Method for recognizing partial discharge mode of electric reactor based on hologram atlas and contrast learning

Abstract

The application relates to a reactance partial discharge pattern recognition method based on hologram and contrast learning, relates to the technical field of partial discharge detection, and is used for solving the problem that current dry-type air-core reactor state monitoring cannot fuse multiple physical semantics, efficiently filters out field working condition noise, and can capture long-range correlations between macroscopic topologies and microscopic textures of partial discharge spectra under a light-weight architecture. The application removes redundant features based on a Spearman coefficient, reconstructs a feature vector into an RGB three-channel feature hologram, in a self-supervised training stage, designs sample pairs, and inputs the sample pairs into a neural network model ResNet-50 in parallel for contrast; finally, a light-weight network architecture LightViT-PD directly receives high-level semantic feature maps output by the backbone network, and a classification head maps and outputs a partial discharge defect type of the dry-type air-core reactor. The application can effectively filter out non-essential physical fluctuation interference in a substation field, and improve the generalization robustness of partial discharge recognition under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of partial discharge detection technology, specifically a method for partial discharge pattern recognition of reactors based on holographic spectra and contrastive learning. Background Technology

[0002] Under long-term electromagnetic, thermal, mechanical vibration, and combined environmental stress, defects can easily develop inside the insulation of reactors, leading to partial discharge and potentially causing insulation breakdown or burnout. Therefore, developing efficient partial discharge mode identification is of great engineering significance for timely detection of early insulation defects in dry-type air-core reactors and for preventing catastrophic failures.

[0003] However, existing methods for partial discharge pattern recognition in reactors still face the following serious challenges in practical engineering applications: First, the incompleteness and redundancy of multi-dimensional heterogeneous feature extraction. Partial discharge exhibits extremely complex physical mechanisms under different excitation conditions. Existing diagnostic methods often rely solely on single time-domain amplitude, frequency-domain spectrum, or pulse group statistical features, making it difficult to comprehensively characterize the deep-seated physical laws of partial discharge pulses in their microscopic pulse morphology, macroscopic pulse group distribution, and transient time-frequency expansion. Simultaneously, there is often high nonlinear redundancy and mutual interference among multi-source features, making it difficult for traditional feature selection methods to accurately capture the core features that truly contribute to pattern recognition while eliminating redundancy. Second, the large interference in substation field conditions leads to poor model generalization robustness. The industrial field monitoring environment is extremely complex, affected by physical factors such as sensor triggering errors, power grid frequency fluctuations, strong electromagnetic interference, and gain differences caused by sensor installation distance. The acquired partial discharge signals often exhibit random phase drift, local feature loss and observation window drift, as well as nonlinear fluctuations in amplitude intensity. Traditional deep learning diagnostic models often rely on rote memorization of shallow visual features or absolute feature values ​​for identification. This makes them prone to misclassifying non-essential physical fluctuations and operational disturbances as different fault categories, leading to high false alarm rates and severely insufficient generalization ability under complex conditions. Thirdly, there is an extreme scarcity of labeled partial discharge samples in the field, coupled with the limitations of traditional models' local perception. In actual substation operation, acquiring partial discharge data with clearly labeled defect categories is extremely costly, making it difficult for traditional fully supervised learning models to extract highly separable deep intrinsic topological features from scarce labeled samples. Furthermore, while traditional convolutional neural networks (CNNs) possess local perception capabilities, their limited receptive field makes it difficult to effectively uncover long-range global dependencies between different spatial and temporal frequency features in the holographic map. Traditional deep learning models like ViT directly slice the original image pixels, losing the inherent physical map induction bias of partial discharge features, and their large number of parameters makes them unsuitable for the lightweight deployment requirements of edge devices in industrial fields.

[0004] Therefore, how to construct a partial discharge pattern recognition method that can integrate multiple physical semantics, efficiently filter out on-site operating noise, and capture the long-range correlation between macroscopic topology and microscopic texture of partial discharge maps under a lightweight architecture is an urgent problem to be solved in the field of condition monitoring of dry-type air-core reactors. Summary of the Invention

[0005] The purpose of this invention is to provide a method for partial discharge pattern recognition of reactors based on holographic maps and contrastive learning, which solves the problem that current dry-type hollow reactor condition monitoring does not have the ability to both integrate multi-dimensional physical semantics, efficiently filter out on-site operating noise, and capture the long-range correlation between macroscopic topology and microscopic texture of partial discharge maps under a lightweight architecture.

[0006] The technical solution adopted by the present invention to solve its technical problem is: a reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning, which includes the following steps.

[0007] Partial discharge signal data acquisition for S1 dry-type air-core reactor.

[0008] S2. Multivariate feature extraction and RGB feature map construction of partial discharge data.

[0009] S2.1 Multivariate Feature Extraction: Feature extraction of partial discharge signals is performed at the pulse-level, statistical-level, and time-frequency-level. The original discharge signal is transformed into a high-dimensional heterogeneous feature vector based on pulse-level and statistical-level features. We employ an improved self-supervised learning framework, SimCLR, to construct physically meaningful feature holograms. Combined with a domain-adaptive enhancement strategy, this enables efficient decoupling and representation of partial discharge features.

[0010] S2.2 RGB Feature Map Construction: The physical characteristics of the partial discharge signal are reconstructed into an image form, that is, the pulse-level features, statistical features and time-frequency features of the signal are mapped to the three channels R, G and B, so that each channel carries specific physical information.

[0011] S3 model comparison between reinforcement learning and lightweight visual Transformer partial discharge classification.

[0012] S3.1 Model Contrastive Augmentation Learning: A contrastive augmentation strategy based on SimCLR physical constraints is designed to learn robust invariant representations by maximizing the information between different augmented views of the same source samples. This forces the model to strip away non-essential physical fluctuations and focus on the inherent topological structure of the partial discharge signal.

[0013] S3.2 Partial discharge pattern recognition and classification based on the lightweight deep learning model ViT architecture.

