A transformer partial discharge intelligent identification method based on deep learning

Abstract

The application belongs to the technical field of transformer partial discharge identification, and discloses a transformer partial discharge intelligent identification method based on deep learning, which constructs an electroacoustic thermal three-mode acquisition architecture, captures 1ns level discharge pulses through a high-frequency current sensor, avoids sound field superposition interference by reasonably arranging an ultrasonic sensor on the oil tank wall, focuses on the easy discharge area to collect thermal signals through an infrared thermal imager, and synchronously obtains working condition parameters such as load rate and oil temperature; subsequently, the effective signal segments are retained through a multi-modal signal cross-correlation matrix, and the interference signals with large deviations are removed, then different networks are used to extract electroacoustic thermal characteristics, the coupling characteristics are calculated by combining the physical constraint layer weight and the cross-field interaction layer, and the key characteristics are retained; the feature overlap of air gap and surface discharge is effectively reduced, the influence of high-frequency interference and signal distortion is reduced, the partial discharge identification precision is improved, the samples are expanded and the fault and normal sample ratio is reasonable through time stretching and additive noise processing.

Description

Technical Field

[0001] This invention belongs to the field of transformer partial discharge identification technology, specifically a deep learning-based intelligent identification method for transformer partial discharge. Background Technology

[0002] Transformers are crucial equipment in power systems, and their insulation condition determines the reliability of the power grid. Partial discharge signals are a key characteristic reflecting insulation degradation. By analyzing the characteristic parameters of discharge signals, insulation defects such as air gaps, floating potentials, and metal particles can be detected in advance. However, in field operation and maintenance practice, existing deep learning-based partial discharge identification technologies still have the following technical problems:

[0003] Partial discharge generates electrical, acoustic, and thermal signals, but there are multiple fields of interference in the field. The oil temperature gradient causes the ultrasonic signal propagation delay distortion to be greater than ±5μs, and the cross-coupling of the high-voltage electromagnetic field and the current signal introduces high-frequency interference. Existing methods such as single wavelet denoising only process single field signals and do not take into account the correlation between electrical, acoustic, and thermal signals. For example, during normal discharge, the time difference between current and ultrasound should be ≤10μs, which leads to a large deviation between the extracted features and the actual physical laws. The overlap between the air gap and surface discharge features exceeds 30%, resulting in a high misjudgment rate and affecting the recognition accuracy.

[0004] The actual operating conditions of transformers change dynamically with load, voltage transients, temperature and humidity. However, existing models are trained based on static laboratory conditions and do not learn the mapping relationship between operating conditions and features. For example, when the oil temperature rises under high load, the discharge pulse interval is shortened by 20% to 30%, and the model cannot adapt to this, resulting in a significant decrease in the accuracy of on-site identification and difficulty in adapting to dynamic operating characteristics.

[0005] The existing model only outputs the fault type, without feedback on the feature contribution, and the feature contribution is not well matched with the physical laws of discharge. For example, the diameter of the metal particle is 0.1mm → pulse amplitude of 300mV, 1mm → 800mV. Maintenance personnel do not know the basis of the model's judgment, making it difficult to accurately locate the defect and thus unable to formulate targeted maintenance strategies. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based intelligent identification method for transformer partial discharge, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based intelligent identification method for transformer partial discharge, the specific steps of which are as follows: Step S1: Collect current signals, sound signals, thermal signals and operating parameters, set dual labels of operating parameters and discharge type for signal segments, generate scarce samples, and finally establish a dataset containing electricity, sound, heat and operating conditions. Step S2: Fit the mapping relationship between oil temperature, load rate and noise, and set a threshold; remove interference signals from the dataset, and use the fitting formula of load rate and discharge pulse interval to make the ratio of faulty to normal samples reach 1:1.1, and obtain the processed signal data. Step S3: Extract the three-branch features of electrical parameters, acoustic parameters and thermal parameters, set the weights of the physical constraint layer, calculate the coupling features and concatenate them with the three-branch features, retain the key features, and finally output the fused feature vector. Step S4: Improve the ResNet-18 backbone network by adding a working condition adaptation layer before the output layer; determine the contribution of output features to the results, correlate features with fault parameters to generate a source tracing report; design a joint loss function to improve the model's classification accuracy and working condition adaptability. Step S5: Preprocess the signal data processed in step S2, and use the AdamW optimizer and the domain adversarial neural network to optimize the model; perform pruning and distillation on the optimized model to adapt it to edge computing power and obtain a lightweight model. Step S6: Verify the model's accuracy, adaptability to operating conditions, and interpretability; then deploy the model on-site to verify its generalization ability; iteratively adjust the model based on the verification results until the indicators meet the requirements. Step S7: Load the model to the edge gateway and design a mechanism to balance computing power and accuracy; it has real-time monitoring, maintenance guidance and fault response functions, and iterates the model through collaborative updates between the edge and the cloud.

[0008] Furthermore, step S1 is specifically as follows: A three-modal acquisition architecture of electrical, acoustic, and thermal is established. A high-frequency current sensor is directly connected to the transformer grounding wire to collect current signals. An ultrasonic sensor is placed on the tank wall to collect acoustic signals. For easily discharged areas, including bushings and the top cover of the tank, thermal signals are collected by an infrared thermal imager. Operating parameters, such as load rate, oil temperature, voltage level, and ambient humidity, are collected synchronously through the transformer intelligent terminal. Signal segments were extracted based on the half-cycle of the power system's power frequency. The operating parameters and discharge type were dual-labeled and cross-reviewed by two testing engineers. Meanwhile, rare samples of metal particle discharge were generated on a 110kV simulation test bench. Finally, a labeled dataset of three modes and operating conditions was established.

