Power transformer discharge fault identification and severity grading method based on wavelet denoising and SE attention mechanism

Abstract

The invention discloses a power transformer discharge fault identification and severity grading method based on wavelet denoising and an SE attention mechanism, and belongs to the technical field of intelligent diagnosis, operation and maintenance of power equipment. Comprising the steps of collecting partial discharge signals in real time through a sensor network arranged at a key position of a power transformer, preprocessing the collected partial discharge signals, constructing a data set, constructing a fault recognition model based on an attention-enhanced lightweight convolutional neural network, training the fault recognition model based on the data set, and obtaining the fault recognition result of the power transformer. And inputting a to-be-detected partial discharge signal into the trained fault recognition model, outputting a fault recognition result and severity level, and generating a differentiated operation and maintenance suggestion based on the severity level. According to the method, accurate identification and severity automatic grading of the transformer discharge fault are realized through wavelet denoising and an attention mechanism, the diagnosis reliability in a complex electromagnetic environment is remarkably improved, and efficient support is provided for operation and maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis and operation and maintenance technology for power equipment, specifically to a method for identifying and classifying the severity of discharge faults in power transformers based on wavelet denoising and SE attention mechanism. Background Technology

[0002] As a core piece of equipment in the power transmission and transformation chain, the insulation condition of power transformers directly determines the safety of power grid operation and the reliability of power supply. With the development of power grids towards ultra-high voltage, large capacity, and a high proportion of renewable energy integration, transformer operating conditions are becoming increasingly complex. Partial discharge caused by insulation aging has become a major cause of equipment failure. Timely and accurate identification of partial discharge fault types and assessment of severity are crucial for avoiding insulation failure and reducing power outage losses. Current transformer fault diagnosis technologies encompass various methods such as infrared thermal imaging, ultrasonic testing, and ultraviolet imaging. Among these, partial discharge detection has become the mainstream technology due to its ability to directly reflect the evolution process of insulation defects.

[0003] Currently, existing detection methods have the following limitations:

[0004] Weak anti-interference capability of signal acquisition: The partial discharge signal acquired on site is easily affected by electromagnetic interference and harmonic noise, resulting in a low signal-to-noise ratio. Traditional acquisition systems do not have a specific noise reduction mechanism designed, which leads to a decrease in the accuracy of subsequent feature extraction and an increase in the fault misjudgment rate.

[0005] Multimodal data fusion is crude: existing multimodal solutions mostly use simple channel splicing to fuse PRPD (phase-resolved partial discharge) and PRPS (phase-resolved pulse sequence) data, without considering the feature correlation between modes, failing to fully explore complementary information, and limiting generalization ability;

[0006] Poor model deployment adaptability: Traditional deep learning models have a large parameter scale, which requires high computing power from edge operation and maintenance terminals, making it difficult to meet the needs of real-time on-site diagnosis, and they are prone to overfitting in small sample scenarios;

[0007] The diagnostic results are not practical enough: they can only identify the type of fault, but do not combine discharge characteristics to classify the severity of the fault, and cannot provide maintenance personnel with maintenance priority suggestions, resulting in low efficiency of maintenance decision-making.

[0008] Weak adaptability to renewable energy scenarios: The impact of harmonic interference and load fluctuations caused by renewable energy grid connection is not considered. Fault identification accuracy decreases significantly under extreme operating conditions, making it difficult to adapt to the needs of a high proportion of renewable energy grids.

[0009] Furthermore, technical solutions similar to existing technologies also have limitations to varying degrees:

[0010] 1. Traditional multimodal deep learning model: It adopts a three-channel input method that directly concatenates PRPD and PRPS data, and combines basic convolutional neural networks to realize fault identification. However, it does not perform signal denoising preprocessing and does not have an attention mechanism to optimize feature weights. Under complex interference scenarios, the recognition accuracy is less than 90%, and it cannot classify the severity of faults.