[0014] S3.2.1 Feature Dimensionality Reduction: The high-level semantic feature map output by the SimCLR backbone network is spatially flattened and converted into a labeled sequence. , where the sequence length O and T are the original feature map dimensions; a linear projection layer is introduced to map the high-dimensional features to the low-dimensional embedding space;

[0015] S3.2.2 Lightweight Self-Attention Encoding: The Multi-Head Self-Attention (MSA) mechanism projects the input into a vector of query o, key r, and value s;

[0016] S3.2.3 Classification and Recognition: A global average pooling strategy is used to classify the encoder output sequence. Aggregation is performed by averaging all tokens along the sequence dimension Y to generate a unique global feature vector. ,Will The input is fed into a multilayer perceptron (MLP) classification head containing a softmax activation function, and the classifier maps the latent features to the posterior probability distributions of four types of partial discharge defects. .

[0017] Further, in step S1, partial discharge signals are acquired, the captured original partial discharge signals are parsed into structured three-dimensional feature vectors, and stored in a standardized CSV format. Each record contains three core attributes: discharge amplitude A, discharge phase φ, and occurrence time t.

[0018] Furthermore, in step S2.1, the original discharge signal is transformed into a high-dimensional heterogeneous feature vector based on the pulse-level and statistical-level characteristics. We employ an improved SimCLR self-supervised architecture to construct physically meaningful feature holograms, and combine this with a domain adaptive enhancement strategy to achieve efficient decoupling and characterization of partial discharge features.

[0019] Furthermore, in step S2.1, pulse-level feature extraction involves introducing pulse energy features. (1) Characterizes the overall energy level of a single pulse, where A(n) is the amplitude of the discrete signal of the partial discharge pulse after denoising within the observation window, n=1,2,…,M; M is the number of pulse sampling points; the peak amplitude of the pulse is used. With effective value range The transient intensity and average energy level of the pulse are described. (2); (3); where max is the maximum value selected from the absolute values ​​of the amplitudes of the M discrete partial discharge pulse signals within the observation window; a waveform factor is introduced. With peak factor Characterizes pulse morphology, (4); (5).

[0020] Furthermore, in step S2.1, the statistical features are extracted by introducing the phase mean. With phase variance Statistical description of pulse phase distribution. (6); (7); among which, To determine the power frequency phase corresponding to the nth pulse within the observation window, a phase skew is introduced. Phase steepness , (8); (9) Characterize the asymmetry and concentration pattern of phase distribution; introduce a phase concentration index based on circular statistical theory. (10) To measure the degree of concentration of the pulse phase within the period.

[0021] Further, in step S2.1, the time-frequency level features are extracted by performing j-level wavelet packet decomposition on A(n) to obtain the wavelet packet coefficient sequences located at nodes (j,k) in different frequency bands. Where j represents the number of decomposition layers, and k represents the corresponding frequency band number; energy index (11); Define the energy percentage characteristics : (12); Introducing energy entropy (13) Statistical representation of the overall time-frequency structure.

[0022] Further, in step S2.2, take For any two features, calculate the correlation between the two indicator features using the Spearman coefficient, and let the number of observed samples for each feature be denoted as . Spearman rank correlation coefficient between two indicator features (14); among which, This represents the difference between the rank sequences of the two features for the i-th sample; a correlation threshold is set. ,like Then, a redundant feature is removed, and finally, only the most redundant features are selected and retained. Core features, reorganized into ; by using a linear projection matrix to Each subset is constructed into three independent one-dimensional feature vectors. , and : (15); where: This represents the weighted pulse-level feature subset; This represents a weighted subset of statistical features. Represents a time-frequency level feature subset; , , It is a linear projection matrix; , , The bias vector is used; in the RGB channel mapping, the R channel represents the pulse characteristics of the partial discharge signal, and the mapping method is as follows: (16); Channel G represents the statistical characteristics of the partial discharge signal, and the mapping method is as follows: (17); Channel B reflects the time-frequency characteristics of the partial discharge signal, and the mapping method is as follows: (18).

[0023] Furthermore, in step S2.2, a linear projection layer is designed to map the one-dimensional feature vector after RGB channels. Mapped to a two-dimensional space: (19); among which, It is a two-dimensional feature mapped with RGB channels; This is the projection weight matrix; As a bias term; introducing position embeddings enhances the model's feature learning ability while preserving the spatial relationships of physical features: (21); among which, Features after adding position encoding; This is the position embedding matrix.

[0024] Further, in step S3.1, the first type of enhancement strategy includes phase translation and physically-aware clipping; phase translation: for a given displacement... Its mathematical mapping is: (22); among which, This represents the element value located at the w-th phase index and the c-th channel in the newly generated positive sample feature map after phase shifting; mod represents the modulo operation. For phase index; is the width dimension of the matrix; For channel indexing; Physically-aware clipping: Define the mask for the random clipping window as... The area ratio of element 1 satisfies : (23); where θ represents the area proportion of the random cropping window; θ max θ represents the maximum allowed area percentage threshold during the cropping operation. min This indicates the minimum area percentage threshold allowed during the cropping operation; The feature map is enhanced by physical perception cropping. This represents the Hadamard product of multiplying each pair of feature map elements one by one. The bilinear interpolation resampling operator restores the cropped region to the original image size. The second type of enhancement strategy is achieved through feature strength dithering: (24); among which, This is a newly generated enhanced feature map after the feature intensity jitter operation; This is the channel scaling factor; This is a random noise bias.

[0025] Furthermore, in step S3.1, the deep residual network ResNet-50 is selected as the core backbone network for deep feature analysis: First, the first convolutional layer of ResNet-50 is adapted. In the self-supervised training stage, positive sample pairs generated from the same standard map after being processed by two random processes from three physical constraint enhancement operators are converted into parallel inputs. The data flows through a ResNet-50 residual block consisting of a dimensionality reduction layer, a feature extraction layer, a dimensionality increase layer, and residual connections. (25); among which, It is an upgraded matrix; For feature extraction matrix, It is a dimension-reduced matrix; The bias of the three convolutional layers; This is feature data used for subsequent Vit classification and discrimination; The input sample is a partially discharged holographic feature map after physical constraint data augmentation. At the end of the backbone network, a global average pooling layer is used to compress the spatial dimension of the high-dimensional feature map, and finally, a high-dimensional feature vector containing profound partial discharge mechanism is extracted. (26); where h q for The high-dimensional feature vector extracted by ResNet-50; h m for High-dimensional feature vectors extracted by ResNet-50; This indicates ResNet-50 and pooling layer operations.