[0009] Furthermore, step S2 is specifically as follows: Based on the dataset from step S1, the mapping relationship between oil temperature, load rate, and noise is first fitted using measured transformer data. The cross-correlation coefficient threshold and wavelet threshold adjustment range are set according to the mapping relationship. Then, the labeled dataset is processed using a denoising process. In the first stage, the cross-correlation matrix of the three-mode signals is calculated. Segments with a current-ultrasonic time difference ≤10μs and a current-infrared hotspot synchronization ≥80% are retained, and interference signals with a deviation exceeding 20% ​​are removed. In the second stage, improved wavelet denoising is applied to the located segments, and the threshold is output by the operating condition and noise model. Based on the fitting formula of load rate and discharge pulse interval, the air gap discharge sample with a load rate of 50% is transformed into an equivalent sample of 60%~100%. Combined with time stretching and additive noise processing, the processed signal data is obtained. The fitting formula for load rate and discharge pulse interval is: 1.2; where, Indicates the discharge pulse interval; This indicates the transformer load rate, reflecting the actual operating load level of the transformer; the constant term 1.2 indicates the reference pulse interval, in milliseconds (ms).

[0010] Furthermore, step S3 is specifically as follows: Based on the signal data processed in step S2, a three-branch feature extraction structure is first designed. The current branch uses a 3-layer 1D-CNN with ReLU activation to extract key electrical parameters including pulse peak value and rising edge slope and output 64-dimensional features. The ultrasonic branch uses a 2-layer BiLSTM to extract acoustic parameters including propagation delay and amplitude attenuation and output 64-dimensional features. The infrared branch uses a 1-layer Transformer to extract thermal parameters including temperature rise gradient and hot spot area and output 64-dimensional features. A fault-oriented fusion mechanism is established, and the weights of the physical constraint layer are set. For metal particle discharge, the corresponding current peak value is 0.9, sound velocity is 0.8, and temperature rise gradient is 0.7. For surface discharge, the corresponding sound attenuation is 0.8, temperature rise area is 0.7, and current rise time is 0.6. The correlation features including current peak value and temperature rise gradient, sound velocity and pulse interval are calculated through the cross-field interaction layer. These features are then concatenated with the three-branch features to form a fusion feature vector, thus establishing a cross-physical field feature fusion module. Finally, the feature mutual information matrix is ​​calculated, and features with mutual information values ​​≥0.7 are retained. The fused feature vector is then output to enhance the feature differentiation capability of different fault types.

[0011] Furthermore, step S4 is specifically as follows: Based on the fusion features of step S3, the ResNet-18 backbone network is first improved by inserting the cross-physics feature fusion module of step S3 into the third bottleneck layer; a working condition adaptation layer is added before the output layer to normalize the four types of working condition parameters collected in step S1 into a 16-dimensional vector, which is then concatenated with the fusion features and input into the classification layer. After the classification layer, a SHAP interpretable module is set up, which is connected to the SHAP TreeExplainer to output the contribution of each feature to the result. A matching threshold is set. If a specific combination of feature contributions is met, it is determined that it conforms to the physical law of metal particle discharge. Otherwise, an early warning is triggered. At the same time, features and fault parameters are automatically associated to generate a source tracing report. A joint loss function is designed, and the Focal Loss algorithm is used to set the loss weight of hard-to-distinguish samples to 2.0 to balance the sample distribution. The working condition adaptation loss calculates the prediction bias of the model under different working conditions and sets the weight to 0.4. The static domain bias loss controls the prediction bias under three voltage scenarios of 110kV, 220kV and 500kV and sets the weight to 0.1 to control the model overfitting. The training process adopts batch size=32 and 100 epochs.

[0012] Furthermore, step S5 is specifically as follows: First, the signal data processed in step S2 is preprocessed and divided into training set, validation set and test set in a ratio of 7:2:1. Z-Score standardization is applied to the three-mode signals and Min-Max standardization is applied to the operating parameters, which are mapped to the 0~1 range. Then, an adaptive training strategy was adopted to optimize the model. The AdamW optimizer was selected, and the learning rate was decayed to 0.7 of the previous round every 5 epochs. A domain adversarial neural network was introduced to treat low load and high load as different domains, and the domain classifier was trained. The trained model is lightweighted by removing channels with a feature contribution of less than 5% through channel pruning. The ResNet-18 trained in step S4 is used as the teacher model, and the student model is distilled using Softmax loss, so that the model can be adapted to the computing power of edge devices, and finally the optimized model is obtained.

[0013] Furthermore, step S6 is specifically as follows: Based on the optimized model in step S5, laboratory verification is first conducted. For accuracy indicators, the accuracy of the test set, the fault recall rate, and the F1 score are evaluated. For operating condition adaptability, the model's ability to resist operating condition drift is verified by simulating dynamic changes in load rate from 30% to 100% and oil temperature from 40℃ to 80℃. For interpretability, the reliability of traceability is evaluated by the matching of SHAP feature contribution with physical laws and the pass rate of expert-reviewed samples. Then, deployment tests were conducted in conjunction with actual on-site scenarios. Two sets of edge devices were deployed in each of the three substations with different voltage levels of 110kV, 220kV and 500kV. The devices were continuously run to collect dynamic on-site operating condition samples to verify whether the model's generalization ability and interpretability meet the requirements of operation and maintenance traceability. During the verification process, if the working condition drift causes the accuracy to drop beyond expectations, return to step S2 to adjust the mapping relationship between the working condition and noise and the wavelet threshold parameters; if the interpretability matching degree does not reach the set threshold, return to step S4 to optimize the weight of the SHAP interpretable module and the physical law matching standard until all indicators meet the standard.

