Load tap changer small sample fault diagnosis method based on wavelet elasticity measurement network

Abstract

The application relates to a small sample fault diagnosis method of an on-load tap changer based on a wavelet elasticity measurement network, relates to the technical field of on-load tap changer fault diagnosis, and is used for solving the problem of recognition precision of on-load tap changer fault diagnosis under the condition of small sample fault of on-load tap changer, which is difficult to obtain, has various fault types and complex differences between categories. The original mechanical vibration signal of the on-load tap changer is taken as input, and sequentially, the following steps are completed: learnable wavelet enhancement, multi-scale feature extraction, elasticity measurement modeling and fault category discrimination, so that end-to-end modeling from the original vibration signal to the diagnosis result is realized. The modules are mutually coordinated in the overall process, the wavelet enhancement improves the feature separability, the multi-scale feature extraction effectively describes the local and overall change features of the vibration signal, and the elasticity measurement mechanism adaptively adjusts the feature distance under the condition of small samples, so that the classification and discrimination ability of the model is improved.

Description

Technical Field

[0001] This invention relates to the field of on-load tap changer fault diagnosis technology, specifically a small-sample fault diagnosis method for on-load tap changers based on wavelet elasticity metric networks. Background Technology

[0002] Existing research on fault identification of vibration signals from on-load tap changers (OLTCs) primarily employs a combination of manual experience and traditional signal processing methods, such as extracting time-domain statistical features, frequency-domain features, and time-frequency analysis features, and integrating support vector machines, k-nearest neighbors, or random forest classification models for fault discrimination. While these methods have validated the feasibility of using vibration signals for OLTC fault diagnosis, their performance is highly dependent on feature engineering design, and their generalization ability is limited under complex operating conditions and multi-fault scenarios. With the development of deep learning technology, convolutional neural networks and recurrent neural network models have been gradually introduced into the field of OLTC fault diagnosis. Through end-to-end learning, they automatically extract discriminative features, achieving better diagnostic results under conditions of sufficient data scale and relatively stable operating conditions. Although deep learning methods have demonstrated strong feature learning capabilities in vibration signal fault diagnosis, their effectiveness is usually based on a large-scale and well-labeled dataset. However, in practical engineering scenarios, the cost of acquiring OLTC fault samples is high, and field testing conditions are limited, making it difficult to construct sufficient fault sample datasets through active experimentation. Fault sample acquisition relies on long-term operational accumulation or manual annotation, resulting in a limited scale and uneven distribution of fault vibration data available for modeling. Under these conditions, directly using traditional deep learning models often leads to overfitting, limiting the model's generalization ability. Modeling approaches that rely solely on expanding data scale are insufficient to meet the engineering requirements of OLTC fault diagnosis; therefore, there is an urgent need to introduce fault diagnosis strategies that can still adapt well under limited sample conditions.

[0003] Unlike traditional supervised learning, which relies on large-scale labeled data, few-shot learning (FSL) enables models to maintain strong generalization ability even under data-constrained conditions by uncovering common structures and latent patterns among tasks. Among numerous FSL methods, meta-learning, by learning transferable prior knowledge across multiple related tasks, allows models to quickly adapt to unknown tasks with only a small number of new samples, thus demonstrating unique advantages in engineering scenarios with limited samples. The meta-learning framework has achieved good results in few-shot fault diagnosis tasks in rotating machinery and wind power equipment. Overall, meta-learning shows strong potential in few-shot diagnosis of mechanical equipment, but its performance depends on capturing the discriminative features of signals. When the diagnostic object shifts from rotating machinery with stable periodic characteristics to switching equipment with more complex operating mechanisms, existing methods still have limitations in terms of effectiveness.

[0004] Although existing technologies have introduced small-sample learning methods into the field of mechanical or electrical equipment fault diagnosis, these methods mostly focus on the design of general model structures and do not adequately consider the strong transient, strong noise, and operating condition differences inherent in OLTC vibration signals. Their applicability in practical OLTC small-sample fault diagnosis still needs further research. Summary of the Invention

[0005] The purpose of this invention is to provide a small-sample fault diagnosis method for on-load tap changers based on wavelet elasticity metric networks, which solves the problem of identification accuracy in the diagnosis of small-sample faults in on-load tap changers where it is difficult to obtain fault samples, fault types are diverse, and differences between categories are complex.

[0006] The technical solution adopted by the present invention to solve its technical problem is: a small-sample fault diagnosis method for on-load tap changers based on wavelet elasticity metric networks, comprising the following steps.