[0011] 2. Lightweight Fault Identification System: Based on a simplified convolutional network structure to reduce the model parameter scale, but it only uses single PRPD modal data, does not integrate PRPS temporal features, has weak generalization ability, and is not adapted to the low computing power environment of edge devices, with inference time exceeding 100ms.

[0012] 3. Partial Discharge Detection and Diagnosis Platform: It collects discharge signals through high-frequency sensors and uses an oscilloscope and partial discharge instrument to display data and perform simple feature analysis. However, manual judgment of the fault type is required afterward. It lacks automatic identification and classification functions, resulting in low efficiency.

[0013] 4. Fault type and severity correlation method: The correlation between discharge amplitude, frequency and fault severity is established based on expert experience. This method is highly subjective, with a classification deviation of more than 15% in different scenarios. Furthermore, it does not combine deep learning models to achieve automated classification, which limits its practicality.

[0014] To address the aforementioned issues, there is an urgent need for a power transformer discharge fault identification and severity classification method based on wavelet denoising and SE attention mechanism, which can solve the problems existing in traditional methods. Summary of the Invention

[0015] The purpose of this invention is to provide a method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism. Through wavelet denoising and attention mechanism, the method achieves accurate identification and automatic severity classification of transformer discharge faults, significantly improves the diagnostic reliability in complex electromagnetic environments, and provides efficient support for operation and maintenance decisions.

[0016] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0017] A method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism includes:

[0018] Step 1: Collect partial discharge signals in real time through a sensor network deployed at key locations on the power transformer;

[0019] Step 2: Preprocess the collected partial discharge signals to construct a dataset;

[0020] Step 3: Construct a fault identification model based on an attention-enhanced lightweight convolutional neural network;

[0021] Step 4: Train the fault identification model based on the dataset;

[0022] Step 5: Input the partial discharge signal to be detected into the trained fault identification model, and output the fault identification result and severity classification;

[0023] Step 6: Generate differentiated operation and maintenance recommendations based on severity level.

[0024] Furthermore, in step 1, partial discharge signals are collected in real time through a sensor network deployed at key locations on the power transformer, specifically as follows:

[0025] A high-frequency pulse current sensor is installed at the bottom of the transformer tank where the power is connected. The current sensor transmits the signal to the partial discharge instrument via a shielded twisted pair cable, and transmits the signal along with the synchronous voltage signal to an oscilloscope to acquire the partial discharge signal.

[0026] Furthermore, the partial discharge signal includes PRPD data and PRPS data.

[0027] Furthermore, in step 2, the acquired partial discharge signal is preprocessed, specifically as follows:

[0028] The acquired partial discharge signals were sequentially subjected to wavelet threshold denoising, data standardization, and adaptive weighted fusion.

[0029] Furthermore, the fault identification model includes an input module, a feature extraction module, an attention optimization module, a feature dimensionality reduction and encoding module, an overfitting suppression module, and a classification and grading module.

[0030] Furthermore, the feature extraction module includes three convolutional units, each of which consists of a convolutional layer, a batch normalization layer, and a ReLU activation layer.

[0031] Furthermore, the convolutional layer has 3×3 convolutional kernels, with the number of kernels being 16, 32, and 64 respectively, a stride of 1, and a padding method of "same".

[0032] Furthermore, the attention optimization module adopts the SE channel attention mechanism, which dynamically adjusts the feature channel weights through squeeze-excitation operations.

[0033] Furthermore, the feature dimensionality reduction and encoding module includes two hybrid pooling layers and two fully connected layers.

[0034] Furthermore, the classification and grading module adopts a dual-output structure, including a Softmax layer and a Sigmoid layer. The Softmax layer is used to output the probability of the fault type, and the Sigmoid layer is used to output the severity score.