[0014] Furthermore, step S7 is specifically as follows: Based on the lightweight model in step S5 and the technical solution verified in step S6, the edge system is deployed. First, the hardware device is configured, and the lightweight model is packaged into ONNX format and loaded into the industrial-grade edge gateway to control the input latency. The design balances computing power and accuracy. In high computing power scenarios, the cross-physical field feature fusion module in step S3 is activated, while in low computing power scenarios, non-core feature layers, including the cross-field interaction layer, are automatically pruned. During real-time monitoring, the operation and maintenance interface displays a heatmap of fault type, confidence level, and SHAP feature contribution. In the maintenance guidance stage, targeted operation suggestions are output. In the fault response stage, audible and visual alarms are triggered, and fault information and SHAP contribution data are uploaded to the power operation and maintenance platform via the MQTT protocol. An edge-cloud collaborative update mechanism is established, with the edge end uploading no less than 50 new field samples per month. The cloud uses these new samples to train an incremental model and pushes it to the edge gateway for model iteration.

[0015] The beneficial effects of this invention are as follows: 1. This invention establishes a three-modal acquisition architecture encompassing electricity, sound, and heat. It acquires current signals via a high-frequency current sensor, sound signals via an ultrasonic sensor on the tank wall, and heat signals via an infrared thermal imager focusing on easily discharged areas. Simultaneously, it obtains operating parameters such as load rate and oil temperature. A multimodal signal cross-correlation matrix retains effective signal segments while removing interference signals with significant deviations. Different networks are then used to extract electrical, acoustic, and thermal parameter features. The physical constraint layer weights and cross-field interaction layer are combined to calculate correlation features, retaining key features. This effectively reduces the feature overlap between air gap and surface discharge, minimizes high-frequency interference and signal distortion, and improves the accuracy of partial discharge identification.

[0016] 2. This invention expands the sample size and makes the ratio of faulty to normal samples reasonable by using a fitting formula for load rate and discharge pulse interval, combined with time stretching and additive noise processing, covering more than 80% of field operating conditions. In terms of model improvement, a working condition adaptation layer is added to ResNet-18, which processes the working condition parameters and combines them with fused features, enabling the model to learn the mapping relationship between working conditions and features. It also introduces a related network to distinguish different load domains, reducing the interference of dynamic working conditions on classification, and designs a computing power and accuracy balance mechanism to adapt to different equipment. This allows the model to operate stably even when the load rate and oil temperature change significantly, solving the identification deviation caused by pulse interval changes under high load and ensuring the accuracy of field deployment.

[0017] 3. This invention sets up an interpretable SHAP module after the classification layer to output the contribution of each feature to the result. It sets a physical law matching threshold based on relevant standards. If the threshold is not met, an early warning is triggered. At the same time, it automatically associates features with fault parameters to generate a source tracing report and clarifies key fault information. In the operation and maintenance phase, the interface displays a heatmap of fault type, confidence level, and feature contribution. During maintenance, it outputs targeted suggestions based on relevant procedures. Fault information is uploaded to the operation and maintenance platform through a specific protocol. This allows operation and maintenance personnel to clearly understand the judgment basis of the model, accurately locate the defect, avoid blind maintenance, and improve maintenance efficiency. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent identification method for transformer partial discharge based on deep learning according to the present invention; Figure 2 This is a flowchart of the data acquisition and preprocessing process of the present invention; Figure 3 This is a flowchart of the feature extraction and fusion process of the present invention; Figure 4 This is a flowchart illustrating the model optimization and lightweighting process of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figures 1 to 4 As shown in the figure, this invention provides a method for intelligent identification of transformer partial discharge based on deep learning. The specific steps of this method are as follows: Step S1: Collect current signals, sound signals, thermal signals and operating parameters, set dual labels of operating parameters and discharge type for signal segments, generate scarce samples, and finally establish a dataset containing electricity, sound, heat and operating conditions. Step S2: Fit the mapping relationship between oil temperature, load rate and noise, and set a threshold; remove interference signals from the dataset, and use the fitting formula of load rate and discharge pulse interval to make the ratio of faulty to normal samples reach 1:1.1, and obtain the processed signal data. Step S3: Extract the three-branch features of electrical parameters, acoustic parameters and thermal parameters, set the weights of the physical constraint layer, calculate the coupling features and concatenate them with the three-branch features, retain the key features, and finally output the fused feature vector. Step S4: Improve the ResNet-18 backbone network by adding a working condition adaptation layer before the output layer; determine the contribution of output features to the results, correlate features with fault parameters to generate a source tracing report; design a joint loss function to improve the model's classification accuracy and working condition adaptability. Step S5: Preprocess the signal data processed in step S2, and use the AdamW optimizer and the domain adversarial neural network to optimize the model; perform pruning and distillation on the optimized model to adapt it to edge computing power and obtain a lightweight model. Step S6: Verify the model's accuracy, adaptability to operating conditions, and interpretability; then deploy the model on-site to verify its generalization ability; iteratively adjust the model based on the verification results until the indicators meet the requirements. Step S7: Load the model to the edge gateway and design a mechanism to balance computing power and accuracy; it has real-time monitoring, maintenance guidance and fault response functions, and iterates the model through collaborative updates between the edge and the cloud.