[0026] Furthermore, in step S3.1, a nonlinear projection head is cascaded after the backbone network: a high-dimensional feature vector h q After nonlinear projection mapping, the feature is further normalized and projected into a low-dimensional hyperspherical feature space to obtain the feature representation. (27); among which, and This is the weight matrix of the projection head; and It is the bias vector; Represents the ReLU activation function; Then it was normalized. High-dimensional feature vectors and After passing through a nonlinear projection head, the result is... and .

[0027] Furthermore, in step S3.1, SimCLRv3 ​​uses the normalized temperature-scaled cross-entropy loss function as the network's reward and punishment mechanism: First, it defines any two feature vectors in the space. and Cosine similarity metric between (28); among which, To calculate the L2 norm; suppose a training batch contains N original partial discharge samples, which are then enhanced into 2N enhanced samples after dual-path physical enhancement. For a specific pair of positive samples... The corresponding projection vector is Then the loss function of the sample pair (29); among them, These are temperature control parameters; For an indicator function, if and only if The value is 1; exp(·) is a power function of e; the joint optimization objective of the network is to minimize the total contrastive loss of the entire batch through gradient backpropagation. (30), among which, Indicates the first The enhanced sample is used as the anchor point, with the first enhanced sample as the anchor point. The contrastive loss calculated when each augmented sample is used as a positive sample target; Indicates the first The enhanced sample is used as the anchor point, with the first enhanced sample as the anchor point. The contrastive loss calculated when an augmented sample is used as the positive sample target.

[0028] Further, in step S3.2.1, the projection matrix is ​​defined. Introducing learnable position embedding vectors Initial input sequence (31); among which, Initialize as an orthogonal matrix to maintain the variance stability of the features; The relative positional relationships of encoded partial discharge features in the grid are dynamically updated through backpropagation.

[0029] Furthermore, in step S2.2.2, in the first... In the layer encoder, Input sequence First, layer normalization is performed, followed by the Multi-Head Self-Attention (MSA) module. The mathematical expression of MSA is as follows: (32); (33); and: (34); where, orT Perform a matrix dot product operation between the query vector and the key vector; represents the intermediate output sequence of the l-th layer encoder; Softmax is the normalized exponential activation function; The transformation matrix is ​​automatically learned and dynamically updated by the network through backpropagation under end-to-end supervised training. Subsequently, the feature sequence enters an MLP layer consisting of two linear transformations and a GELU activation function for position-by-position nonlinear feature transformation. (35); among which, This represents the weight matrix of the first linear transformation in a multilayer perceptron. This represents the weight matrix of the second-layer linear transformation in a multilayer perceptron; and These represent the bias term vectors corresponding to the linear transformations; LN represents layer normalization; and GELU represents the activation function.

[0030] Furthermore, in step S3.2.3, (36); (37); among which, For the first The feature vectors output by the MLP at each position; This is the weight matrix for the classification heads; This is the bias term for the classification header.

[0031] The beneficial effects of this invention are as follows: This invention eliminates redundant features by using Spearman coefficients and reconstructs RGB three-channel two-dimensional holographic maps, achieving efficient fusion and structured representation of multiple physical features; then, by leveraging the SimCLR self-supervised contrastive learning framework, a physical constraint enhancement strategy forces the backbone network to identify deep partial discharge topological features with essential invariance; finally, the lightweight network architecture LightViT-PD directly handles high-level semantic feature maps, retaining global long-range dependency modeling capabilities while incorporating convolutional inductive bias, achieving collaborative extraction of macroscopic topology and microscopic texture of partial discharge maps with extremely low parameter count. This invention proposes a highly feasible solution to the problems of feature redundancy, poor generalization robustness, scarce labels, and difficult model deployment in traditional diagnostics. Attached Figure Description

[0032] Figure 1 A schematic diagram of the ResNet-50 backbone network architecture and bottleneck residual units;

[0033] Figure 2 Design diagram of LightViT-PD, a lightweight network architecture for partial discharge characteristics. Detailed Implementation

[0034] This invention, based on the feature extraction of the original data, removes redundant features using Spearman coefficients, and then reconstructs the feature vectors into an RGB three-channel two-dimensional feature hologram, achieving efficient data representation. In the self-supervised training phase, sample pairs are designed and input into a Siamese ResNet-50 backbone network for comparison, optimizing the backbone network parameters. Finally, the lightweight network architecture LightViT-PD directly receives the high-level semantic feature map output from the backbone network, and the classification head maps and outputs the partial discharge defect type of the dry-type hollow reactor. This invention can effectively filter out non-essential physical fluctuation interference in substations and significantly improve the generalization robustness of partial discharge pattern recognition under complex operating conditions. A reactor partial discharge pattern recognition method based on holograms and contrastive learning includes the following steps.

[0035] Partial discharge signal data acquisition for S1 dry-type air-core reactor.

[0036] Based on a scaled-down model defect test platform for dry-type air-core reactors, a systematic partial discharge signal acquisition experiment was conducted. The captured raw partial discharge signals were analyzed into structured three-dimensional feature vectors and stored in standardized CSV format. Each record contains three core attributes: discharge amplitude A, discharge phase φ, and occurrence time t.

[0037] S2. Multivariate feature extraction and RGB feature map construction of partial discharge data.

[0038] S2.1 Multivariate feature extraction of data.

[0039] S2.1.1 Extraction of pulse-level features.