[0035] In summary, the present invention has at least one of the following beneficial technical effects:

[0036] 1. Dual improvement in signal quality and feature extraction accuracy: Through wavelet threshold denoising preprocessing (db4 wavelet basis + 3-level decomposition), the signal-to-noise ratio of the discharge signal is improved from 20dB to over 35dB, the invalid interference information removal rate exceeds 80%, the proportion of effective information during feature extraction is increased by 15%, and the fault identification accuracy on the test set reaches 94.2%, which is 4%-6% higher than the existing no-denoising scheme;

[0037] 2. Significantly optimized multimodal fusion efficiency: Adaptive weighted fusion (α=0.52) is used to replace simple channel splicing, dynamically balancing the feature weights of PRPD and PRPS, improving the utilization rate of intermodal correlation by 20%, enhancing generalization ability, and reducing accuracy fluctuation from 8% to below 2.5% in complex interference scenarios;

[0038] 3. Excellent model performance and deployment adaptability: The introduction of a channel attention mechanism focuses on key features, reducing the risk of overfitting by 30% in small sample scenarios (500 samples), and the accuracy fluctuation of the test set is ≤2%; at the same time, the network parameter scale is optimized, and the inference time is 78ms, which is 22% shorter than the existing lightweight model, and can be deployed on edge operation and maintenance terminals.

[0039] 4. Enhanced functionality to support operational and maintenance decisions: The new fault severity classification function has been added, with a classification accuracy rate of 92.5%. Based on the classification results, targeted operation and maintenance suggestions are output (such as immediate shutdown for severe faults and monthly monitoring for moderate faults), reducing fault response time by 30% and preventing the escalation of incidents.

[0040] 5. Strong adaptability and practicality: The addition of a new energy harmonic correction module improves accuracy by 3% under extreme harmonic conditions, making it suitable for high-proportion new energy power grids; at the same time, 80 sets of field measurement data are added for verification, with the accuracy deviation between the laboratory and the field being ≤0.4%, significantly improving its engineering practicality. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0042] Figure 2 This is a schematic diagram of the anti-interference fault signal acquisition and preprocessing process;

[0043] Figure 3 A schematic diagram of an attention-enhanced lightweight convolutional neural network structure;

[0044] Figure 4 This is a schematic diagram of the fault identification and severity classification decision-making process. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0046] like Figure 1 As shown, this invention provides a method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism, including:

[0047] Step 1: Collect partial discharge signals in real time through a sensor network deployed at key locations on the power transformer;

[0048] Step 2: Preprocess the collected partial discharge signals to construct a dataset;

[0049] Step 3: Construct a fault identification model based on an attention-enhanced lightweight convolutional neural network;

[0050] Step 4: Train the fault identification model based on the dataset;

[0051] Step 5: Input the partial discharge signal to be detected into the trained fault identification model, and output the fault identification result and severity classification;

[0052] Step 6: Generate differentiated operation and maintenance recommendations based on severity level.

[0053] In step 1, partial discharge signals are collected in real time through a sensor network deployed at key locations on the power transformer. Specifically:

[0054] This application adopts an integrated data acquisition system of "sensor-anti-interference transmission-data storage", the core components of which include:

[0055] 1. High-frequency pulse current sensor: The HFCT-30MHz model is used and is deployed at the bottom grounding wire of the transformer tank (5cm away from the grounding terminal). It couples 3-30MHz high-frequency pulse current signals, with an acquisition accuracy of ±0.08μL / L, and supports the acquisition of four types of typical discharge fault signals.

[0056] 2. Anti-interference transmission link: Shielded twisted pair cable (95% shielding coverage) is used to transmit signals to reduce the impact of electromagnetic interference. At the same time, a 50Ω matching resistor is connected in series at the sensor output to reduce signal reflection.

[0057] 3. Dual-channel data acquisition: One signal is transmitted to the improved partial discharge instrument (with a newly added signal amplification module and a gain of 20dB) for spectrum data acquisition, and the other signal is transmitted together with the synchronization voltage signal (sampling frequency 1MHz) to the oscilloscope to realize the synchronous acquisition of real-time waveform and synchronization voltage signal with a sampling interval of 0.8ms. Finally, the partial discharge signal, namely PRPD data and PRPS data, is obtained.