[0021] Specifically, step S1 is as follows: A three-modal acquisition architecture (electric, acoustic, and thermal) was established according to power equipment testing standards. A high-frequency current sensor was directly associated with the transformer grounding wire to capture 1ns-level discharge pulses for current signal acquisition. An ultrasonic sensor was placed at three measuring points on the tank wall with a spacing of 1.5m to avoid sound field superposition interference for acoustic signal acquisition. For easily discharged areas, including bushings and the top cover of the tank, thermal signals were acquired using an infrared thermal imager at a frame rate of 30fps. Operating parameters, including load rate (30%~100%), oil temperature (40℃~80℃), voltage level (110kV / 220kV / 500kV), and ambient humidity (30%~90%), were synchronously acquired through the transformer intelligent terminal (IEC61850 protocol) at a sampling interval of 100ms. The high-frequency current sensor is model CT-200M, with a sampling rate of 200MHz and a bandwidth of 1MHz~100MHz; The ultrasonic sensor is model US-65, with a sensitivity of -65dB and a range of 0~100dB. The infrared thermal imager is model IR-640, with a resolution of 640×512 and a temperature measurement range of -20℃ to 150℃. To ensure the accuracy of data labeling, signal segments were extracted at 10ms half-cycle of the power system frequency. Each segment contained 5000-10000 sampling points to match the 200MHz sampling rate. Dual labels were used for operating parameters and discharge types, including five categories of discharge types: normal, air gap, floating potential, surface discharge, and metal particles. The labels were cross-reviewed by two testing engineers. At the same time, rare metal particle discharge samples were generated on a 110kV simulation test bench. The particle diameter ranged from 0.1mm to 1mm, and 500 segments were collected for each diameter. Finally, a labeled dataset of three modes and operating conditions was established, with fault samples accounting for about 35%.

[0022] On the 110kV test bench dataset, the low-load domain samples consist of 1200 sets of data with a load rate of 30%-50%, and the high-load domain samples consist of 1500 sets of data with a load rate of 60%-100%. The training data of the domain classifier are combined according to the source domain (high-load labeled data) and the target domain (low-load unlabeled data). Adversarial learning is achieved through a gradient inversion layer to ensure that the feature distribution difference of the output features of the feature extractor is reduced by more than 40% in different load domains.

[0023] Specifically, step S2 is as follows: Based on the dataset from step S1, the mapping relationship between oil temperature, load rate, and noise is first fitted using measured data from 110kV~500kV transformers. Within the range of 40℃~80℃, the ultrasonic noise amplitude shows a regular change for every 10℃ increase in oil temperature, due to the change in sound propagation loss caused by the decrease in oil viscosity. Similarly, the background noise amplitude of the current signal shows a regular change for every 20% increase in load rate. Based on the mapping relationship, a cross-correlation coefficient threshold is set: ≤10μs at 40℃ and ≤8μs at 80℃. A wavelet threshold adjustment range is also set. For low load (30%~50%), the threshold is 0.1~0.2; for high load (80%~100%), the threshold is 0.4~0.5. Then, the labeled data is processed using a denoising process. In the first stage, the cross-correlation matrix of the three-mode signals is calculated. Segments with a current-ultrasound time difference ≤10μs and a current-infrared hotspot synchronization ≥80% are retained, and interference signals with a deviation of more than 20% are removed. In the second stage, the localized segments are denoised using an improved wavelet denoising method, and the threshold is output by the working condition and noise model. The denoising effect is verified by comparing with the standard discharge signal. To address the issue of insufficient coverage of fault samples and operating conditions in step S1, based on the fitting formula between load rate and discharge pulse interval, the air gap discharge samples with a load rate of 50% are transformed into equivalent samples with a load rate of 60% to 100%. Combined with time stretching at 0.8 to 1.2 times the speed and additive noise processing with a signal-to-noise ratio of 20 dB to 30 dB, the ratio of fault samples to normal samples is finally reduced to approximately 1:1.1, covering more than 80% of the field operating conditions. Specifically, this includes 110kV load rates of 30% to 100%, 220kV load rates of 40% to 90%, and 500kV load rates of 50% to 100%, as well as temperature and humidity ranges of 30% to 90%RH and oil temperatures of 40℃ to 80℃, covering a total of 24 core operating condition combinations, resulting in denoised signal data with comprehensive operating condition coverage.

[0024] Fitting formula for load rate and discharge pulse interval: 1.2 In the formula: It represents the discharge pulse interval, measured in milliseconds (ms). It refers to the time interval between two adjacent partial discharge pulse signals and is a key timing characteristic for distinguishing different discharge types (such as air gap discharge and metal particle discharge). This indicates the transformer load rate, expressed as a percentage, ranging from 30% to 100%. It is collected in real time through the transformer's intelligent terminal, reflecting the actual operating load level of the transformer. The coefficient 0.02 represents the influence coefficient of the load rate on the pulse interval, which is obtained by fitting measured data of 110kV~500kV transformers. It means that for every 1% increase in load rate, the discharge pulse interval increases by an average of 0.02ms. The constant term 1.2 represents the reference pulse interval, in milliseconds (ms).

[0025] Formula for cross-correlation coefficient of multimodal signals:

[0026] In the formula: The cross-correlation coefficient represents the cross-correlation coefficient between two modal signals, with a value range of [-1, 1]. It is used to determine the degree of linear correlation between the signals. A value of 0.8 indicates a strong positive correlation, which is determined to be a genuine discharge signal. A value of 0.5 indicates a weak correlation, which is considered an interference signal. 、 This represents the two-mode signal sequence to be calculated, such as... It is a current signal sequence. It is an ultrasonic signal sequence, with a length of [missing information]. The number of sampling points corresponding to a 10ms signal segment is 5000~10000; express and The covariance, expressed as the product of the units of the two signals, such as mV of current signal × dB of ultrasonic signal, reflects the overall correlation trend between the two signals. This represents the mathematical expectation operator, used to calculate the mean of the expression within the parentheses; 、 They represent signals respectively. 、 The mean value, with units consistent with the corresponding signal; 、 They represent signals respectively. 、 The standard deviation, with units consistent with the corresponding signal, reflects the degree of dispersion of the signal.