[0007] S1, Data Acquisition.

[0008] Mechanical vibration signals from on-load tap changers (OLTCs) are collected and divided into training and test sets. The training and test sets are then further divided into support and query sets, respectively.

[0009] S2. Construct a wavelet-enhanced multi-scale feature encoder.

[0010] S2.1, Regarding the OLTC vibration signal x i Perform learnable wavelet transform on the vibration signal x i Mapping to a multi-scale time-frequency representation space highlights key frequency bands related to fault states.

[0011] S2.2 Construct a lightweight multi-scale convolution to extract features from the wavelet-enhanced vibration signal to obtain the features of the vibration signal at different time scales, and then concatenate the features to obtain the fused features.

[0012] S2.3, Introduction of channel attention mechanism.

[0013] Assume fusion features First, feature compression is performed on each channel using global average pooling: (1); where, This represents the feature value of the c-th channel at position i; channel importance is modeled using a gating mechanism: (2); where, The attention weights for each channel; Represents the Sigmoid function; and These are the learnable parameters for the fully connected layer; Represents the ReLU activation function; The channel description vector is used; the channel weights are scaled channel-wise with respect to the original features to obtain the attention-enhanced feature representation: (3); where β c This represents the attention weight corresponding to the c-th channel; This represents the feature vector of the c-th channel.

[0014] S3, Elasticity Measurement Modeling.

[0015] After obtaining the feature representation output by the wavelet-enhanced multi-scale feature encoder, an elasticity measure (EMM) is further introduced to adaptively adjust the measure scale for different categories.

[0016] Let the query set sample embedding after wavelet-enhanced multi-scale feature encoder be represented as follows: (4); where d is the dimension of the feature space; the prototype vector of class c is calculated from the mean of the samples in the support set. (5); where S c For the support set of class c; |S c | represents the sample size; E φ (·) represents the mapping function of the feature encoder; X j It is S c The j-th original input sample; define the unnormalized Euclidean distance. (6); where z im This represents the value of the embedding vector of the i-th query set sample in the m-th dimension; m is the feature dimension, m=1,2…d; P cm It is the component of the prototype vector of class c in the m-th dimension; an independent elasticity factor λ is introduced for each feature dimension m of each class c. cm >0, forming the elastic weight vector (7); Obtain the query set sample z i With the prototype p of class c c elastic distance (8); The probability P that a sample in the query set belongs to the k-th class k P is calculated using the Softmax function. k = (9); among which, For the query set sample; for Elastic distance between the prototype and category k N is the number of current task categories; d ij * It is the query set sample z i With the j-th prototype p j The elastic distance between them is calculated using formula (8); exp is the natural exponential function e x ;dk * It is the elastic distance of the k-th prototype; P k It is a query set sample The probability of belonging to the k-th class; the elastic classification loss of the model is defined as: (10); where Q k Let Φ be the query set of class k, and Ф be all learnable parameters of the feature encoder; a regularization term is introduced: (11).

[0017] S4. Fault type identification.

[0018] Through the objective function (12) Determine the fault type; among which, Let be the elasticity classification loss function of the elasticity measure EMM; α is the regularization coefficient. It is the regularization term of the elasticity factor.

[0019] Furthermore, the specific content of step S2.1 is as follows: Let the input vibration signal be... Where L is the signal length, the output of the j-th wavelet neuron is expressed as: (13); among which, For wavelet basis functions, Using Morlet wavelets: (14); w ji b is the weight matrix of the wavelet neuron; j a represents the wavelet translation coefficient; j is the wavelet scaling factor; M is the number of wavelet neurons.

[0020] Furthermore, in step S2.2, the multi-scale convolution content is as follows: S2.2.1 Let the input features be processed by learnable wavelet as follows: (15), C represents the number of channels; the multi-scale convolution module includes several parallel one-dimensional convolution branches. For the s-th convolution branch, s=1,2…S, the one-dimensional convolution operation is: (16); among which, k represents the kernel parameters for the s-th branch. s C is the kernel length corresponding to this branch; s b is the number of output channels for this branch; (s) This represents the bias term; * indicates a one-dimensional convolution operation. The feature representation output by the s-th convolutional branch.

[0021] Furthermore, in step S2.2, batch normalization (S2.2.2) is introduced after the convolution operation to standardize the features: (17); (18); among which, These are standardized features; Representation of features The mean; Representation of features The variance; A smoothing constant is used to prevent numerical instability; It is the final feature after translation and scaling; A learnable scale; These are learnable bias parameters.