[0040] Under the excitation of oscillating waves, partial discharge typically manifests as transient pulses with abrupt amplitude changes. Their time-domain waveforms are characterized by steep rising edges, short durations, and a decaying oscillating envelope. These pulses reflect the rapid release of local electric field energy at the moment of discharge, containing key information about discharge intensity and oscillation structure. Therefore, we first characterize the partial discharge signal at the pulse level. Let the amplitude of the denoised discrete signal of the partial discharge pulse within the observation window be A(n), where n = 1, 2, ..., M; M is the number of pulse sampling points. To characterize the overall energy level of a single pulse, pulse energy characteristics are introduced. : (1). In terms of amplitude characterization, the peak pulse amplitude is used. With effective value range The transient intensity and average energy level of the pulse are described. (2); (3). Where max is the maximum value selected from the absolute values ​​of the amplitudes of the M discrete partial discharge pulse signals within the observation window. A pulse shape factor is further introduced to characterize the pulse shape. This invention selects a waveform factor that reflects the relationship between the pulse's effective value and its average amplitude. Peak factor, which characterizes the prominence of peak amplitude relative to the overall energy level. As an indicator of pulse morphology. (4); (5).

[0041] S2.1.2 Extraction of statistical features.

[0042] In engineering measurement environments, partial discharge is not an isolated event, but rather occurs repeatedly in the form of pulse sequences. The information reflected by a single pulse has certain limitations and is difficult to reveal the overall distribution pattern of discharge activity on the time and phase scales. Therefore, it is necessary to describe the partial discharge signal at the pulse group level to characterize the differences in phase distribution, amplitude level, and repeatability of different discharge sources.

[0043] Let the power frequency phase corresponding to the nth pulse within the observation window be . The amplitude is First, the pulse phase distribution is statistically described, introducing the phase mean. With phase variance , (6); (7). The phase mean reflects the overall concentration of the discharge pulse within the power frequency cycle, while the phase variance describes the dispersion of the phase distribution. To further characterize the asymmetry and concentration pattern of the phase distribution, phase skewness is introduced. Phase steepness , (8); (9). Among them, phase skewness is used to describe the offset trend of phase distribution relative to the mean, and phase steepness is used to reflect the concentration and tail characteristics of phase distribution. and It can supplement the distribution pattern information that is difficult to reflect by the mean and variance. Considering the periodicity of the phase variable, a phase concentration index based on circular statistics theory is further introduced. : (10) is used to measure the degree of concentration of pulse phase within a period. This index is sensitive to the concentration or dispersion of phase distribution and can reflect the concentration characteristics of pulse groups in the power frequency phase dimension from an overall perspective.

[0044] S2.1.3 Extraction of time-frequency features.

[0045] Considering the signal characteristics under oscillating wave excitation conditions, partial discharge pulses generally contain multiple periodic high-frequency oscillation components, whose energy expands and decays across multiple frequency bands within a short period. This energy expansion process exhibits significant non-stationarity, and relying solely on time-domain amplitude or the overall spectrum is insufficient to reflect its frequency band structure differences. To reveal the distribution characteristics of the oscillating components of partial discharge pulses on the frequency scale, it is necessary to introduce time-frequency analysis methods to characterize the pulse energy expansion process at different scales.

[0046] Wavelet packet transform can recursively decompose signals in the time-frequency plane, achieving fine frequency band division while maintaining time resolution. It is suitable for analyzing partial discharge pulse signals with transient oscillation characteristics. By performing j-level wavelet packet decomposition on the amplitude signal A(n) within the observation window, wavelet packet coefficient sequences located at different frequency band nodes (j,k) are obtained. , where j represents the number of decomposition layers and k represents the corresponding frequency band number. (11). a j，k As an energy index, it reflects the energy concentration of a partial discharge pulse within its corresponding frequency band. Differences in the oscillation structure and propagation path of different discharge sources often lead to different energy distribution characteristics across frequency bands. Based on this, to eliminate the influence of overall pulse amplitude differences on energy comparison, the energy of each frequency band is normalized, defining an energy proportion characteristic. : (12). Examining only the energy proportion of a single frequency band is insufficient to fully describe the dispersion of energy distribution. Therefore, energy entropy H is further introduced as a statistical representation of the overall time-frequency structure. (13). Energy entropy is used to characterize the uniformity of partial discharge pulse energy distribution across frequency bands. When the energy is mainly concentrated in a few frequency bands, the energy entropy value is small; when the energy is more dispersed across multiple frequency bands, the energy entropy increases accordingly.

[0047] Based on the constructed multidimensional feature system of partial discharge, and targeting the pulse-level and statistical-level features obtained under oscillating wave excitation conditions, this invention transforms the original sample data into a high-dimensional heterogeneous feature vector composed of multiple physical quantities, denoted as... Given that the statistical distribution of partial discharge characteristic parameters typically exhibits non-Gaussian properties, and the dependencies between features are often nonlinear monotonic correlations, this invention employs an improved SimCLR self-supervised architecture to construct a physically meaningful feature hologram. Combined with a domain adaptive enhancement strategy, this achieves efficient decoupling and characterization of partial discharge features.

[0048] S2.2 RGB Feature Map Construction.

[0049] To overcome the limitations of multi-level feature recognition, which fails to fully reveal the deep structure and potential laws of signals, the physical characteristics of partial discharge signals are reconstructed into image form. This involves mapping the pulse-level, statistical, and time-frequency features of the signal onto the R, G, and B channels, so that each channel carries specific physical information.

[0050] Given that the statistical distribution of partial discharge characteristic parameters typically exhibits non-Gaussian properties, and the dependencies between features are often non-linear monotonic correlations, non-parametric statistical methods are used to measure feature correlation. Specifically, the Spearman rank correlation coefficient is introduced as an evaluation index for the degree of feature correlation. By comparing the consistency between the ranks of variables to measure their monotonic correlation, the influence of outliers and distribution patterns on the correlation estimation results can be effectively avoided.

[0051] The high-dimensional feature vectors constructed by the aforementioned method are known. Take any two features from the feature vector set, and use the Spearman coefficient to calculate the correlation between the two feature indicators. Let the number of observed samples for each feature be denoted as . After arranging the original observations of the corresponding features in each sample from largest to smallest and converting them into the corresponding rank sequence, the Spearman rank correlation coefficient between the two is calculated. for: (14). Among them, This represents the difference between the rank sequences of the two features for the i-th sample; the correlation threshold is set to... If it appears Then, a redundant feature is removed, and finally, only the most redundant features are selected and retained. Core features, reorganized into .

