A fault diagnosis method for on-load tap changers based on geometric feature mapping and enhanced EEGNet

Abstract

The application discloses a kind of based on geometry feature mapping and enhanced EEGNet on-load tap-changer fault diagnosis method, it is related to on-load tap-changer fault diagnosis technical field, for solving industrial field vibration signal non-stationary, weak fault is easy to cover and depth model decision is not transparent operation and maintenance field problem.The application first carries out Hilbert-Savgol filter noise reduction and phase space plotting signal preprocessing operation, then the data after preprocessing is carried out nesting convex hull mask and distance field calculation geometry feature coding operation, finally, data processing is carried out using enhanced attention EEGNet network.The application eliminates the black box attribute of deep learning model by constructing the strong coupling explanation logic of thermal distribution feature-geometric topological motion-physical operation mechanism, provides intuitive, traceable diagnostic evidence for operation and maintenance personnel.

Description

Technical Field

[0001] This invention relates to the field of on-load tap changer fault diagnosis technology, specifically to an on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet. Background Technology

[0002] Non-invasive diagnostic techniques based on vibration signal analysis have become a research hotspot due to their advantages of not interfering with normal switch operation and providing rich information. Existing technologies utilize Fast Fourier Transform, Wavelet Transform, or Hilbert-Huang Transform to extract the time-frequency features of signals and combine them with deep learning models for fault classification. However, in practical industrial applications, the field of on-load tap changer faults still faces the following shortcomings.

[0003] First, there is insufficient mining of spatial topological information in the evolution of signal features. Existing diagnostic methods mostly focus on one-dimensional time-domain or frequency-domain features, neglecting the nonlinear dynamic evolution characteristics of vibration signals in phase space during on-load tap changer operation. When the signal is damaged or a fault shift occurs, the geometric shape of its phase trajectory will undergo slight distortion, and traditional feature representation methods are unable to quantify such high-dimensional geometric topological changes.

[0004] Secondly, the robustness of features under strong background noise is poor. The instantaneous operation of an on-load tap changer is accompanied by strong electromagnetic interference and mechanical shock, and weak early fault features are easily masked by non-stationary noise. Traditional preprocessing methods often cause secondary loss of feature information while reducing noise, resulting in limited generalization ability of the model under complex operating conditions and difficulty in accurately capturing weak faults.

[0005] Third, the black-box nature of deep learning models limits their credibility in industry. Although convolutional neural network models perform well in terms of recognition accuracy, their feature extraction and decision-making processes lack physical interpretability. In the power production field, where safety requirements are extremely high, if the physical basis for the model's fault determination cannot be intuitively explained, such as the specific time period for locating abnormal signals, the credibility of the diagnostic results will be reduced, making it difficult to effectively guide maintenance personnel to carry out accurate repairs.

[0006] To address the aforementioned challenges, there is an urgent need to explore an intelligent fault diagnosis solution for on-load tap changers that can deeply integrate signal geometric topology features, possess strong noise immunity, and has a transparent and interpretable diagnostic process, in order to meet the actual needs of power equipment operation and maintenance. Summary of the Invention

[0007] The purpose of this invention is to provide a fault diagnosis method for on-load tap changers based on geometric feature mapping and enhanced EEGNet, which can solve the problems of non-stationary vibration signals, easy masking of weak faults, and opaque decision-making of deep models in industrial sites.

[0008] The technical solution adopted by the present invention to solve its technical problem is: an on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet, comprising the following steps.

[0009] S1. On-load tap changer vibration signal acquisition and preprocessing.

[0010] S1.1 Acquires vibration signals from on-load tap changers .

[0011] S1.2 for vibration signals Perform Hilbert transform to extract the instantaneous envelope that reflects the energy evolution of the signal. ,against An SG filter is used for denoising to obtain a smoothed signal sequence. .

[0012] S1.3 Perform phase space mapping.

[0013] S2 constructs a representation based on a computational geometry-based dual-channel topological map.

[0014] S2.1 uses the scan-line filling algorithm to generate nested convex hull masks. .

[0015] S2.2 Topological distance field modeling based on Euclidean distance transformation (EDT).

[0016] S2.3 Dual-channel feature fusion and tensor encapsulation.

[0017] By concatenating nested convex hull masks and topological distance fields along the depth dimension, a dual-channel input feature tensor is constructed. .

[0018] S3. Fault diagnosis of on-load tap changers of power transformers based on MS-EEGNet-AM network.

[0019] Fault diagnosis is performed using an improved EEGNet model. The improved EEGNet uses depthwise separable convolution to independently model the mask and range field channels, and uses an attention mechanism to achieve optimal stimulus allocation among feature channels.

[0020] Further, step S1.2 includes: S1.2.1 (1); where, The sequence is the Hilbert sequence obtained after performing a Hilbert transform on the original vibration signal; j is the imaginary unit. (2); t is the current time; H() is the Hilbert transform; It is an integral variable; It is the vibration signal at time. The specific value; (3); S1.2.2 For envelope signals medium length is The sliding window fitting polynomial is first based on (4) Calculate the coefficients of the SG filter. Where A(i+m) is the instantaneous envelope data; m is the data index; n is... The nth order polynomial; then, the smoothed value of the final output is calculated using the filter coefficients. The smoothed signal sequence is obtained. .