[0022] Furthermore, in step S2.2, after step S2.2.2, step S2.2.3 is executed: the output features of all branches are concatenated along the channel dimension to obtain the fused features. (19); where Concat(⋅) represents the splicing operation along the channel dimension.

[0023] The beneficial effects of this invention are as follows: using the original OLTC mechanical vibration signal as input, it sequentially completes learnable wavelet enhancement, multi-scale feature extraction, elasticity measurement modeling, and fault category discrimination, achieving end-to-end modeling from the original signal to the diagnostic result. Each module collaborates with the others in the overall process. Wavelet enhancement improves feature separability, multi-scale feature extraction effectively characterizes the local and global changes in the vibration signal, and the elasticity measurement mechanism adaptively adjusts the feature distance under small sample conditions, thereby improving the model's classification and discrimination capabilities. In summary, addressing the problems of limited samples and insufficient feature discriminability in OLTC fault diagnosis, a prototype learning network composed of wavelet-enhanced multi-scale feature encoding and an elasticity measurement mechanism is constructed. Compared with fixed metrics or simple similarity measurement methods, WEMNet can more flexibly characterize the differences between samples in the feature space, providing a more discriminative distance metric basis for small sample classification, and offering an efficient solution with good generalization ability for small-sample mechanical fault diagnosis of OLTCs. Attached Figure Description

[0024] Figure 1 A structural diagram of a wavelet-enhanced multi-scale feature encoder;

[0025] Figure 2 This is a schematic diagram of an elasticity measuring device.

[0026] Figure 3 This is a flowchart of the overall WEMNet fault diagnosis process. Detailed Implementation

[0027] This invention focuses on the vibration signal of on-load tap changers and proposes a wavelet elasticity metric network (WEMNet) for small-sample fault diagnosis scenarios. This improves the recognition accuracy and model generalization ability of OLTCs in such scenarios. Under a meta-learning framework, learnable wavelet layers and multi-scale convolutional feature extraction structures are introduced to effectively enhance the expressive power of fault-related features. Simultaneously, an elasticity metric mechanism is introduced to adaptively learn feature dimension weights for different categories, enabling the model to dynamically adjust the distance scale based on intra-class distribution characteristics, thereby alleviating inter-class confusion and overfitting problems under small-sample conditions. Figure 3 As shown, the present invention provides a method for small-sample fault diagnosis of on-load tap changers based on wavelet elasticity metric networks, comprising the following steps.

[0028] S1, Data Acquisition

[0029] The mechanical vibration signal of the on-load tap changer (OLTC) is collected by a sensor, and the collected vibration signal is divided into a training set and a test set. The training set and the test set are then further divided into a support set and a query set, respectively.

[0030] S2. Construct a wavelet-enhanced multi-scale feature encoder.

[0031] To address the challenges of limited sample size, complex operating conditions, and significant non-stationary signal characteristics in OLTC fault diagnosis, a wavelet-enhanced multi-scale feature encoder (WMFE) is designed to avoid reliance on manual feature extraction. Figure 1 As shown, the OLTC vibration signal is frequency-adaptively mapped through a set of learnable wavelet kernels with different scales and center frequencies to obtain multi-channel wavelet enhancement features. On this basis, the features of the signal in different time ranges are captured by parallel one-dimensional convolutional branches using convolutional kernels of different lengths and fused in the channel dimension. Finally, a channel attention mechanism is introduced to adaptively weight the features of different channels and output the feature representation for metric learning.

[0032] S2.1, Regarding the OLTC vibration signal x i Perform learnable wavelet transform.

[0033] To enhance the model's ability to adaptively select different frequency components in the OLTC signal, the original OLTC vibration signal x was processed. i By performing a learnable wavelet transform and using a trainable wavelet kernel, the model automatically adjusts the scale and translation parameters of each wavelet neuron during training, transforming the original vibration signal x... i Mapping to a multi-scale time-frequency representation space highlights key frequency bands relevant to the fault state and suppresses irrelevant or interfering components, providing a frequency-friendly input representation for feature extraction.

