearest feature center with the smallest distance. That is, the pseudo label of the qth unlabeled sample and Then, the pseudo-labeled dataset It can be merged into the previous labeled dataset for the next round of training.

[0080] 4. Network fine-tuning:

[0081] In the network fine-tuning module, distribution loss (CL) and distribution alignment loss (DA) are used to optimize labeled and unlabeled samples to prevent the labels generated by the pseudo-label module from being biased towards the majority class.

[0082] Classification Loss (CL): As the number of training times in the deep metric learning module increases, pseudo labels from unlabeled data are gradually added to the original labeled data, resulting in changes in the number of majority class samples and minority class samples. Therefore, based on the gradually expanding dataset, an improved classification loss is designed for the classification performance of the model.

[0083] For unlabeled datasets Any unlabeled sample in , whose feature representation can be obtained from the deep metric learning module and pseudo labels They are pseudo-label modules. Two fully connected layers FC and a Softmax layer σ are added on top of the deep metric network trained in the deep metric learning module to estimate The probability of the class is as follows:

[0084] (16)

[0085] in, is the output of FC, so the CL loss is defined as follows:

[0086] (17)

[0087] in is the scaling factor, the weight parameter It reflects the change of imbalance ratio during training, t represents the tth epoch in training, Indicates the current imbalance ratio in the tth epoch, represents the initial imbalance ratio before training. With the addition of pseudo-labeled samples, the class imbalance problem in the training set is alleviated (i.e. The scaling factor is , which means that the number of easy-to-classify samples (i.e., samples with a high probability of being correctly classified in the majority class) will also increase. This is because in the presence of class imbalance, pseudo-labeled samples are more likely to be classified as the majority class. It is easy to classify correctly, while minority class samples are difficult to classify correctly. Can reduce The contribution of samples that are easy to classify in the model causes the model to pay more attention to samples that are difficult to classify correctly. Can reduce The contribution of majority class samples helps the model pay more attention to minority classes.

[0088] Distribution Alignment Loss (DA): When the majority class dominates in the unlabeled data, the class imbalance in the unlabeled data may make it difficult to align the minority class distribution. In order to ensure that the labeled data and pseudo-labeled data follow the same imbalanced class distribution, this paper designs an improved distribution alignment loss based on Kullback-Leibler divergence, which aims to minimize the distribution difference between intra-class data and maximize the distribution difference between inter-class data.

[0089] For pseudo-labeled datasets , the present invention calculates the qth unlabeled sample The Kullback-Leibler divergence between the original sample feature distribution and

[0090] (18)

[0091] in Represents the top layer of the trained deep metric network in deep metric learning The characteristic representation of Indicates the pseudo label module The kth estimated label of . Similarly for the labeled dataset ,Pseudo-labeled dataset It can also be divided into intra-class collections and inter-class collections Then, the DA loss can be expressed as:

[0092] (19)

[0093] Where C1 and C2 are trade-off parameters, Used to maximize the distribution difference between inter-class data, Used to minimize the intra-class distribution difference.

[0094] Dynamic addition. Using ML loss in the deep metric learning module After optimizing the depth metric network, the present invention uses CL loss and DA loss Refine the network parameters. For a given extended labeled dataset , the objective function is:

[0095] (20)

[0096] Where t represents the tth epoch in the training process, T represents the total number of training epochs, is a weight parameter that changes according to the number of training epochs. As the training process progresses, the goal of the network fine-tuning module gradually shifts from distribution alignment to classification optimization. As a result, the learning network can expand the decision boundary around minority samples, helping the model pay more attention to minority categories.

[0097] This invention combines the TCN-based metric network with the k-means clustering algorithm for the first time and applies it to the field of hydropower unit fault diagnosis. TCN is used to efficiently extract deep temporal and spatial features from time series data, while the k-means algorithm uses these features to automatically cluster and generate pseudo labels. This combination significantly reduces the reliance on a large amount of labeled data, and effectively improves the efficiency and accuracy of fault diagnosis through automatic clustering of deep features.