[0021] Further, step S1.3 includes: constructing a phase point set. (5); among which, for The rate of change of amplitude; define the mapping function. (6); (7); P and All represent components in the phase point set S; Res is the preset geometric resolution. To map the horizontal components in the phase point set S to the discrete pixel coordinate axes; To map the ordinate components of the phase point set S to the discrete pixel coordinate axes; min(P) is the minimum value of P; max(P) is the maximum value of P; min(ΔP) is... The minimum value; max(ΔP) is The maximum value.

[0022] Furthermore, in step S2.1, (8); among which, The indicator function is Polygon(V); Polygon(V) represents a closed polygonal region bounded by the vertex sequence V.

[0023] Furthermore, in step S2.2, for Each foreground pixel in Find its closest point to the background area. The Euclidean distance, its transformation operator (9); among which, These are the coordinates of the pixels within the Polygon (V); It is the distance to the background region outside of Polygon(V). The coordinates of the nearest background point to the pixel; normalize the distance matrix by applying the global maximum value to obtain the normalized topological distance field. (10); among which, The internal coordinates of Polygon (V) are The minimum distance between the point and the background point outside the Polygon (V); Points inside a Polygon (V) The maximum value.

[0024] Furthermore, in step S2.3, (11); among which, This represents a cascade operator in the depth dimension.

[0025] Furthermore, in step S3, the parallel multi-scale convolutional layers of the improved EEGNet use convolutions of different sizes. Simultaneously extract features, the output features of the convolutional layer (12); where ELU is the activation function; For batch normalization layer; It is a convolution kernel.

[0026] Furthermore, in step S3, the improved EEGNet introduces two learnable parameters: the scale factor. and offset factor The normalized data is reconstructed using affine transformation: (13); among which, This is the final output of the batch normalization layer; This is the original sample channel after normalization. The data.

[0027] Furthermore, in step S3, the improved EEGNet introduces deep convolutional layers to handle multi-scale input feature tensors. The forward propagation process of depthwise convolution is as follows: (14); among which, The index represents the feature channel; r and u are the spatial indices of the feature map; is the spatial size of the convolution kernel; f and j are the kernel indices.

[0028] Furthermore, in step S3, the improved EEGNet introduces pointwise convolutional layers. The pointwise convolutional forward propagation process is as follows: (15); among which, The output is a pointwise convolution; represents the channel index of the output feature map; ELU is the exponential linear unit activation function; For corresponding Channel bias; BN is batch normalization. This is the weight matrix.

[0029] Furthermore, in step S3, the improved EEGNet introduces an adaptive channel attention mechanism for the first channel. One channel, (16); among which, This represents the output feature map after pointwise convolution. Each channel is in coordinates The pixel value at that location; the first Dynamic incentive weights for each channel (17); Sigmoid is the activation function; Weights for the dimensionality reduction layer; For activation functions; Weights for the dimension-upgrading layer; The first channel description vector Each component; finally, the weights are multiplied channel by channel. Apply to the original feature channel Above, the feature reconstruction is completed; (18); The feature results of the reconstruction; The first point of the output result of the pointwise convolution Feature maps of each channel; For the first Dynamic incentive weights for each channel.

[0030] Furthermore, in step S3, global average pooling is performed on the reconstructed features. (19); among which, The first term after global average pooling The weight coefficients for each channel are determined; subsequently, the feature vector is input into the fully connected layer. (20); among which, These are the weights of the fully connected layer; For the output of the fully connected layer for the first Bias in the results of fault classes; using the Softmax function to map high-dimensional features to the fault probability space: (21); among which, The sample was ultimately determined to be The probability of a type of failure; For the classification layer targeting the first The logical value output by the fault type; Index for fault categories; This represents the total number of fault types.

[0031] Furthermore, in step S3, the cross-entropy loss function is used as the objective optimization function: (22); among which, The true label after one-hot encoding; The w-th sample is judged as the w-th sample. The probability of a fault class; Total is the total number of samples used in this training.

[0032] Furthermore, in step S3, a cosine annealing strategy with hot restart is introduced, which participates in all learnable parameters within the network architecture throughout the process. (23); among which, The learning rate used in the current round; This is the lower bound of the learning rate in the i-th cycle; This is the upper bound of the learning rate in the i-th cycle; This refers to the number of training rounds already performed within the current cycle; Let be the total length of the i-th cycle.

[0033] Furthermore, in step S3, the attribution analysis method SHAP is introduced for a given input map feature. Its prediction function The additive interpretation model is as follows: (24); among which, This represents the total number of pixels. The baseline predicted mean of the model across all samples; For the first Shapley values ​​of 1 pixel; Characterizing the first in the original map The activation state of each pixel feature; P SHAP Determine the probability for a specific fault category; (25); among which, The sample obtained through random sampling does not include the first... 1 pixel A subset of pixels; For the current subset The number of features; Indicates when features The predicted probability of the model for a specific fault category when it is in an active or inactive state; ! represents factorial.