[0027] Specifically, step S3 is as follows: To extract features that effectively distinguish fault types from the signal data processed in step S2, a three-branch feature extraction structure is first designed. The current branch uses a 3-layer 1D-CNN (3×1, 5×1, 7×1 kernels, stride 1) with ReLU activation to extract 12 key electrical parameters such as pulse peak value and rising edge slope, and outputs 64-dimensional features. The ultrasonic branch uses a 2-layer BiLSTM (128-dimensional hidden layer, dropout=0.2, sequence length 2000) to extract 8 acoustic parameters such as propagation delay and amplitude attenuation, and outputs 64-dimensional features. The infrared branch uses a 1-layer Transformer (4 attention heads, 64 feature dimensions, input is a 16×16 hotspot matrix) to extract 6 thermal parameters such as temperature rise gradient and hotspot area, and outputs 64-dimensional features. Considering the differences in physical characteristics among different fault types, a fault-oriented fusion mechanism is established based on the IEC 60270 standard. Weights are set for the physical constraint layer: peak current of metal particle discharge corresponds to 0.9, sound velocity to 0.8, and temperature rise gradient to 0.7; surface discharge corresponds to sound attenuation to 0.8, temperature rise area to 0.7, and current rise time to 0.6. A cross-field interaction layer calculates the correlation features, including peak current and temperature rise gradient, and sound velocity and pulse interval, totaling 128 dimensions. These are then concatenated with the three-branch features to form a 320-dimensional fusion feature vector, establishing a cross-physical field feature fusion module. Finally, the feature mutual information matrix is ​​calculated, retaining features with mutual information values ​​≥ 0.7, such as peak current minus infrared hotspot area, and outputting the fusion feature vector to enhance the feature differentiation capability for different fault types.

[0028] Specifically, step S4 is as follows: Based on the fusion features of step S3, the ResNet-18 backbone network is first improved by inserting the cross-physics feature fusion module of step S3 into the third bottleneck layer to replace the original residual connection and enhance the mid-to-high layer feature expression. Considering the influence of transformer dynamic operating conditions on the classification results, an operating condition adaptation layer is added before the output layer. The four operating condition parameters of load rate, oil temperature, voltage level and ambient humidity collected in step S1 are normalized into 16-dimensional vectors, concatenated with the fusion features and input into the classification layer. The classification layer adopts a fully connected layer 528→256→5 and outputs the probability of 5 types of discharge. After the classification layer, a SHAP interpretable module is set up, which is connected to the SHAP TreeExplainer to output the contribution of each feature to the result, such as the contribution of the peak current greater than 600mV in metal particle discharge. A matching degree threshold is set, and it is calibrated with 1000 expert-annotated samples. If it meets the specific feature contribution combination, such as the peak current contribution ≥60% and the sound velocity contribution ≥50%, it is determined to meet the physical law of metal particle discharge. Otherwise, an early warning is triggered. At the same time, the feature and fault parameters are automatically associated to generate a source tracing report, such as the peak current of 650mV corresponding to a particle diameter of 0.6mm and the temperature rise gradient of 5℃ / mm corresponding to an air gap thickness of 0.2mm. To improve the model's classification accuracy and adaptability to different operating conditions, a joint loss function was designed. The Focal Loss algorithm was used, with the loss weight for difficult-to-classify samples set to 2.0. This was used to balance the sample distribution for samples confused by metal particles / surface discharge. The operating condition adaptation loss calculated the prediction bias of the model under different operating conditions, with a weight set to 0.4. The load rate bias weight was 0.2, the oil temperature bias weight was 0.15, and the voltage bias weight was 0.05. The static domain bias loss controlled the prediction bias under three voltage scenarios (110kV, 220kV, and 500kV) with a weight set to 0.1 to control overfitting. The training process used a batch size of 32 and 100 epochs to further ensure the model's convergence stability.

[0029] Joint loss function formula:

[0030]

[0031]

[0032]

[0033] In the formula: This represents the total loss function, which comprehensively measures the model classification error, operating condition adaptation error, and voltage level deviation. 、 、 These represent the weighting coefficients for each loss term, taken as 1.0, 0.4, and 0.1 respectively. =0.4 corresponds to the working condition adaptation loss weight. =0.1 corresponds to the static domain bias loss weight; This represents the focusing loss, used to address the imbalanced sample problem. When the weight is 2.0, it is difficult to distinguish sample weights, especially for samples confused by metal particles / surface discharge. When the value is 2, focus on the parameter and reduce the proportion of easily distinguishable samples lost. For the model to the first The predicted probability of each sample, with a value ranging from 0 to 1; This represents the adaptation loss under different operating conditions, quantifying the prediction bias under dynamic operating conditions. =8 represents the number of operating condition ranges, divided into 10% increments based on load rate (30%~100%) and 10°C increments based on oil temperature (40°C~80°C). For the first The weights for each operating condition interval are allocated according to the proportion of samples in that interval. For the first The average prediction deviation for each operating condition range, expressed in % This represents the static domain bias loss, used to control prediction bias at different voltage levels. =3, Voltage levels: 110kV, 220kV, 500kV For the first Each voltage level is weighted at 1 / 3 to ensure that each voltage level is equally important. For the first Prediction error rate for each voltage level, in units of %.