[0052] To accommodate subsequent image channel dimension mapping requirements, a linear projection matrix is ​​used to... Each subset is constructed into three independent one-dimensional feature vectors. , and : (15). Among them: This represents the weighted pulse-level feature subset; This represents a weighted subset of statistical features. Represents a time-frequency level feature subset; , , It is a linear projection matrix; , , This is the bias vector. Through the above cascaded processing, the original high-dimensional sample feature vectors are transformed into three specific sets of feature vectors. , and These are used for two-dimensional spatial texture mapping of the R-channel pulse features, G-channel statistical features, and B-channel time-frequency features in subsequent RGB holograms. In the RGB channel mapping, the R-channel represents the pulse features of the signal, and the specific mapping method is as follows: (16). The G channel represents the statistical characteristics of the partial discharge signal, specifically mapped as follows: (17); Channel B reflects the time-frequency characteristics of the signal, with the specific mapping method as follows: (18).

[0053] Since the feature space of partial discharge signals is usually low-dimensional, while the processing requires a two-dimensional image format, a linear projection layer needs to be designed to map the one-dimensional feature vector after RGB channels. This is mapped to a two-dimensional space. This process transforms the feature values ​​of each partial discharge signal into a spatial texture distribution, enabling subsequent learning models to effectively extract more complex spatial features. The specific formula is: (19). Among them, It is a two-dimensional feature mapped from the RGB channels; This is the projection weight matrix; This is the bias term. To further enhance the model's feature learning ability and preserve the spatial relationships of physical features, position embedding is introduced. The effectiveness of the receptive field is closely related to the spatial structure of the input features. To help the model better understand the spatial relationships of the data, a unique encoding value is assigned to each feature location, enabling the model to fully consider their relative positions when processing features. The role of position embedding is similar to position encoding in the Transformer model, a deep learning architecture based on self-attention mechanisms. It can effectively preserve the spatial structure of features and enhance the model's sensitivity to the correlations between features. The formula for position embedding is: (21). Among them, Features after adding position encoding; This is the position embedding matrix.

[0054] S3 model comparison between reinforcement learning and lightweight visual Transformer partial discharge classification.

[0055] S3.1 Model Comparative Reinforcement Learning.

[0056] To further explore the essential representation of partial discharge signals in deep feature space, a contrast enhancement strategy based on physical constraints of the self-supervised learning framework SimCLR is designed. Its core is to learn robust invariant representations by maximizing the information between different enhanced views of the same source samples, forcing the model to strip away non-essential physical fluctuations and focus on the inherent topological structure of the partial discharge signal.

[0057] The first type of enhancement strategy is spatiotemporal topological consistency enhancement, including phase shifting and physically-aware clipping. Phase shifting: Considering the physical characteristic of the random phase distribution of partial discharge signals within the power frequency cycle, a cyclic shift operator along the phase dimension is designed, assuming the width dimension of the matrix is... For a given displacement Its mathematical mapping is: (22). Among them, For phase index; For channel indexing; This represents the element value located at the w-th phase index and c-th channel in the newly generated positive sample feature map after phase shifting. It inherits from the original feature map, but the phase position has been shifted. `mod` is the modulo operation to ensure the phase shift is correct. When the image width exceeds the specified width, the signal can be cyclically returned to the starting position, thereby maintaining the continuous periodicity of the partial discharge signal in the time domain and preventing data distortion due to out-of-bounds errors.

[0058] This operation constructs positive samples with significant phase differences but consistent physical topology, forcing the network to eliminate its dependence on absolute phase position and instead focus on capturing the periodicity and morphological regularity of the hologram in the time domain. Physically perceptual clipping: To simulate the loss of local features or observation window drift in on-site monitoring, clipping and resampling operators were designed. The mask for the random clipping window is defined as... The area ratio of element 1 satisfies The final output after completing the physical perception clipping. (23). Where θ represents the area ratio of the random cropping window, i.e., the mask. The ratio of the area of ​​the region with a value of 1 to the total area u×x of the original feature map. θ max θ represents the maximum allowed area percentage threshold during the cropping operation. min This indicates the minimum area percentage threshold allowed during the cropping operation; This is the enhanced feature map after physical perception cropping. Let θ be... max and θ min This is to prevent the cropping area from being too large, resulting in the complete loss of key physical features, or too small, thus failing to achieve the desired data augmentation effect. Among these, This represents the Hadamard product of multiplying each pair of feature map elements one by one. The bilinear interpolation resampling operator restores the cropped region to the original image size. This operator cuts off the model's over-reliance on global statistics, forcing it to extract global representations from local texture details.

[0059] The second type of enhancement strategy is amplitude invariance enhancement, mainly achieved through characteristic intensity jitter. Although the data is normalized, different discharge intensities and sensor gains can still cause overall fluctuations in the holographic spectrum colors. To address this issue, characteristic intensity jitter applies random linear gain perturbations to the channels, simulating nonlinear changes in intensity. (24); among which, This represents the complete two-dimensional spatial matrix of the newly enhanced feature map generated after feature intensity dithering, located in the c-th channel. This is the channel scaling factor; As a random noise bias, this operator avoids overfitting the model to the absolute numerical value of the features by simulating the nonlinear change of the intensity, and significantly improves the generalization robustness of the model under varying conditions.

[0060] A deep residual network, ResNet-50, was selected as the core backbone network for deep feature analysis. Considering the unique three-dimensional structure of the partial discharge feature hologram, the first convolutional layer of ResNet-50 was first adapted to directly handle RGB multi-channel inputs containing multi-source physical semantics. During the self-supervised training phase, positive sample pairs generated from the same standard graph after being processed by two randomized methods of three physical constraint enhancement operators are converted into parallel inputs. The data flows through a ResNet-50 residual block consisting of a dimensionality reduction layer, a feature extraction layer, a dimensionality increase layer, and residual connections. (25). Among them, It is a dimension-reduced matrix; For feature extraction matrix, It is an upgraded matrix; The bias of the three convolutional layers; This is the feature data used for subsequent Vit classification; ReLU is a non-linear activation function. The input sample is a partially discharged holographic feature map after physical constraint data augmentation. A global average pooling layer at the end of the backbone network is used to compress the spatial dimension of the high-dimensional feature map, ultimately extracting a high-dimensional feature vector containing profound insights into the mechanism of partially discharged discharge. (26). Among them, h q for The high-dimensional feature vector extracted by ResNet-50; h m for High-dimensional feature vectors extracted by ResNet-50; This indicates ResNet-50 and pooling layer operations.