[0034] The beneficial effects of this invention are: by constructing a strongly coupled explanatory logic of thermal distribution characteristics, geometric topological anomalies, and physical operating mechanisms, this invention eliminates the black-box nature of deep learning models, provides operation and maintenance personnel with intuitive and traceable diagnostic evidence, and improves the industrial credibility of evaluating the on-load tap changer status of power transformers based on large models. Attached Figure Description

[0035] Figure 1 This is a flowchart of the on-load tap changer fault diagnosis method of the present invention;

[0036] Figure 2 This is a flowchart of the SHAP attribution analysis process of the present invention. Detailed Implementation

[0037] This invention first addresses the problem of insufficient mining of topological information in the signal feature evolution space. It utilizes phase space reconstruction to upgrade the dimensionality of one-dimensional signals and constructs a dual-channel geometric feature encoding mechanism to quantify trajectory distortion from both the outer envelope and internal density dimensions. Secondly, to address the issue of poor feature robustness under strong background noise, an enhanced EEGNet integrating multi-scale parallel convolution and adaptive attention mechanisms is designed. It utilizes convolution operators with different receptive fields to decouple and extract heterogeneous features, achieving accurate fault capture under complex interference. Finally, to address the challenge of insufficient industrial trust caused by the black-box nature of deep learning, a game theory-based SHAP attribution analysis algorithm is introduced. This algorithm maps model decisions to physically oriented pixel-level heatmaps, making the diagnostic process transparent and effectively guiding operational practices. Figure 1 As shown, the on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet of the present invention first performs signal preprocessing operations such as Hilbert-Savgol filtering for noise reduction and phase space plotting, then performs geometric feature encoding operations such as nested convex hull masking and distance field calculation on the preprocessed data, and finally uses an enhanced attention EEGNet network for data processing. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet of the present invention includes the following steps.

[0038] S1 On-load tap changer vibration signal preprocessing.

[0039] S1.1 Acquisition of the original vibration signal of the on-load tap changer.

[0040] Using a high-sensitivity piezoelectric accelerometer mounted on the outer wall of the transformer tank, the transient vibration response, i.e., the vibration signal, during the on-load tap changer switching process is synchronously acquired. By setting specific mechanical disturbance parameters, four typical operating states of on-load tap changers were simulated on an experimental platform: normal operation, gear mechanism jamming, contact failure, and insufficient preload of fastening components. A raw time-series database containing multiple independent samples was constructed to ensure that the training data covers the dynamic evolution characteristics of the entire operating cycle of the on-load tap changer, providing underlying data support for subsequent feature extraction and model training.

[0041] S1.2 Hilbert-Savgol filter.

[0042] S1.2.1 The acquired raw one-dimensional vibration signal Perform Hilbert transform to construct analytic signals. : (1); where, The sequence is the Hilbert sequence obtained after performing a Hilbert transform on the original vibration signal; j is the imaginary unit. ; (2). Where t is the current time; H() is the Hilbert transform; It is an integral variable; It is the original one-dimensional vibration signal at time t. The specific numerical value is then used to extract the instantaneous envelope reflecting the evolution of signal energy. : (3).

[0043] S1.2.2 For the extracted envelope signal An SG filter is used for noise reduction.

[0044] The principle of the SG filter is to utilize within a sliding window... A polynomial of order n is used to perform local least-squares fitting on the sampling points. For a signal with a length of n... The sliding window fitting polynomial, firstly, according to (4) Calculate the coefficients of the SG filter. Where A(i+m) is the instantaneous envelope data; m is the data index; n is... The nth order polynomial. Then, the smoothed value of the final output is calculated using the obtained filter coefficients. The smoothed signal sequence is obtained. Compared to simple averaging filtering, polynomial constraints effectively preserve the second-order and higher-order derivative features of the envelope, such as peaks and steep edges, while smoothing the signal by filtering out high-frequency glitches.

[0045] S1.3 smooths the signal sequence Mapped to the state phase space.

[0046] To capture the evolution trend of the signal, the current amplitude and its first-order difference are selected as spatial basis vectors to construct a phase point set. : (5). Among them, Characterization system in The smoothed signal amplitude at any given time; for The rate of change of amplitude. To transform the phase point set into a geometric map with fixed resolution, a linear mapping is performed on the coordinates. Define the mapping function. (6); (7). Among them, P and All represent components in the phase point set S; Res is the preset geometric resolution. To map the horizontal components of the phase point set S to discrete pixel coordinate axes, To map the ordinate components of the phase point set S to the discrete pixel coordinate axes, a normalized point set is obtained; min(P) is the minimum value of P; max(P) is the maximum value of P; min(ΔP) is... The minimum value; max(ΔP) is The maximum value.

[0047] The deep coupling of Hilbert-Savgol filtering and phase space mapping effectively suppresses interference from complex electromagnetic environments and random mechanical noise in industrial settings. It also extracts high-order dynamic features hidden in non-stationary envelopes and transforms one-dimensional time series into two-dimensional dynamic trajectories that reflect the evolution of system states, thus providing high-quality feature signal components for subsequent fault identification.

[0048] S2. Construction of representation based on computational geometry dual-channel topological map.

[0049] S2.1 Nested convex hull mask generation based on Graham scan algorithm.