[0034] Specifically, step S5 is as follows: To improve the adaptability and deployment efficiency of the interpretable dynamic classification model in step S4 under actual dynamic working conditions, the signal data processed in step S2 is first preprocessed and divided into training set, validation set and test set in a ratio of 7:2:1. The test set includes 20% of the working conditions that were not trained, such as 110kV load rate of 30% and 500kV oil temperature of 80%. The three-mode signals are standardized using Z-Score with a mean of 0 and a standard deviation of 1. The working condition parameters are standardized using Min-Max and mapped to the 0~1 interval. Then, an adaptive training strategy was adopted to optimize the model. The AdamW optimizer was selected (initial learning rate 1e-3, weight decay 1e-4). Every 5 epochs, the learning rate was decayed to 0.7 of the previous round (adjusted based on the validation set loss curve). A domain adversarial neural network was introduced to treat low load (30%~50%), medium load (50%~80%, classified as low load domain for auxiliary training), and high load (80%~100%) as two domains (low load class: 30%~80%, high load class: 80%~100%). The domain classifier was trained (loss weight 0.3) so that the model learns working condition-independent features and reduces the impact of dynamic working conditions on classification accuracy. The domain adversarial neural network includes a feature extractor, a label predictor, and a domain classifier. The feature extractor reuses the parameters of the first 16 layers of the improved ResNet-18 network. The domain classifier uses a 2-layer fully connected network, with the input being the 64-dimensional features output by the feature extractor. The first layer has 128 neurons, and the second layer has 2 neurons, used to distinguish between two domains with low load (≤50%) and high load (>50%). The objective function for adversarial training is: ,in Cross-entropy classification loss, For domain classification loss, To balance the parameters, the initial value is 0.01, and it decreases by 0.9 every 10 rounds of training; During training, the parameters of the feature extractor and label predictor are fixed to update the domain classifier, and then the parameters of the domain classifier are fixed to update the feature extractor and label predictor. This process is repeated until the loss function converges (≤0.001).

[0035] Considering the computing power limitations of edge devices, lightweight processing is implemented on the trained model. Channel pruning is performed on the residual blocks of layers 3-6 of ResNet-18, removing channels with feature contribution of less than 5%, with a pruning ratio of 40% to 50%. The ResNet-18 trained in step S4 is used as the teacher model, and the student model is distilled using Softmax loss with a temperature coefficient T=5, thereby adapting the model to the computing power of edge devices and finally obtaining the optimized model.

[0036] Z-Score Standardization Formula:

[0037] In the formula: It represents the standardized three-modal signal values, applicable to current signals, ultrasonic signals, and infrared thermal signals. It maps the original signal to a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of signal dimension differences on model training. The original sampled values ​​of the three-mode signals are represented, such as current signals in mV, ultrasonic signals in dB, and infrared signals in °C. This represents the mean of all sampled points of a signal of a certain mode. The calculation range is all samples of that mode in the labeled dataset, and the unit is the same as that of the original signal. Consistent; The standard deviation represents the total number of sampling points of a given modal signal, reflecting the dispersion of the original signal. Its unit is the standard deviation of the original signal. Consistency is required, and outliers, such as interference signals exceeding 3 times the standard deviation, must be excluded during calculation.

[0038] Min-Max normalization formula:

[0039] In the formula: This represents the standardized operating condition parameter value, ranging from 0 to 1. It is used to uniformly map operating condition parameters of different magnitudes to the same interval, making it easier to combine with fused features and input them into the model. The raw values ​​of the operating parameters are represented, including load rate (%), oil temperature (°C), and ambient humidity (%). Voltage level (kV) needs to be coded separately by level before standardization. This represents the minimum value of a parameter under a certain operating condition, obtained from statistical analysis of operating condition data, such as oil temperature. =40℃, load rate =30%, unit and original parameter Consistent; This represents the maximum value of a certain operating condition parameter, obtained from the data collected in step S1, such as oil temperature. =80℃, load rate =100%, unit and original parameter Consistent.

[0040] Specifically, step S6 is as follows: To ensure that the optimized model in step S5 meets the needs of on-site operation and maintenance, laboratory validation is first conducted. For accuracy indicators, the accuracy of the test set, fault recall rate, and F1 score are evaluated, with a focus on the recall effect of easily confused faults such as metal particle discharge and surface discharge. For operating condition adaptability, the model's resistance to operating condition drift is verified by simulating dynamic changes in load rate from 30% to 100% and oil temperature from 40℃ to 80℃. For interpretability, the reliability of traceability is evaluated by the matching of SHAP feature contribution with physical laws and the pass rate of expert-reviewed samples. Then, deployment tests were carried out in combination with actual on-site scenarios. Two sets of edge devices were deployed in each of the three different voltage level substations of 110kV, 220kV and 500kV to continuously collect dynamic on-site operating condition samples, including rare fault samples such as 0.1mm metal particle discharge, to verify whether the model's generalization ability and interpretability meet the requirements of operation and maintenance traceability. During the verification process, if the working condition drift causes the accuracy to drop beyond expectations, return to step S2 to adjust the mapping relationship between the working condition and noise and the wavelet threshold parameters; if the interpretability matching degree does not reach the set threshold, return to step S4 to optimize the weight of the SHAP interpretable module and the physical law matching standard until all indicators meet the standard.