[0061] To eliminate feature redundancy and improve its separability in low-dimensional space, a nonlinear projection head is cascaded after the backbone network. The nonlinear projection head consists of two levels of multilayer perceptrons and embeds the ReLU nonlinear activation function. The high-dimensional feature vector h is mapped by the nonlinear projection head and further normalized and projected onto the low-dimensional hyperspherical feature space to obtain the feature representation. : (27). Among them, and This is the weight matrix of the projection head; and It is the bias vector; This represents the ReLU activation function. Then it was normalized. This is beneficial for subsequent calculations of geometric distance. High-dimensional feature vectors and After passing through a nonlinear projection head, the result was... and .

[0062] To achieve compact intra-class and segregated inter-class feature selection within the hyperspherical latent space, SimCLR employs a normalized temperature-scaled cross-entropy loss function as the network's reward and penalty mechanism. First, it defines any two feature vectors in the space... and Cosine similarity metric between them: (28). Among them, To calculate the L2 norm, suppose a training batch contains N original partial discharge samples, which are then enhanced into 2N enhanced samples after dual-path physical enhancement. For a specific pair of positive samples... Its corresponding projection vector is Then the loss function of the sample pair (29). Among them, This is a temperature adjustment parameter used to control the granularity of the network's penalty for difficult negative samples; For an indicator function, if and only if The value is 1, used to remove the sample's own contrast from the denominator, and exp(·) is a power function of e. The joint optimization objective of the entire network is to minimize the total contrast loss of the entire batch through gradient backpropagation. : (30). Indicates the first The enhanced sample is used as the anchor point, with the first enhanced sample as the anchor point. The contrastive loss calculated when each augmented sample is used as a positive sample target; Indicates the first The enhanced sample is used as the anchor point, with the first enhanced sample as the anchor point. The contrastive loss is calculated when an enhanced sample is used as the positive sample target. Through strong constraint rewards and penalties, the model is forced to maximize the mutual information between different views of the same source sample during the optimization process, while significantly widening the distance between different partial discharge types. This mechanism fundamentally drives the backbone network ResNet-50 to strip away non-essential operating condition interferences and accurately extract the deep topological features of partial discharges with essential invariance.

[0063] S3.2 Partial discharge pattern recognition and classification based on the lightweight deep learning model ViT architecture.

[0064] For the subsequent task of partial discharge pattern recognition in dry-type air-core reactors, a lightweight LightViT-PD architecture oriented towards partial discharge feature representation is adopted. Building upon the advantages of Transformer's global long-range dependency modeling, the LightViT-PD architecture integrates the inductive bias of convolutional neural networks in terms of local feature perception and translation invariance, aiming to achieve efficient collaborative extraction of macroscopic topology and microscopic texture of partial discharge maps.

[0065] S3.2.1 Feature dimensionality reduction processing.

[0066] LightViT-PD does not employ the traditional ViT method of slicing raw pixels at its input; instead, it directly receives the high-level semantic feature map output from the SimCLR backbone network. Let the feature tensor output by SimCLR be... This feature map preserves the spatial topology of the partial discharge signal in the time-frequency domain, and each feature point contains highly abstract physical semantic features. To adapt to the serialization input requirements of the Transformer and remove redundant information in the feature space, a spatial flattening operation is first performed to convert the feature map into a labeled sequence. , where the sequence length O and T are the original feature map dimensions. Subsequently, a linear projection layer is introduced to map the high-dimensional features to a low-dimensional embedding space. Based on the analysis of the low-rank characteristics of partial discharge features, a projection matrix is ​​defined. To achieve model lightweighting, feature dimensions are compressed. To compensate for the loss of spatial location information caused by the flattening operation, learnable location embedding vectors are introduced. Initial input sequence (31). Among them, Initialize as an orthogonal matrix to maintain the variance stability of the features; The relative positional relationships of encoded partial discharge features in the grid are dynamically updated through backpropagation.

[0067] S3.2.2 Lightweight self-attention encoding.

[0068] The feature encoding stage consists of L stacked lightweight Transformer encoder blocks. Each encoding block comprises a multi-head self-attention layer and a multi-layer perceptron layer, supplemented by layer normalization and residual connections to establish stable gradient propagation. In the... In the layer encoder, Input sequence First, layer normalization is performed, followed by the multi-head self-attention mechanism (MSA) module, with the number of heads set to 4, and the feature dimension of each attention head is [not specified]. MSA projects the input as a vector of query o, key r, and value s, and captures long-range dependencies of features at different spatial locations in the partial discharge map by calculating scaled dot product attention between the tags. The mathematical expression of the attention mechanism is: (32); (33); and: (34). Among them, or T Perform a matrix dot product operation between the query vector and the key vector; is the intermediate output sequence of the l-th layer encoder; Softmax is the normalized exponential activation function, which is used to map the similarity scores obtained by scaling the dot product to a probability distribution with a sum of 1, so as to determine the attention allocation weights between different spatial locations in the feature sequence. The transformation matrix is ​​automatically learned and dynamically updated by the network through backpropagation under end-to-end supervised training. The feature sequence then enters an MLP layer consisting of two linear transformation layers and a GELU activation function for position-by-position nonlinear feature transformation. The encoder output sequence... (35). Among them, This represents the weight matrix of the first linear transformation in a multilayer perceptron (MLP). This represents the weight matrix of the second-layer linear transformation in a multilayer perceptron (MLP). and These represent the bias term vectors corresponding to the linear transformations. They are network parameters that are automatically learned and dynamically updated through backpropagation during end-to-end supervised training, aiming to complete the positional nonlinear spatial mapping and extraction of features. LN is layer normalization; GELU is the activation function; through deep interaction of layers L, the model can further explore the overall relationship between features while retaining the physical features extracted by SimCLR, thereby forming a feature representation with strong discriminative power.