[0050] The convex hull is the smallest convex polygon that contains all phase points. The convex hull is used as the macroscopic envelope of the signal energy distribution.

[0051] Vertex search logic: The system first identifies extreme points, such as the point with the smallest ordinate, and uses this as a benchmark to sort other points. The Graham scan algorithm is used to traverse the point set, and by determining the rotation direction of the vector product, all concave points located inside are eliminated, retaining only the vertex sequence V that constitutes the outermost boundary.

[0052] Mask pixel mapping: Utilizing a scanline fill algorithm, at a resolution of... Within the zero matrix, the vertex sequence All pixels inside the enclosed region are assigned a weight of 1, and those outside the region are assigned a weight of 0, thus generating a nested convex hull mask. . (8). Among them, The function is an indicator function. If the condition in parentheses is true, the pixel is assigned a value of 1; otherwise, it is assigned a value of 0. Polygon(V) represents a closed polygonal region enclosed by the vertex sequence V.

[0053] S2.2 Topological distance field modeling based on Euclidean distance transformation (EDT).

[0054] A simple convex hull boundary only reflects the extreme range of the signal and cannot characterize the distribution density of the trajectory inside. This invention introduces a distance transformation to convert discrete distribution properties into a continuous distance field.

[0055] The distance transformation algorithm is implemented through two scan mappings to obtain the distance weight from each pixel to the boundary. For Each foreground pixel in Find its closest point to the background area. The Euclidean distance. Its transformation operator. (9). Among them, These are the coordinates of the pixels within the Polygon (V); It is the distance to the background region outside of Polygon(V). The coordinates of the nearest background point to the pixel.

[0056] Range field normalization: To eliminate the influence of absolute pixel size, the distance matrix is ​​normalized to its global maximum value to obtain a range field that reflects the internal topological depth. (10). Among them, The internal coordinates of Polygon (V) are The minimum distance between the point and the background point outside the Polygon (V); Points inside a Polygon (V) The maximum value; It is the normalized topological distance field.

[0057] S2.3 Dual-channel feature fusion and tensor encapsulation.

[0058] The two independent physical channels generated, namely the nested convex hull mask channel and the topological distance field channel, are concatenated in the depth dimension to construct a dual-channel input feature tensor. : (11). Among them, This represents a cascade operator in the depth dimension, which combines two objects of equal size. The matrices are stacked to form a three-dimensional data block with two channels. Then, the input feature tensor... It is then fed into a subsequent multi-scale EEG diagnostic network.

[0059] Step S2.3 uses a dual-channel encoding mechanism of nested convex hull mask and flexible edge distance field to map the one-dimensional signal into a two-dimensional topological map. This not only utilizes geometric boundaries to lock in the macroscopic evolution trend of signal energy, but also quantifies the density fluctuations and topological instability within the trajectory through the distance field gradient. This cross-dimensional feature representation method effectively overcomes the limitations of traditional one-dimensional time-frequency analysis in mining nonlinear dynamic information when processing non-stationary signals. It can capture geometric distortions hidden under strong background noise, thereby enhancing the model's spatial representation ability and discrimination accuracy for early weak faults in on-load tap changer systems.

[0060] Through the above filtering process, phase space mapping, and dual-channel feature tensor encapsulation, the original vibration signal has become a multidimensional tensor.

[0061] S3. Fault diagnosis of on-load tap changers of power transformers based on MS-EEGNet-AM network.

[0062] EEGNet is a lightweight convolutional neural network specifically designed for processing electroencephalogram (EEG) signals. Its core advantage lies in its decoupling extraction of spatiotemporal features through depthwise separable convolutions, enabling it to capture complex nonlinear dynamic features with a minimal number of parameters. Addressing the characteristics of on-load tap changer vibration signals—strong non-stationarity and susceptibility to industrial background noise—EEGNet leverages its expertise in low signal-to-noise ratio feature extraction accumulated in EEG signal processing. By utilizing depthwise separable convolutions to decouple spatiotemporal features, it can accurately capture weak fault pulses in transient impacts while effectively suppressing overfitting risks in small sample environments through its lightweight design. Therefore, this invention employs the core architecture of EEGNet for neural network fault diagnosis.

[0063] The first layer of EEGNet is a two-dimensional temporal convolutional layer, which uses a long, narrow convolutional kernel to slide along the temporal dimension of a single channel to extract frequency domain features of the signal, similar to bandpass filtering, capturing transient fluctuation patterns in vibration signals or EEG signals. The second layer is a deep convolutional layer, which does not perform cross-channel feature fusion but instead assigns an independent convolutional kernel to each input channel, learning the unique spatial weights of each physical channel when processing multi-source signals. The third layer is a separable convolutional layer, utilizing... The pointwise convolution nonlinearly combines the previously extracted heterogeneous features. This step achieves cross-channel information integration, ultimately mapping the features in the boundary and gradient dimensions to a high-dimensional fault feature space.

[0064] To address the issue of traditional networks' limited representation of non-stationary signals, an improved EEGNet model with multi-scale feature decoupling and adaptive saliency enhancement is proposed. Its core logic lies in using depthwise separable convolutions to independently model the mask and range field channels, and employing an attention mechanism to achieve optimal excitation allocation among feature channels. This hierarchical evolution strategy aims to explicitly strengthen the geometric topological features that critically contribute to the state evolution of on-load tap changers.