[0098] The present invention designs a new semi-supervised learning framework, which can use limited labeled samples and a large number of unlabeled samples to train the model. Through the iterative process of deep feature extraction, clustering, pseudo-label generation and model update, the accuracy and robustness of the fault diagnosis model are continuously improved. It effectively alleviates the dependence of traditional deep learning methods on large-scale labeled data and reduces the labor cost of data labeling.

[0099] The present invention uses clustering loss and distribution alignment loss as dual loss functions, and combines true labels and pseudo labels to optimize the model. By minimizing the dual losses, the generalization ability and robustness of the model are further improved, ensuring the accuracy of fault diagnosis.

[0100] The present invention uses the ICEEMDAN-PE-SVD algorithm to perform noise reduction processing on the original measurement point signal to improve the data quality. High-quality data is the basis of model accuracy, and noise reduction processing can effectively reduce the interference of noise on fault diagnosis.

[0101] The present invention can dynamically update cluster centers and model parameters, and continuously learn new "knowledge" as new samples are added. Through continuous learning, the diagnostic model can adapt to the ever-changing operating environment and improve the accuracy and timeliness of diagnosis.

[0102] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0103] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0104] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A semi-supervised fault diagnosis system based on TCN network and deep clustering, characterized by: Applied to fault diagnosis of hydropower units, including: The data processing module is used to obtain the peak-to-peak value of the vibration signal of the stability measurement point of the hydropower unit and the unit over-alarm fault alarm signal, build and annotate the data set, and divide the training data into labeled data and unlabeled data; A model training module is used to perform semi-supervised model training based on the labeled data and unlabeled data after noise reduction through a TCN metric network, extract deep features in time and space, and generate pseudo labels based on a k-means clustering algorithm; A model optimization and fault diagnosis module, which is used to optimize the model based on the pseudo-labels and the annotated true labels, using distribution loss and distribution alignment loss as dual loss functions, and construct a fault diagnosis model for fault diagnosis of the hydropower unit; The model optimization and fault diagnosis module is also used to optimize labeled samples and unlabeled samples through distribution loss CL and distribution alignment loss DA, so as to prevent the labels generated by the pseudo-label module from being biased towards the majority class; The model optimization and fault diagnosis module is further used to obtain the distribution alignment loss DA according to the Kullback-Leibler divergence between the feature distributions of the unlabeled samples and the original samples; DA loss is expressed as: ， Where C1 and C2 are trade-off parameters, Used to maximize the distribution difference between inter-class data, Used to minimize the intra-class distribution differences; The data processing module is further used to perform noise reduction processing on the acquired signal of the hydropower unit through the ICEEMDAN algorithm; The data processing module is also used to extract effective alarm signal subsequences based on the data set after noise reduction processing as prior knowledge and mark them as fault data. Similarly, the sequence signal when the unit is running stably is selected and marked as normal data; The model training module is further used to divide the labeled data into intra-class data and inter-class data, and to construct a distribution alignment loss by maximizing the distribution difference between inter-class data and minimizing the distribution difference between intra-class data; The loss function of the TCN metric network is expressed as: ， In the formula, and represents the subset of the kth fault class The samples in represents the subset of the kth normal class The samples in represents data sets from the same category, Represents data sets from different categories, M is the number of network layers, represents the trade-off parameter that controls the degree to which the k-th class samples are brought closer, It represents the trade-off parameter that controls the degree to which the k-th class sample is away from other class samples. represents the Frobenius norm of matrix A, λ is the regularization parameter, and They are the weight matrix and bias term of the deep metric network, and the weighted parameter and is a weighting parameter designed to balance the impact of imbalanced classes in the loss function. Pull similar pairs closer to each other, the loss function Separate dissimilar pairs further apart; The model training module is further used to cluster the labeled features and the unlabeled features using a K-means clustering algorithm, generate pseudo labels for the unlabeled data by nearest neighbor selection (NNS), and assign the label of the feature center to the unlabeled sample with the smallest distance from the feature center; The model optimization and fault diagnosis module is also used to optimize the model by using CL loss and DA loss to refine network parameters after optimizing the deep metric network in the deep metric learning module based on the pseudo-label and the true label.

Images