[0058] Among them, the test environment is controlled as follows: the laboratory simulates an ambient temperature of 20-23℃ and a humidity of 50%-60%. During on-site data collection, the environmental influence is corrected through a temperature compensation module (compensation range -10-40℃) to ensure data stability.

[0059] In step 2, the acquired partial discharge signals are preprocessed, specifically as follows:

[0060] The acquired partial discharge signals were sequentially subjected to wavelet threshold denoising, data standardization, and adaptive weighted fusion processing, which are described in detail below:

[0061] Wavelet threshold denoising: The acquired partial discharge signal is decomposed into three layers using the db4 wavelet basis. The high-frequency detail coefficients are processed by hard thresholding (threshold λ=1.2σ, ​​where σ is the noise standard deviation). The denoised partial discharge signal is reconstructed, which improves the signal-to-noise ratio to more than 35dB and eliminates high-frequency interference.

[0062] Modal data standardization: The PRPD and PRPS data in the denoised partial discharge signal are standardized separately, and the process is described in detail below:

[0063] PRPD data: Construct a two-dimensional matrix Y1 (horizontal axis phase 0-360°, vertical axis discharge amplitude 0-1000pC, matrix value is discharge pulse density), and map the amplitude range to [0, 1] by linear normalization;

[0064] PRPS data: Construct a three-dimensional matrix Y2 (horizontal axis phase 0-360°, vertical axis time period 0-100, matrix value is discharge amplitude), extract time series features and normalize to the range [0, 1];

[0065] Adaptive weighted fusion: A dynamic weight coefficient α (optimized to α=0.52 through cross-validation of the training set) is introduced to perform weighted fusion of the standardized PRPD and PRPS data, as shown in the formula Y. fuse =α・Y1+(1-α)・Y2, and then expanded into a three-channel input image Z=[Y] through feature mapping. fuse [;Y1;Y2], with a uniform size of 288×288 (zero fill adjustment), to enhance the feature correlation between modes.

[0066] In step 3, a fault identification model is constructed based on an attention-enhanced lightweight convolutional neural network, specifically as follows:

[0067] This section introduces the attention-enhanced lightweight convolutional neural network, which comprises a four-level architecture: "input module - feature extraction module - attention optimization module - classification and hierarchical module." The specific layers and functions are as follows:

[0068] 1. Input module: Receives a weighted fusion multimodal image of 288×288×3, providing standardized input for subsequent feature extraction;

[0069] 2. Feature extraction module: Contains 3 convolutional units (each unit consists of a convolutional layer + a batch normalization layer + a ReLU activation layer):

[0070] (1) Convolutional layer: 3×3 convolutional kernels are used, with the number of kernels being 16, 32, and 64 respectively, stride 1, padding method is same, and the output feature map size is maintained at 288×288; the convolution operation formula is:

[0071] (1)

[0072] In the formula, This is the output of the k-th layer. For trainable weights, For bias terms;

[0073] (2) Batch Normalization Layer: Normalizes the convolution output using the following formula:

[0074] (2)

[0075] In the formula,

[0076] (3)

[0077] In the formula, μ is the mini-batch mean, σ² is the variance, and ε=1e-5 avoids numerical explosion and accelerates model convergence;

[0078] (3) ReLU activation layer: realizes nonlinear transformation, the formula is:

[0079] (4)

[0080] Used to enhance the model's ability to fit complex features.