[0041] Specifically, step S7 is as follows: Based on the lightweight model in step S5 and the technical solution verified in step S6, the edge system is deployed. First, the hardware device is configured, and the lightweight model is packaged into ONNX format and loaded onto the industrial-grade edge gateway. It supports 4G / 5G and Ethernet communication, controls input latency, and includes signal preprocessing and model inference time. The industrial-grade edge gateway model is EG-500, with a computing power of 5 TOPS, 4GB of memory, and 32GB of storage. Considering the differences in computing power of edge devices in different substations, a computing power and accuracy balance mechanism is designed. In high computing power scenarios (≥5TOPS), the cross-physical field feature fusion module in step S3 is activated, and in low computing power scenarios (3TOPS), non-core feature layers such as cross-field interaction layers are automatically pruned. Finally, the system achieves full-process operation and maintenance functionality. During real-time monitoring, the operation and maintenance interface displays the fault type, confidence level (≥90%), and a heatmap of SHAP feature contribution, such as metal particle discharge, where the current peak contribution is relatively high. In the maintenance guidance stage, it outputs targeted operation suggestions in conjunction with the "Transformer Partial Discharge Maintenance Procedures," such as prioritizing the inspection of the tank top cover insulation when the infrared hotspot temperature is high and the contribution is high. In the fault response stage, it triggers audible and visual alarms and uploads fault information and SHAP contribution data to the power operation and maintenance platform via the MQTT protocol. To ensure the long-term adaptability of the model, a collaborative update mechanism between the edge and the cloud is established. The edge uploads no less than 50 new field samples per month, and the cloud trains an incremental model based on the new samples. When the accuracy of the incremental model on the validation set improves by ≥2% compared to the current edge model, it is pushed to the edge gateway via 4G / 5G. If the improvement threshold is not reached, the samples are accumulated until the next (2 months) for retraining and evaluation, which is used for model iteration.

[0042] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of partial discharge in transformers based on deep learning, characterized in that: The specific steps of this method are as follows: Step S1: Collect current signals, sound signals, thermal signals and operating parameters, set dual labels of operating parameters and discharge type for signal segments, generate scarce samples, and finally establish a dataset containing electricity, sound, heat and operating conditions. Step S2: Fit the mapping relationship between oil temperature, load rate and noise, and set a threshold; By removing interference signals from the dataset, and using the fitting formula of load rate and discharge pulse interval, the ratio of faulty to normal samples is made to 1:1.1, thus obtaining the processed signal data. Step S3: Extract the three-branch features of electrical parameters, acoustic parameters and thermal parameters, set the weights of the physical constraint layer, calculate the coupling features and concatenate them with the three-branch features, retain the key features, and finally output the fused feature vector. Step S4: Improve the ResNet-18 backbone network by adding a working condition adaptation layer before the output layer; determine the contribution of output features to the results, correlate features with fault parameters to generate a source tracing report; design a joint loss function to improve the model's classification accuracy and working condition adaptability. Step S5: Preprocess the signal data processed in step S2, and use the AdamW optimizer and the domain adversarial neural network to optimize the model; perform pruning and distillation on the optimized model to adapt it to edge computing power and obtain a lightweight model. Step S6: Verify the model's accuracy, adaptability to operating conditions, and interpretability; then deploy the model on-site to verify its generalization ability; iteratively adjust the model based on the verification results until the indicators meet the requirements. Step S7: Load the model to the edge gateway and design a mechanism to balance computing power and accuracy; it has real-time monitoring, maintenance guidance and fault response functions, and iterates the model through collaborative updates between the edge and the cloud.

2. The intelligent identification method for transformer partial discharge based on deep learning according to claim 1, characterized in that: The specific steps of S1 are as follows: A three-modal acquisition architecture of electrical, acoustic, and thermal is established. A high-frequency current sensor is directly connected to the transformer grounding wire to collect current signals. An ultrasonic sensor is placed on the tank wall to collect acoustic signals. For easily discharged areas, including bushings and the top cover of the tank, thermal signals are collected by an infrared thermal imager. Operating parameters, such as load factor, oil temperature, voltage level, and ambient humidity, are collected synchronously through the transformer's intelligent terminal. Signal segments were extracted based on the half-cycle of the power system's power frequency. The operating parameters and discharge type were dual-labeled and cross-reviewed by two testing engineers. Meanwhile, rare samples of metal particle discharge were generated on a 110kV simulation test bench. Finally, a labeled dataset of three modes and operating conditions was established.

3. The intelligent identification method for transformer partial discharge based on deep learning according to claim 2, characterized in that: Step S2 is as follows: Based on the dataset from step S1, the mapping relationship between oil temperature, load rate, and noise is first fitted using measured transformer data. The cross-correlation coefficient threshold and wavelet threshold adjustment range are set according to the mapping relationship. Then, the labeled dataset is processed using a denoising process. In the first stage, the cross-correlation matrix of the three-mode signals is calculated. Segments with a current-ultrasonic time difference ≤10μs and a current-infrared hotspot synchronization ≥80% are retained, and interference signals with a deviation exceeding 20% ​​are removed. In the second stage, improved wavelet denoising is applied to the located segments, and the threshold is output by the operating condition and noise model. Based on the fitting formula of load rate and discharge pulse interval, the air gap discharge sample with a load rate of 50% is transformed into an equivalent sample of 60%~100%. Combined with time stretching and additive noise processing, the processed signal data is obtained. The fitting formula for load rate and discharge pulse interval is: 1.2; where, Indicates the discharge pulse interval; This indicates the transformer load rate, reflecting the actual operating load level of the transformer; the constant term 1.2 indicates the reference pulse interval, in milliseconds (ms).