[0034] Assume the input vibration signal Where L is the signal length, the output of the j-th wavelet neuron is expressed as: (1). Among them, For wavelet basis functions, Using Morlet wavelets: (2); w ji b is the weight matrix of the wavelet neuron; j These are the wavelet translation coefficients used to control the center frequency; a j denoted by the wavelet scaling coefficient, used to control the bandwidth; M represents the number of wavelet neurons. By introducing learnable wavelets, the model no longer relies on fixed partitions, but can adaptively select key frequency components in the vibration signal during training, thereby reducing the burden on the convolutional module to learn multi-scale features.

[0035] S2.2 Construct a lightweight multi-scale convolution to extract features from the wavelet-enhanced vibration signal, so as to obtain the features of the vibration signal at different time scales.

[0036] S2.2.1 Multi-scale convolution uses parallel one-dimensional convolution branches to model transient impacts, periodic fluctuations, and slow-changing trends in vibration signals using different receptive fields, thus preserving multi-scale structural differences at the feature level. Let the input features after learnable wavelet processing be... (3). Where C represents the number of channels. The multi-scale convolution module consists of several parallel one-dimensional convolution branches, each branch using a convolution kernel of a different size to correspond to a different temporal receptive field. For the s-th convolution branch, s=1,2…S, the one-dimensional convolution operation is expressed as: (4). Among them, k represents the kernel parameters for the s-th branch. s C is the kernel length corresponding to this branch; s b is the number of output channels for this branch; (s) This represents the bias term; * indicates a one-dimensional convolution operation. It is the feature representation output by the s-th convolutional branch.

[0037] S2.2.2 To alleviate the training instability caused by differences in feature distribution between different convolutional branches, batch normalization is introduced after the convolution operation to standardize the features: (5); (6). Among them, These are standardized features; Representation of features The mean; Representation of features The variance; A smoothing constant is used to prevent numerical instability; It is the final feature after translation and scaling; A learnable scale; These are learnable bias parameters.

[0038] The ReLU activation function is used to introduce nonlinear mapping capability: (7). After processing by each convolutional branch, feature representations at different time scales are obtained.

[0039] S2.2.3 To achieve multi-scale information fusion, the output features of all branches are concatenated along the channel dimension to obtain the fused features. (8). Wherein, Concat(⋅) represents the splicing operation along the channel dimension.

[0040] By using multi-scale convolution, the model can simultaneously capture short-term impact features and long-term evolution patterns on the wavelet-enhanced vibration signal, providing richer multi-scale information for feature discrimination.

[0041] S2.3, Introduction of channel attention mechanism.

[0042] Although multi-scale convolutional modules can extract rich feature representations, the contribution of features at different scales and frequency bands to fault diagnosis tasks varies. To further enhance the model's ability to focus on key information, a channel attention mechanism is introduced after multi-scale feature fusion. This mechanism aims to effectively suppress interference and random noise from invalid frequency bands by capturing the correlation between feature channels, thereby outputting cleaner multi-scale discriminative features.

[0043] Let the output features of the multi-scale convolution module be... First, feature compression is performed on each channel using global average pooling: (9). Among them, Let represent the feature value of the c-th channel at position i. Channel importance is modeled using a gating mechanism consisting of two fully connected layers: (10). Among them, The attention weights for each channel; This represents the Sigmoid function, used to normalize channel weights to the (0,1) interval; and These are the learnable parameters for the fully connected layer; Represents the ReLU activation function; This is the channel description vector. The channel weights are then scaled channel-wise with the original features to obtain the attention-enhanced feature representation: (11). Among them, β c This represents the attention weight corresponding to the c-th channel; This represents the feature vector of the c-th channel.

[0044] The introduction of the attention mechanism enables the model to adaptively emphasize frequency band features that are more critical for OLTC fault identification, and improve the effectiveness of feature representation discrimination while suppressing redundant or noisy features. Through wavelet-enhanced multi-scale feature encoder, the model maps the original OLTC vibration signal into a feature representation with more complete structural information and higher stability, providing a reliable feature foundation for subsequent metric learning-based fault diagnosis.

[0045] S3, Elasticity Measurement Modeling.

[0046] After obtaining the feature representation output by the wavelet-enhanced multi-scale feature encoder, an elasticity measurer (EMM) is further introduced to enhance the ability of the metric space to characterize class differences. Traditional prototype networks typically use a fixed Euclidean distance to measure the similarity between query set samples and class prototypes, assuming that each class has a consistent scale in the feature space. However, in OLTC fault diagnosis scenarios, the compactness of feature distributions often varies significantly among different fault types due to differences in operating conditions, signal non-stationarity, and noise interference. A fixed metric method cannot simultaneously account for all classes, easily leading to unreasonable discrimination boundaries between classes. Therefore, an adjustable elasticity factor is introduced in the metric stage of the prototype network to adaptively adjust the metric scale for different classes.