[0065] Parallel multi-scale operators are employed to capture the transient and steady-state distortions of on-load tap changer vibration signals. To address the scale diversity exhibited by the on-load tap changer vibration signals in phase space, such as overall centroid shift and local trajectory spikes, parallel multi-scale convolutional layers are designed. These parallel multi-scale convolutional layers utilize convolutional kernels of different sizes. Simultaneous feature extraction: (12). Among them, The output features of the multi-kernel convolutional layer are represented by ELU; ELU is the activation function. For batch normalization layer; For dual-channel input feature tensors; It's a convolution kernel. Three different sizes of convolution kernels are used, including small-sized convolution kernels. Capable of capturing microscopic geometric features reflecting local distortions in phase space trajectories caused by early, minor faults induced by mechanical shock; medium-sized convolution kernels It is responsible for identifying local curvature or regional density changes in phase trajectories, and recognizing moderate changes in the envelope signal; large-size convolutional kernels. It focuses on recognizing global macroscopic features such as the overall center of gravity shift or large-scale geometric contour deformation of the phase space map with a larger receptive field.

[0066] After each convolutional layer outputs, the distribution of feature maps often suffers from internal covariate shift as network depth increases. Batch normalization standardizes each mini-batch of data, mapping the feature distribution to a standard normal distribution region with a mean of 0 and a variance of 1. The specific calculation process is as follows: Let the current mini-batch samples be... ,in, This represents the number of samples stacked during training. For each channel of the feature map, first calculate the mean value of pixels in all identical channels of that batch of sample data. With variance Using the calculated mean With variance For each sample, the data in the same channel after multi-scale convolution output is normalized, and a very small positive number is introduced. This is to prevent the denominator from being zero and to ensure the stability of numerical calculations. (13). Among them, This is the original sample channel before normalization. Data; This is the original sample channel after normalization. The data.

[0067] To ensure that the normalization operation does not destroy the feature representation capabilities already learned by the improved EEGNet model, i.e., to prevent all activation values ​​from being restricted to the linear region of the non-linear activation function, batch normalization (BN) introduces two learnable parameters: the scale factor. and offset factor The normalized data is reconstructed using affine transformation: (14); among which, This is the final output of the batch normalization layer. Scale factor. and offset factor By optimizing the backpropagation algorithm, the network can adaptively adjust the amplitude and center of the feature distribution, thereby preserving the key nonlinear features that contribute to the identification of on-load tap changer faults.

[0068] The traditional ReLU function The output is always zero, which causes the mean output of neurons to always be greater than zero. This non-zero mean phenomenon creates a bias shift between neural network layers, forcing the weights of the next layer to be constantly adjusted during training to offset this shift, thus slowing down the convergence speed of backpropagation. In contrast, the Exponential Linear Unit (ELU) introduces parameterized... The exponential term of the algorithm exhibits soft saturation in the negative region, causing the output mean to spontaneously tend towards zero, producing a modulation effect similar to batch normalization. This zero-mean directionality reduces the overall shift in neuron activation, making the gradient descent path in the parameter space smoother and shortening the training convergence cycle of the improved EEGNet model on complex OLTC fault samples.

[0069] In the processing of the improved EEGNet with multi-scale feature decoupling and attention mechanisms, a deep convolutional layer based on the EEGNet architecture is introduced to address the limitation of traditional convolutional operators in handling feature mixing when dealing with multi-source heterogeneous features. The core logic of the deep convolutional layer is that instead of treating the input feature map as a whole tensor for cross-dimensional weighted summation, it independently allocates a dedicated set of spatial convolutional kernels for each physical channel in the input feature map, namely the mask channel component and the distance field channel component in the multi-scale feature tensor. .

[0070] For the multi-scale input feature tensor composed of convex hull mask and Euclidean distance transform (EDT) The forward propagation process of its depthwise convolution can be described as a spatial convolution operation for each independent channel: (15). Among them, The index represents the feature channel; r and u are the spatial indices of the feature map; is the spatial size of the convolution kernel; f and j are the kernel indices. Compared to standard convolution, depthwise convolution is mathematically characterized by assigning an independent two-dimensional convolution kernel to each channel, with each kernel responsible only for the convolution calculation of its assigned channel.

[0071] After achieving spatial decoupling of heterogeneous features in deep convolutional layers, in order to establish the intrinsic nonlinear relationship between mask boundary information and range field gradient information, a method based on... Pointwise convolutional layers with kernel-sized features facilitate the transition from local spatial perception to global logical classification. Pointwise convolution does not alter the spatial dimension; instead, it extracts features through a kernel applied to depthwise convolution. Each independent channel undergoes linear combination and nonlinear activation. Its forward propagation process is as follows: (16). Among them, It is a pointwise convolution output, which linearly combines the decoupled features of the previous segment according to the weights; The channel index is used for the output feature map; ELU is the exponential linear unit activation function, which enhances the expressive power of feature combinations through the exponential properties of its negative region; For corresponding Channel bias; BN is batch normalization. This is the weight matrix, responsible for weighting the number of input channels. Mapped to the target feature space dimension.