[0081] 3. Attention Optimization Module: A channel attention submodule is added after the 2nd and 3rd convolutional units. It dynamically adjusts the feature channel weights through a squeeze-encouragement (SE) mechanism. The operation is described below:

[0082] (1) Squeezing operation: Perform global average pooling on the feature map to obtain the channel descriptor, which is:

[0083] (5)

[0084] In the formula, H and W are the height and width of the feature map;

[0085] (2) Activation operation: Output channel weights through a fully connected layer and a Sigmoid function:

[0086] (6)

[0087] The optimized features obtained by weighting the original feature map are:

[0088] (7)

[0089] Ultimately, the focus is on key discharge characteristics;

[0090] 4. Feature Dimensionality Reduction and Encoding Module: It contains two hybrid pooling layers (first layer max pooling + second layer average pooling, kernel size 2×2, stride 2) to reduce the feature map size to 72×72; then it connects to two fully connected layers (with 256 and 128 neurons respectively) to encode the features in low dimension.

[0091] 5. Overfitting suppression module: A Dropout layer (dropout rate 0.45) is added after the fully connected layer to randomly deactivate some neurons and reduce the risk of overfitting;

[0092] 6. Classification and Grading Module: Adopts a dual-output structure, specifically as follows:

[0093] (1) Fault type output: The type probability is output through a Softmax layer (4 neurons, corresponding to four types of faults), and the formula is:

[0094] (8)

[0095] The maximum value corresponds to the fault type;

[0096] (2) Severity output: The grading is achieved by combining the Sigmoid layer (1 neuron, outputting 0-1 value) with the grading threshold (0-0.33 for mild, 0.33-0.66 for moderate, and 0.66-1 for severe). The threshold is obtained by training and optimization with 500 sets of samples.

[0097] In step 4, the fault identification model is trained based on the dataset, specifically as follows:

[0098] 1. Training parameters and strategy design

[0099] (1) Weight initialization: The Xavier uniform distribution initialization method is used, and the weights are:

[0100] (9)

[0101] In the formula, The number of input neurons, (Number of output neurons), ensuring initial gradient stability;

[0102] (2) Loss function: A combined loss function is used, which includes the cross-entropy loss for type classification and the mean squared error loss for severity regression. The formula is as follows:

[0103] (10)

[0104] In the formula, For cross-entropy loss, The mean squared error is given, and λ = 0.0003 is the L2 regularization coefficient.

[0105] (3) Optimizer and hyperparameters: The AdamW optimizer was selected (learning rate 0.008, β1=0.9, β2=0.999, ε=1e-7), batch size 64, training iterations 80 epochs, and the learning rate was exponentially decayed to 0.5 times the original value every 20 epochs;

[0106] (4) Dataset construction: The sample set consists of 500 multimodal input images (125 images for each of the four types of faults, including 80 sets of field measurement data), which are divided into training set (325 images), validation set (75 images), and test set (100 images) according to 65%, 15%, and 20% respectively; the training set is augmented with data (adding ±3% Gaussian noise and time series flipping) to suppress overfitting;

[0107] 2. Performance verification and adaptation optimization

[0108] Laboratory performance verification: The test set showed a fault type identification accuracy of 94.2%, a severity classification accuracy of 92.5%, and an inference time of 78ms, which is 4% higher in accuracy and 22% faster inference speed than the existing solution.

[0109] On-site testing and verification: Ten 110kV transformers with different operating years (5-20 years) were selected, and 30 sets of fault data were collected for testing. The average identification accuracy was 93.8%, and the classification accuracy was 91.2%, which met the on-site requirements.

[0110] New energy operating conditions adaptation: Introducing a harmonic distortion rate correction coefficient (coefficient 0.92 when harmonic distortion rate > 8%, 0.95 when harmonic distortion rate is 5%-8%, and 0.98 when harmonic distortion rate is < 5%), the accuracy fluctuation under extreme harmonic operating conditions is ≤ 2.5%, adapting to new energy grid connection scenarios.