[0047] like Figure 2 As shown, the support set samples are mapped to the embedding space through a wavelet-enhanced multi-scale feature encoder to form prototype representations for each category. After obtaining feature representations for the query set samples through the wavelet-enhanced multi-scale feature encoder, elastic distances are calculated with all category prototypes, and probabilities are obtained through a normalized exponential function, Softmax. This process uses backpropagation via a loss function to jointly update the feature encoder and elasticity metric parameters.

[0048] Let the query set sample embedding after wavelet-enhanced multi-scale feature encoder be represented as follows: (12). Where d is the dimension of the feature space. The prototype vector of class c is calculated from the mean of the samples in the support set. (13). Among them, S c For the support set of class c; |S c | represents the sample size; E φ (·) represents the mapping function of the feature encoder; X j It is S c The j-th original input sample in the dataset. To provide a more adequate elastic space, an unnormalized Euclidean distance is defined. (14). Among them, z im This represents the value of the embedding vector of the i-th query set sample in the m-th dimension; m is the feature dimension, m=1,2…d; P cmIt is the component of the prototype vector of class c in the m-th dimension. An independent elasticity factor λ is introduced for each feature dimension m of each class c. cm >0, forming the elastic weight vector (15). Thus, the query set sample z is obtained. i With the prototype p of class c c elastic distance (16). Wherein, the weight λ cm Adaptive adjustment is made based on the importance of feature dimensions: dimensions with higher information content have increased weights, i.e., λ. cm >1; the weights of dimensions with low information content are reduced, i.e., λ cm <1. Different categories can independently depend on different feature subsets. The probability P that a sample in the query set belongs to the k-th class. k P is calculated using the Softmax function. k = (17). Among them, For the query set sample; for Elastic distance between the prototype and category k N is the number of current task categories; d ij * It is the query set sample z i With the j-th prototype p j The elastic distance between them is calculated using formula (16); exp is the natural exponential function e x ;d k * It is the elastic distance of the k-th prototype; P k It is a query set sample The probability of belonging to the k-th class. The elastic classification loss of the model is defined as: (18). Among them, Q k Let be the query set of the k-th class, and Ф be all learnable parameters of the feature encoder.

[0049] To prevent the elasticity factor from deviating excessively from its unit value and to promote training stability, a regularization term is introduced: (19).

[0050] S4. Fault type identification.

[0051] The final training objective function is: (20). Where α is the regularization coefficient, and α takes a value of 0.1; Ф and all categories of elasticity factors. Joint optimization is achieved through stochastic gradient descent. By employing a class-specific elasticity factor, the proposed elasticity measure enables adaptation to the discrimination pattern of each sample class, thereby improving the model's generalization ability in OLTC fault identification tasks with small sample sizes and under different operating conditions. This is the elastic classification loss function of the elasticity measure EMM, used to measure the difference between the model's prediction results and the true labels; It is the regularization term of the elasticity factor.

[0052] This invention uses raw OLTC mechanical vibration signals as input and sequentially completes learnable wavelet enhancement, multi-scale feature extraction, elasticity metric modeling, and fault category discrimination, achieving end-to-end modeling from raw vibration signals to diagnostic results. Each module collaborates within the overall process: wavelet enhancement improves feature separability, multi-scale feature extraction effectively characterizes the local and global variations of the vibration signal, and the elasticity metric mechanism adaptively adjusts feature distances under small sample conditions, thereby enhancing the model's classification and discrimination capabilities. In summary, addressing the issues of limited samples and insufficient feature discriminability in OLTC fault diagnosis, a prototype learning network composed of wavelet-enhanced multi-scale feature encoding and an elasticity metric mechanism is constructed. Compared to fixed metrics or simple similarity measurement methods, WEMNet can more flexibly characterize differences between samples in the feature space, providing a more discriminative distance metric basis for small sample classification, and offering an efficient and well-generalized solution for small-sample mechanical fault diagnosis in OLTCs.