[0072] To identify key representations highly correlated with the evolution of on-load tap changer faults from multi-scale fused features, an adaptive channel attention mechanism is introduced. This mechanism achieves secondary weight calculation of the feature map along the channel dimension by explicitly modeling the dependencies between channels. Its core logic consists of three stages: compression, activation, and feature reconstruction.

[0073] Traditional convolutional operations can only utilize spatial information within a local receptive field, making it difficult to capture the global distribution characteristics between channels. The compression phase in the adaptive channel attention mechanism uses global average pooling to pool each channel... The feature channels of the size are compressed into a single channel description vector representing the global receptive field. For the first One channel, (17). Among them, This represents the output feature map after pointwise convolution. Each channel is in coordinates The model extracts the geometric features of the mask boundary and the global energy distribution of the distance field gradient, compressing the spatial dimension into a channel description vector. .

[0074] The attention mechanism module employs a bottleneck structure consisting of two fully connected layers to capture the nonlinear dependencies between channels. This process aims to learn and generate dynamic activation weights for each channel. Dynamic incentive weights for each channel (18). Wherein, Sigmoid is the activation function that maps the output to... The interval represents the relative importance of each channel; Weights for the dimensionality reduction layer; This is an activation function used to introduce non-linear interaction elements; To increase the weights of the dimension-upgrading layer, the features are restored to their original dimensions; The first channel description vector Each component.

[0075] Finally, the learned activation weights are multiplied channel by channel. Apply to the original feature channel Then, the feature reconstruction is completed. (19). Among them, The feature results of the reconstruction; The first point of the output result of the pointwise convolution Feature maps of each channel; For the first Dynamic incentive weights for each channel.

[0076] The attention mechanism module ends by performing global average pooling on the reconstructed features. (20) Replaces the traditional flattening operation to significantly reduce parameter redundancy and enhance robustness to feature space distortion. Among them, The first term after global average pooling The weight coefficients for each channel are then assigned. Subsequently, the feature vectors are input into the fully connected layer. (21). Among them, These are the weights of the fully connected layer; For the output of the fully connected layer for the first Bias in the outcome of fault classes. The Softmax function is used to map high-dimensional features to the fault probability space: (22). Among them, The sample was ultimately determined to be The probability of a type of failure; For the classification layer targeting the first The logical value output by the fault type; Index for fault categories; This represents the total number of fault types.

[0077] To characterize the discrepancy between the diagnostic results and the true labels, the cross-entropy loss function is used as the objective optimization function: (23). Among them, The true label after one-hot encoding; The w-th sample is judged as the w-th sample. The probability of a fault class; Total is the total number of samples used in this training. This cross-entropy loss function penalizes deviations from the probability distribution, causing the model to continuously compress the intra-class distance in the feature space during training.

[0078] In small-sample, non-stationary data scenarios for OLTC fault diagnosis, a fixed learning rate can easily cause the model to get stuck in local minima or stagnate at saddle points. To address this, a cosine annealing strategy with hot restart is introduced for all learning parameters within the network architecture, participating throughout the entire learning process. (24). Among them, The learning rate used in the current round; This is the lower bound of the learning rate in the i-th cycle; This is the upper bound of the learning rate in the i-th cycle; This refers to the number of training rounds already performed within the current cycle; Let be the total number of rounds in the i-th cycle. The core of the cosine annealing strategy lies in the periodic activation of global search capability; that is, at the end of each decay cycle, the learning rate will momentarily jump back to its maximum value. This hot restart mechanism generates perturbations that allow the model to escape potential local traps or flat boundaries in the non-convex loss function, enabling it to re-search for optimal feature combinations within a larger parameter space. Simultaneously, with the current cycle round... approaching the target value The learning rate decays smoothly following a cosine curve. This local convergence characteristic allows the model to fine-tune weights with minimal steps at the bottom of the loss space in the later stages of training, ensuring that the parameters are positioned at the optimal point with strong generalization ability. Furthermore, the continuous smoothness of the cosine descent trajectory alleviates the convergence oscillations caused by traditional step-down descent, improving the convergence robustness of the MS-EEGNet-AM architecture against complex electromagnetic noise and mechanical background interference in industrial environments. This allows the model to maintain stable training efficiency even in non-stationary random signal environments.

[0079] To address the lack of interpretability in deep learning models for power equipment condition assessment, an attribution analysis method based on game theory, SHAP, is introduced. SHAP establishes a logical connection between map features and physical operating conditions by quantifying the contribution of each input pixel feature to the diagnostic decision of MS-EEGNet-AM.

[0080] Step S3 achieves deep spatial decoupling of heterogeneous features. By combining an attention mechanism to saliency-enhancing key topological features with a cosine annealing strategy that includes hot restart, the model can periodically escape local minima during training, significantly enhancing its global optimal parameter retrieval capability and convergence stability in small sample sizes and strong background noise environments. In summary, this architecture, through the synergistic effect of multi-scale decoupling and adaptive weight allocation, improves the model's sensitivity to transient and steady-state distortions in on-load tap changer vibration signals, effectively overcoming the limitation of traditional networks' singular representation capabilities under non-stationary random disturbances. Thus, it achieves accurate capture and robust identification of weak early faults while maintaining a low parameter count.