[0069] S3.2.3 Classification and Recognition.

[0070] In the output stage, to overcome the local information bias that may be caused by a single class label, LightViT-PD employs a global average pooling strategy for the encoder output sequence. Aggregation is performed. A global average pooling operation averages all tokens along the sequence dimension Y, generating a unique global feature vector. This strategy can integrate feature information from the entire image, enhancing the model's robustness to shifts in the location of partial discharge occurrences. (36); among them, For the first The feature vectors at each position are processed by the MLP output. Finally, the feature vectors... The input is fed into an MLP classification head containing a Softmax activation function. The classifier maps the latent features to posterior probability distributions of four partial discharge defects. : (37). Among them, This is the weight matrix for the classification heads; This is the bias term for the classification head. The LightViT-PD architecture achieves efficient decoupling and classification of SimCLRv3 ​​pre-trained features by minimizing cross-entropy loss through end-to-end supervised training.

[0071] In summary, this invention proposes a method for partial discharge pattern recognition of dry hollow reactors that integrates contrastive learning and a lightweight Vision Transformer. By eliminating redundant features using Spearman coefficients and reconstructing RGB three-channel two-dimensional holographic maps, it achieves efficient fusion and structured representation of multiple physical features. Then, leveraging the SimCLR self-supervised contrastive learning framework, a physical constraint enhancement strategy forces the backbone network to identify deep partial discharge topological features with inherent invariance. Finally, the lightweight LightViT-PD directly handles the high-level semantic feature map, retaining global long-range dependency modeling capabilities while incorporating convolutional inductive bias. This achieves collaborative extraction of macroscopic topology and microscopic texture of the partial discharge map with extremely low parameter count. This invention offers a highly feasible solution to the problems of feature redundancy, poor generalization robustness, scarce labels, and difficult model deployment in traditional diagnostic methods.

Claims

1. A method for partial discharge pattern recognition of reactors based on holographic maps and contrastive learning, characterized in that, Includes the following steps: S1 acquires data on the partial discharge signal of the dry-type air-core reactor; Multivariate feature extraction and RGB feature map construction of S2 partial discharge data; S2.1 Multivariate feature extraction of data: Feature extraction of partial discharge signals is performed from three levels: pulse level, statistical level, and time-frequency level; S2.2 RGB Feature Map Construction: Reconstructing the physical features of the partial discharge signal into an image format; taking... For any two features, let the number of observed samples for each feature be denoted as . Spearman rank correlation coefficient between two indicator features ;in, This represents the difference between the rank sequences of the two features for the i-th sample; a correlation threshold is set. ,like Then, a redundant feature is removed, and finally, only the most redundant features are selected and retained. Core features, reorganized into ; by using a linear projection matrix to Each subset is constructed into three independent one-dimensional feature vectors. , and : ;in: This represents the weighted pulse-level feature subset; This represents a weighted subset of statistical features. Represents a time-frequency level feature subset; , , It is a linear projection matrix; , , The bias vector is used; in the RGB channel mapping, the R channel represents the pulse characteristics of the partial discharge signal, and the mapping method is as follows: The G channel represents the statistical characteristics of the partial discharge signal, mapped as follows: Channel B reflects the time-frequency characteristics of the partial discharge signal, and the mapping method is as follows: Design a linear projection layer to map the one-dimensional feature vector after RGB channels. Mapped to a two-dimensional space: ;in, It is a two-dimensional feature mapped with RGB channels; This is the projection weight matrix; As a bias term; introducing position embeddings enhances the model's feature learning ability while preserving the spatial relationships of physical features: ;in, Features after adding position encoding; For the position embedding matrix; S3 Model Comparison with Reinforcement Learning and Lightweight Visual Transformer Partial Discharge Classification; S3.1 Model Contrast Augmentation Learning: Design a contrast augmentation strategy based on the physical constraints of the self-supervised learning framework SimCLR, and learn robust invariant representations by maximizing the information between different augmented views of the same source samples. S3.2 Partial discharge pattern recognition and classification based on the lightweight deep learning model ViT architecture; S3.2.1 Feature Dimensionality Reduction: The high-level semantic feature maps output by the SimCLR backbone network are spatially flattened and converted into labeled sequences. ; where, sequence length O and T are the original feature map dimensions; a linear projection layer is introduced to map the high-dimensional features to the low-dimensional embedding space; S3.2.2 Lightweight Self-Attention Encoding: The multi-head self-attention mechanism MSA is used to project the input into a vector of query o, key r, and value s; S3.2.3 Classification and Recognition: A global average pooling strategy is used to classify the encoder output sequence. Aggregation is performed by averaging all tokens along the sequence dimension Y to generate a unique global feature vector. ,Will The input is fed into a multilayer perceptron (MLP) classification head containing a softmax activation function, mapping the latent features to the posterior probability distributions of four types of partial discharge defects. .

2. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 1, characterized in that, In step S1, partial discharge signals are acquired, and the captured raw partial discharge signals are parsed into structured three-dimensional feature vectors and stored in a standardized CSV format. Each record contains three core attributes: discharge amplitude A, discharge phase φ, and occurrence time t.

3. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 2, characterized in that, In step S2.1, the original partial discharge signal is transformed into a high-dimensional heterogeneous feature vector based on the pulse-level and statistical-level characteristics. The SimCLR self-supervised learning framework is used to construct feature holograms, and combined with a domain adaptive enhancement strategy, to achieve efficient decoupling and representation of partial discharge features.

4. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 3, characterized in that, Step S2.1 Pulse-level feature extraction: Introducing pulse energy features The overall energy level of a single pulse is represented by A(n), where A(n) is the amplitude of the discrete signal of the partial discharge pulse after denoising within the observation window, n=1,2,…,M; M is the number of pulse sampling points; and the peak amplitude of the pulse is used. With effective value range The transient intensity and average energy level of the pulse are described; where max is the maximum value selected from the absolute values ​​of the amplitudes of M discrete partial discharge pulse signals within the observation window; a waveform factor is introduced. With peak factor Characterizes pulse morphology.

5. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 4, characterized in that, Extraction of statistical features in step S2.1: Introducing the phase mean With phase variance The pulse phase distribution is statistically described; among which, To determine the power frequency phase corresponding to the nth pulse within the observation window, a phase skew is introduced. Phase steepness Characterizing the asymmetry and concentration pattern of phase distribution; introducing a phase concentration index based on circular statistics theory. It measures the degree of concentration of the pulse phase within a period.

6. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 5, characterized in that, Step S2.1 Extraction of time-frequency level features: Perform j-level wavelet packet decomposition on A(n) to obtain wavelet packet coefficient sequences located at nodes (j,k) in different frequency bands. Where j represents the number of decomposition layers, and k represents the corresponding frequency band number; energy index Define the energy percentage characteristics Introducing energy entropy As a statistical representation of the overall time-frequency structure.

7. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 1, characterized in that, In step S3.1, the first type of enhancement strategy includes phase translation and physically-aware clipping; phase translation: for a given displacement... Its mathematical mapping is: ;in, This represents the element value located at the w-th phase index and the c-th channel in the newly generated positive sample feature map after phase shifting; mod represents the modulo operation. For phase index; is the width dimension of the matrix; For channel indexing; Physically-aware clipping: Define the mask for the random clipping window as... The area ratio of element 1 satisfies : Where θ represents the area proportion of the randomly cropped window; θ max θ represents the maximum allowed area percentage threshold during the cropping operation. min This indicates the minimum area percentage threshold allowed during the cropping operation; The feature map is enhanced by physical perception cropping. This represents the Hadamard product of multiplying each pair of feature map elements one by one. The bilinear interpolation resampling operator restores the cropped region to the original image size. The second type of enhancement strategy is achieved through feature strength dithering: ;in, This is a newly generated enhanced feature map after the feature intensity jitter operation; This is the channel scaling factor; This is a random noise bias.

8. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 7, characterized in that, In step S3.1, the deep residual network ResNet-50 is selected as the core backbone network for deep feature analysis. First, the first convolutional layer of ResNet-50 is adapted. During the self-supervised training stage, positive sample pairs generated from the same standard map after being processed by two randomized methods from three physical constraint enhancement operators are converted into parallel inputs. The data flows through a ResNet-50 residual block consisting of a dimensionality reduction layer, a feature extraction layer, a dimensionality increase layer, and residual connections. ;in, It is an upgraded matrix; For feature extraction matrix, It is a dimension-reduced matrix; The bias of the three convolutional layers; This is feature data used for subsequent Vit classification and discrimination; The input sample is a partially discharged holographic feature map after physical constraint data augmentation. At the end of the backbone network, a global average pooling layer is used to compress the spatial dimension of the high-dimensional feature map, and finally, a high-dimensional feature vector containing profound partial discharge mechanism is extracted. (26); where h q for The high-dimensional feature vector extracted by ResNet-50; h m for High-dimensional feature vectors extracted by ResNet-50; This indicates ResNet-50 and pooling layer operations.

9. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 8, characterized in that, In step S3.1, a nonlinear projection head is cascaded after the backbone network: h q Feature representation is obtained through nonlinear projection head mapping. ;in, and This is the weight matrix of the projection head; and It is the bias vector; Represents the ReLU activation function; Then it was normalized. , and After passing through a nonlinear projection head, the result is... and .

10. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 9, characterized in that, In step S3.1, SimCLRv3 ​​uses the normalized temperature-scaled cross-entropy loss function as the network's reward and punishment mechanism: First, define any two feature vectors in the space. and Cosine similarity metric between ;in, To calculate the L2 norm; suppose a training batch contains N original partial discharge samples, which are then enhanced into 2N enhanced samples after dual-path physical enhancement. For a specific pair of positive samples... The corresponding projection vector is Then the loss function of the sample pair ;in, These are temperature control parameters; Let be the indicator function; exp(·) be the power function of e; the joint optimization objective of the network is to minimize the total contrastive loss of the entire batch through gradient backpropagation. ;in, Indicates the first The enhanced sample is used as the anchor point, with the first enhanced sample as the anchor point. The contrastive loss calculated when each augmented sample is used as a positive sample target; Indicates the first The enhanced sample is used as the anchor point, with the first enhanced sample as the anchor point. The contrastive loss calculated when an augmented sample is used as the positive sample target.

11. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 10, characterized in that, In step S3.2.1, the projection matrix is ​​defined. Introducing learnable position embedding vectors Initial input sequence ;in, Initialize as an orthogonal matrix to maintain the variance stability of the features; The relative positional relationships of encoded partial discharge features in the grid are dynamically updated through backpropagation.

12. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 11, characterized in that, In step S3.2.2, at the first In the layer encoder, Input sequence First, layer normalization is performed, followed by MSA (Maximum Subtraction Anomaly). The MSA is expressed as follows: ; And: ; among which, or T Perform a matrix dot product operation between the query vector and the key vector; represents the intermediate output sequence of the l-th layer encoder; Softmax is the normalized exponential activation function; The transformation matrix is ​​automatically learned and dynamically updated by the network through backpropagation under end-to-end supervised training. Subsequently, the feature sequence enters an MLP layer consisting of two linear transformations and a GELU activation function for position-by-position nonlinear feature transformation to obtain the feature vector. ;in, This represents the weight matrix of the first linear transformation in a multilayer perceptron. This represents the weight matrix of the second-layer linear transformation in a multilayer perceptron; and These represent the bias term vectors corresponding to the linear transformations; LN represents layer normalization; and GELU represents the activation function.

13. The reactor partial discharge pattern recognition method based on holographic spectrum and contrastive learning according to claim 12, characterized in that, In step S3.2.3, ; ;in, For the first The feature vectors output by the MLP at each position; This is the weight matrix for the classification heads; This is the bias term for the classification header.

Images