Claims

1. A method for small-sample fault diagnosis of on-load tap changers based on wavelet elasticity metric networks, characterized in that, Includes the following steps: S1, Data Acquisition Collect mechanical vibration signal x from on-load tap changer i The dataset is divided into a training set and a test set, and then the training set and test set are further divided into a support set and a query set, respectively. S2. Construct a wavelet-enhanced multi-scale feature encoder S2.1, Regarding the vibration signal x i Perform a learnable wavelet transform, and then apply the vibration signal x i Mapped to a multi-scale time-frequency representation space; S2.2 A lightweight multi-scale convolution is constructed to extract features from the wavelet-enhanced vibration signal, obtaining the features of the vibration signal at different time scales. These features are then concatenated to obtain the fused feature F. ms ; S2.3 Introduction of Channel Attention Mechanism Assume fusion features Feature compression is performed on each channel using global average pooling; S3, Elasticity Measurement Modeling An elasticity measure (EMM) is introduced to adaptively adjust different categories of measurement scales; S4. Fault Category Determination Through the objective function (1) Determine the fault type; among which, Let be the elasticity classification loss function of the elasticity measure EMM; α is the regularization coefficient. It is the regularization term of the elasticity factor; In step S2.2, the multi-scale convolution content is as follows: S2.2.1 Let the input features be processed by learnable wavelet as follows: (2), C represents the number of channels; the multi-scale convolution module includes several parallel one-dimensional convolution branches. For the s-th convolution branch, s=1,2…S, the one-dimensional convolution operation is: (3); among which, k represents the kernel parameters for the s-th branch. s C is the kernel length corresponding to this branch; s b is the number of output channels for this branch; (s) This represents the bias term; * indicates a one-dimensional convolution operation. Let S represent the feature representation output by the s-th convolutional branch; after the convolution operation, batch normalization S2.2.2 is introduced to standardize the features: (4); (5); among which, These are standardized features; Representation of features The mean; Representation of features The variance; A smoothing constant is used to prevent numerical instability; It is the final feature after translation and scaling; A learnable scale; The bias parameters are learnable; after step S2.2.2, execute S2.2.3: concatenate the output features of all branches along the channel dimension to obtain the fused features. (6); where Concat(⋅) represents the concatenation operation along the channel dimension; In step S4, the elastic classification loss function (7); where Q k Let be the query set of the k-th class; Ф be all learnable parameters of the feature encoder; For query set samples The elastic distance between the prototype of category k; N is the number of current task categories; exp is the natural exponential function e x ;d ij * It is the query set sample z i With the j-th prototype p j The elastic distance between them; The calculation formula is: (8); where λ cm Let λ be the elasticity factor, and λ be the elasticity factor. cm >0; For the elastic weight vector, .

2. The method for small-sample fault diagnosis of on-load tap changers based on wavelet elasticity metric networks according to claim 1, characterized in that, The specific content of step S2.1 is as follows: Let the input vibration signal be... Where L is the signal length, the output of the j-th wavelet neuron is expressed as: (9); among which, The wavelet basis functions are Morlet wavelets. (10); w ji b is the weight matrix of the wavelet neuron; j a represents the wavelet translation coefficient; j is the wavelet scaling factor; M is the number of wavelet neurons.

3. The method for small-sample fault diagnosis of on-load tap changers based on wavelet elasticity metric networks according to claim 1, characterized in that, In step S2.3, the feature compression formula for the channel is as follows: (11); among which, Let m represent the feature value of the c-th channel at position i, where m is the feature dimension, m=1,2…d; channel importance is modeled using a gating mechanism: (12); among which, The attention weights for each channel; Represents the Sigmoid function; and These are the learnable parameters for the fully connected layer; Represents the ReLU activation function; The channel description vector is used; the channel weights are scaled channel-wise with respect to the original features to obtain the attention-enhanced feature representation. (13); where β c This represents the attention weight corresponding to the c-th channel; This represents the feature vector of the c-th channel.

4. The method for small-sample fault diagnosis of on-load tap changers based on wavelet elasticity metric networks according to claim 1, characterized in that, The calculation formula is: (14); Where, d k * E is the elastic distance of the k-th prototype; φ (·) represents the mapping function of the feature encoder; P k It is a query set sample The probability of belonging to the k-th class.

5. The method for small-sample fault diagnosis of on-load tap changers based on wavelet elasticity metric networks according to claim 1, characterized in that, P k Calculated using the Softmax function: P k = (15); Where, d ij * Calculate using the following formula: d ij * = (16); where z im P represents the value of the embedding vector of the i-th query set sample in the m-th dimension; cm It is the component of the c-th class prototype vector in the m-th dimension; P c It is the prototype vector of class c. S c For the support set of class c; |S c | represents the number of samples; X j It is S c The j-th original input sample in the dataset.

Images