[0081] like Figure 2 As shown, the core logic of the attribution analysis method SHAP lies in treating the final prediction of the model as the combined benefit of all input features, and assigning a unique Shapley value to each feature based on its marginal contribution across all possible subset combinations. For a given input map feature... Its prediction function The additive interpretation model is defined as: (25). Among them, This represents the total number of pixels. The baseline predicted mean of the model across all samples; For the first The Shapley value of the nth pixel measures the... The contribution of each pixel to the current prediction result; Characterizing the first in the original map The activation state of each pixel feature; P SHAP Determine the probability for a specific fault category; (26). Among them, The sample obtained through random sampling does not include the first... 1 pixel A subset of pixels; For the current subset The number of features; Indicates when features The predicted probability of the model for a specific fault category when in an active or inactive state; ! represents factorial. Ensure that the probability of judgment for a specific fault category calculated by formula (25) is accurate. The probability of fault judgment is equal to that of the network processing output, achieving a balance between computational efficiency and attribution accuracy. Formula (26) considers feature x in different feature sets. The marginal increments ensured mathematical consistency in the attribution results. Then, a heatmap visualization was created for the Shapley values ​​calculated for each pixel.

[0082] When the highlighted positive contribution regions of the SHAP attribution heatmap are concentrated at the edge of the convex hull mask channel and in its normal offset direction, the judgment model identifies boundary expansion or geometric distortion of the phase space trajectory. At the physical logic level, this feature mapping corresponds to the motion trajectory obstruction caused by a surge in instantaneous resistance during gear meshing. This attribution feature is directly related to the nonlinear evolution of stroke caused by gear tooth wear, lubrication failure, or foreign object embedding, thereby enabling the identification of jamming defects in gear mechanisms.

[0083] When the positive contribution core region of the SHAP attribution heatmap focuses on the geometric centroid of the distance field channel and the high dynamic gradient fluctuations, it indicates that the model has captured the structural instability of the phase space trajectory energy density distribution. From a physical evolution perspective, this attribution characteristic points to the abnormal fluctuations in arc energy or the vibration energy reconstruction caused by increased contact resistance during on-load tap changer switching. Since poor contact leads to a centroid shift and energy dispersion characteristic in phase space, this attribution result enables a quantitative diagnosis of contact failures.

[0084] When the SHAP attribution heatmap exhibits diffuse highlighting across channels, and the high-contribution region is distributed in a scattered pattern within the overall contour of the convex hull mask channel, it indicates that the model has identified high-frequency texture perturbations within the trajectory. At the physics level, insufficient preload can cause uncontrolled secondary vibrations or stray harmonics in the mechanical structure during operation, resulting in microscopic burrs or envelope thickening within the phase space trajectory. This attribution feature directly maps to the decrease in system stiffness caused by component loosening, thus enabling the identification of insufficient preload faults in fastening components.

[0085] In summary, this invention eliminates the black-box nature of deep learning models by constructing a strongly coupled explanatory logic of thermal distribution characteristics, geometric topological anomalies, and physical operating mechanisms. This provides maintenance personnel with intuitive and traceable diagnostic evidence and enhances the industrial credibility of evaluating the on-load tap changer status of power transformers based on large models.

Claims

1. A method for on-load tap changer fault diagnosis based on geometric feature mapping and enhanced EEGNet, characterized in that, Includes the following steps: S1. Vibration signal acquisition and preprocessing of on-load tap changer; S1.1 Acquires vibration signals from on-load tap changers ; S1.2 for vibration signals Perform Hilbert transform to extract the instantaneous envelope that reflects the energy evolution of the signal. ,against An SG filter is used for denoising to obtain a smoothed signal sequence. ; S1.3 Perform phase space mapping; S2 constructs a representation based on a dual-channel topological map using computational geometry; S2.1 uses the scan-line filling algorithm to generate nested convex hull masks. ; S2.2 Topological distance field modeling based on Euclidean distance transformation (EDT); S2.3 Dual-channel feature fusion and tensor encapsulation; By concatenating nested convex hull masks and topological distance fields along the depth dimension, a dual-channel input feature tensor is constructed. ; S3. Fault diagnosis of on-load tap changer of power transformer based on MS-EEGNet-AM network; Fault diagnosis is performed using an improved EEGNet model based on multi-scale feature extraction (MS). The improved EEGNet uses depthwise separable convolution to independently model the mask and range field channels, and uses the attention mechanism (AM) to achieve optimal activation allocation among feature channels.

2. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 1, characterized in that, Step S1.2 includes: S1.2.1 Constructing an analytic signal (1); where, The sequence is the Hilbert sequence obtained after performing a Hilbert transform on the original vibration signal; j is the imaginary unit. (2); t is the current time; H() is the Hilbert transform; It is an integral variable; It is the vibration signal at time. The specific value; (3); S1.2.2 For envelope signals medium length is The sliding window fitting polynomial is first based on (4) Calculate the coefficients of the SG filter. Where A(i+m) is the instantaneous envelope data; m is the data index; n is... The nth order polynomial in the nth order; then, the smoothed value of the final output is calculated using the filter coefficients. The smoothed signal sequence is obtained. .

3. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 2, characterized in that, Step S1.3 includes: constructing a phase point set. (5); among which, for The rate of change of amplitude; define the mapping function. (6); (7); P and All represent components in the phase point set S; Res is the preset geometric resolution. To map the horizontal components in the phase point set S to the discrete pixel coordinate axes; To map the ordinate components of the phase point set S to the discrete pixel coordinate axes; min(P) is the minimum value of P; max(P) is the maximum value of P; min(ΔP) is... The minimum value; max(ΔP) is The maximum value.

4. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 3, characterized in that, In step S2.1, (8); among which, The indicator function is Polygon(V); Polygon(V) represents a closed polygonal region bounded by the vertex sequence V.

5. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 4, characterized in that, In step S2.2, for Each foreground pixel in Find its closest point to the background area. The Euclidean distance, its transformation operator (9); among which, These are the coordinates of the pixels within the Polygon (V); The distance to the background region outside of Polygon(V) The coordinates of the nearest background point to the pixel; normalize the distance matrix by applying the global maximum value to obtain the normalized topological distance field. (10); among which, The internal coordinates of Polygon (V) are The minimum distance between the point and the background point outside the Polygon (V); Points inside a Polygon (V) The maximum value.

6. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 5, characterized in that, In step S2.3, (11); among which, This represents a cascade operator in the depth dimension.

7. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 6, characterized in that, In step S3, the parallel multi-scale convolutional layers of the improved EEGNet use convolutions of different sizes. Simultaneously extract features, the output features of the convolutional layer (12); where ELU is the activation function; For batch normalization layer; It is a convolution kernel.

8. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 7, characterized in that, In step S3, the improved EEGNet introduces two learnable parameters: the scale factor. and offset factor The normalized data is reconstructed using affine transformation: (13); among which, This is the final output of the batch normalization layer; This is the original sample channel after normalization. The data.

9. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 8, characterized in that, In step S3, the improved EEGNet introduces deep convolutional layers to handle multi-scale input feature tensors. The forward propagation process of depthwise convolution is as follows: (14); among which, The index represents the feature channel; r and u are the spatial indices of the feature map; is the spatial size of the convolution kernel; f and j are the kernel indices.

10. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 9, characterized in that, In step S3, the improved EEGNet introduces pointwise convolutional layers. The pointwise convolutional forward propagation process is as follows: (15); among which, This represents the output feature map after pointwise convolution. Each channel is in coordinates Pixel value at; represents the channel index of the output feature map; ELU is the exponential linear unit activation function; For corresponding Channel bias; BN is batch normalization. This is the weight matrix.

11. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 10, characterized in that, In step S3, the improved EEGNet introduces an adaptive channel attention mechanism (AM) for the first channel. One channel, (16); Among them, the Dynamic incentive weights for each channel (17); The first channel description vector There are 1 component; Sigmoid is the activation function; Weights for the dimensionality reduction layer; For activation functions; This is used to assign weights to the dimensionality-upper layer; finally, the weights are multiplied channel by channel to achieve the desired result. Apply to the original feature channel Above, the feature reconstruction is completed; (18); The feature results of the reconstruction; The first point of the output result of the pointwise convolution Feature map of each channel.

12. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 11, characterized in that, In step S3, global average pooling is performed on the reconstructed features. (19); among which, The first term after global average pooling The weight coefficients for each channel are determined; subsequently, the feature vector is input into the fully connected layer. (20); These are the weights of the fully connected layer; For the output of the fully connected layer for the first Bias in the results of fault classes; using the Softmax function to map high-dimensional features to the fault probability space: (twenty one); The sample was ultimately determined to be The probability of a type of failure; For the classification layer targeting the first The logical value output by the fault type; Index for fault categories; This represents the total number of fault types.

13. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 12, characterized in that, In step S3, the cross-entropy loss function is used as the objective optimization function: (22); among which, The true label after one-hot encoding; The w-th sample is judged as the w-th sample. The probability of a fault class; Total is the total number of samples used in this training.

14. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 13, characterized in that, In step S3, a cosine annealing strategy with hot restart is introduced, which participates in all learnable parameters within the network architecture throughout the process. (23); among which, The learning rate used in the current round; This is the lower bound of the learning rate in the i-th cycle; This is the upper bound of the learning rate during the i-th cycle; This refers to the training rounds already performed within the current cycle; Let be the total length of the i-th cycle.

15. The on-load tap changer fault diagnosis method based on geometric feature mapping and enhanced EEGNet according to claim 14, characterized in that, In step S3, the attribution analysis method SHAP is introduced for a given input map feature. Its prediction function The additive interpretation model is as follows: (24); among which, This represents the total number of pixels. The baseline predicted mean of the model across all samples; For the first Shapley values ​​for each pixel; Characterizing the first in the original map The activation state of each pixel feature; P SHAP Determine the probability for a specific fault category; (25); For those obtained through random sampling that do not include the first 1 pixel A subset of pixels; For the current subset The number of features; Indicates when features The predicted probability of the model for a specific fault category when it is in an active or inactive state; ! represents factorial.

Images