4. The intelligent identification method for transformer partial discharge based on deep learning according to claim 3, characterized in that: Step S3 is as follows: Based on the signal data processed in step S2, a three-branch feature extraction structure is first designed. The current branch adopts a 3-layer 1D-CNN with ReLU activation to extract key electrical parameters including pulse peak value and rising edge slope and output 64-dimensional features. The ultrasonic branch uses a 2-layer BiLSTM to extract acoustic parameters including propagation delay and amplitude attenuation and output 64-dimensional features; The infrared branch uses a single Transformer layer to extract thermal parameters including temperature rise gradient and hot spot area and output 64-dimensional features. A fault-oriented fusion mechanism is established, and the weights of the physical constraint layer are set. For metal particle discharge, the corresponding current peak value is 0.9, sound velocity is 0.8, and temperature rise gradient is 0.

7. For surface discharge, the corresponding sound attenuation is 0.8, temperature rise area is 0.7, and current rise time is 0.

6. The correlation features including current peak value and temperature rise gradient, sound velocity and pulse interval are calculated through the cross-field interaction layer. These features are then concatenated with the three-branch features to form a fusion feature vector, thus establishing a cross-physical field feature fusion module. Finally, the feature mutual information matrix is ​​calculated, and features with mutual information values ​​≥0.7 are retained. The fused feature vector is then output to enhance the feature differentiation capability of different fault types.

5. The intelligent identification method for transformer partial discharge based on deep learning according to claim 4, characterized in that: Step S4 is as follows: Based on the fusion features of step S3, the ResNet-18 backbone network is first improved by inserting the cross-physics feature fusion module of step S3 into the third bottleneck layer. Add a working condition adaptation layer before the output layer, normalize the four types of working condition parameters collected in step S1 into a 16-dimensional vector, and input it into the classification layer after concatenating it with the fusion features. After the classification layer, a SHAP interpretable module is set up, which is connected to the SHAP TreeExplainer to output the contribution of each feature to the result. A matching threshold is set. If a specific combination of feature contributions is met, it is determined that it conforms to the physical law of metal particle discharge. Otherwise, an early warning is triggered. At the same time, features and fault parameters are automatically associated to generate a source tracing report. A joint loss function is designed, and the Focal Loss algorithm is used to set the loss weight of hard-to-distinguish samples to 2.0 to balance the sample distribution. The working condition adaptation loss calculates the prediction bias of the model under different working conditions and sets the weight to 0.

4. The static domain bias loss controls the prediction bias under three voltage scenarios of 110kV, 220kV, and 500kV and sets the weight to 0.1 to control the model overfitting. The training process adopts batch size=32 and 100 epochs.

6. The intelligent identification method for transformer partial discharge based on deep learning according to claim 5, characterized in that: Step S5 is as follows: The signal data processed in step S2 is preprocessed and divided into training set, validation set and test set in a ratio of 7:2:

1. The three-mode signals are standardized by Z-Score and the operating parameters are standardized by Min-Max, and mapped to the 0~1 range. Then, an adaptive training strategy was adopted to optimize the model. The AdamW optimizer was selected, and the learning rate was decayed to 0.7 of the previous round every 5 epochs. A domain adversarial neural network was introduced to treat low load and high load as different domains, and the domain classifier was trained. The trained model is lightweighted by removing channels with a feature contribution of less than 5% through channel pruning. The ResNet-18 trained in step S4 is used as the teacher model, and the student model is distilled using Softmax loss, so that the model can be adapted to the computing power of edge devices, and finally the optimized model is obtained.

7. The intelligent identification method for transformer partial discharge based on deep learning according to claim 6, characterized in that: Step S6 is as follows: Based on the optimized model in step S5, laboratory verification is first conducted. For accuracy indicators, the accuracy of the test set, the fault recall rate, and the F1 score are evaluated. For operating condition adaptability, the model's ability to resist operating condition drift is verified by simulating dynamic changes in load rate from 30% to 100% and oil temperature from 40℃ to 80℃. For interpretability, the reliability of traceability is evaluated by the matching of SHAP feature contribution with physical laws and the pass rate of expert-reviewed samples. Then, deployment tests were conducted in conjunction with actual on-site scenarios. Two sets of edge devices were deployed in each of the three substations with different voltage levels of 110kV, 220kV and 500kV. The devices were continuously run to collect dynamic on-site operating condition samples to verify whether the model's generalization ability and interpretability meet the requirements of operation and maintenance traceability. During the verification process, if the working condition drift causes the accuracy to drop beyond expectations, return to step S2 to adjust the mapping relationship between the working condition and noise and the wavelet threshold parameters; if the interpretability matching degree does not reach the set threshold, return to step S4 to optimize the weight of the SHAP interpretable module and the physical law matching standard until all indicators meet the standard.

8. The intelligent identification method for transformer partial discharge based on deep learning according to claim 7, characterized in that: Step S7 is as follows: Based on the lightweight model in step S5 and the technical solution verified in step S6, the edge system is deployed. First, the hardware device is configured, and the lightweight model is packaged into ONNX format and loaded into the industrial-grade edge gateway to control the input latency. The design balances computing power and accuracy. In high computing power scenarios, the cross-physical field feature fusion module in step S3 is activated, while in low computing power scenarios, non-core feature layers, including the cross-field interaction layer, are automatically pruned. During real-time monitoring, the operation and maintenance interface displays a heatmap of fault type, confidence level, and SHAP feature contribution. In the maintenance guidance phase, targeted operational suggestions are provided; In the fault response phase, an audible and visual alarm is triggered, and the fault information and SHAP contribution data are uploaded to the power operation and maintenance platform via the MQTT protocol. Establish a collaborative update mechanism between the edge and the cloud. The edge device uploads no less than 50 new on-site samples every month, and the cloud trains an incremental model based on the new samples and pushes it to the edge gateway for model iteration.

Images