[0111] like Figure 2 As shown, this invention, with its innovative logic of "hardware anti-interference acquisition - signal quality optimization - multi-modal efficient fusion," fully presents the entire process from signal acquisition to standardized input:

[0112] On the hardware side, the HFCT-30MHz sensor is innovatively deployed precisely (at the grounding wire at the bottom of the transformer tank, 5cm away from the grounding terminal), and paired with a twisted-pair transmission link with 95% shielding coverage and a 50Ω matching resistor to build a strong anti-interference acquisition system, solving the defect of the acquisition scheme being susceptible to electromagnetic interference;

[0113] A new 3-layer threshold denoising module based on the db4 wavelet basis (threshold λ=1.2σ) is added to the preprocessing stage, which improves the signal-to-noise ratio of the discharge signal from 20dB to 35dB, effectively eliminating high-frequency interference without losing pulse details;

[0114] The multimodal fusion stage abandons the traditional simple channel splicing. It dynamically balances the PRPD phase-amplitude features and PRPS temporal features through an adaptive weight coefficient α=0.52 that has been cross-validated by 500 sets of samples, and finally generates a 288×288×3 multimodal input image, which improves the modal correlation utilization rate by 20%, significantly optimizes the multimodal fusion efficiency, and comprehensively solves the core pain points of low signal quality and coarse fusion.

[0115] like Figure 3 As shown, this invention takes "precise feature extraction - key information focusing - integrated functional output" as its core innovation and constructs a four-level lightweight network architecture:

[0116] Input layer reception Figure 2 The output 288×288×3 normalized multimodal image is extracted by the feature extraction unit through a combination of 16 / 32 / 64 3×3 convolutional kernels (stride 1, same padding) and ReLU activation function to achieve efficient extraction of basic features;

[0117] An innovative SE channel attention module is introduced, which dynamically adjusts the feature channel weights through a process of "global average pooling (squeezing) - fully connected layer + Sigmoid (excitation)", thereby increasing the weights of key features such as sudden changes in discharge pulse density and timing anomalies by 15%-20%.

[0118] The feature dimensionality reduction process breaks through the single pooling mode and adopts a hybrid pooling design of "pre-layer max pooling + post-layer average pooling" (2×2 kernels, stride 2) to balance feature robustness and detail preservation. It is paired with a 256 / 128 neuron fully connected layer and a 0.45 dropout rate to suppress overfitting. The output layer features an innovative dual-output structure: the left Softmax layer (4 neurons) outputs the fault type, and the right Sigmoid layer (1 neuron) outputs the severity value of 0-1. The overall model inference time is only 78ms and the test set accuracy reaches 94.2%. While achieving lightweight design, it expands the "recognition-grading" integrated function and adapts to the deployment requirements of edge devices.

[0119] like Figure 4 As shown, this invention constructs a complete link from technical diagnosis to operation and maintenance implementation based on the closed-loop innovation logic of "data input-preprocessing-prediction-hierarchy-decision".

[0120] The data input stage supports dual sources: "field-acquired signals" and "laboratory-simulated signals," enhancing the applicability of the solution.

[0121] The preprocessing stage is directly related Figure 2 The "wavelet denoising-weighted fusion" process ensures the quality of input data;

[0122] The model prediction stage relies on Figure 3 The attention-enhancing network synchronously outputs the fault type (four types) and a severity value of 0-1;

[0123] An innovative three-level severity assessment mechanism is designed, which achieves automated classification through visual threshold segmentation (mild 0-0.33, moderate 0.33-0.66, severe 0.66-1), with a classification accuracy rate of 92.5%.

[0124] The decision output process generates differentiated operation and maintenance recommendations based on three severity levels (quarterly monitoring for minor faults, monthly monitoring for moderate faults with overload prevention, and immediate shutdown for severe faults), reducing fault response time by 30%. It also supports multi-terminal output via edge LCD display, mobile APP push notifications, and cloud platform synchronization, forming a complete closed loop of "diagnosis results - operation and maintenance decisions - multi-terminal implementation". Additionally, it includes a "new energy harmonic correction" branch to adapt to high-proportion new energy power grid operating conditions.

[0125] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] The present invention also provides an embodiment:

[0127] Specific Example: A Case Study on Partial Discharge Fault Diagnosis of a 110kV Power Transformer

[0128] Background: A 110kV power transformer with a service life of 12 years collects partial discharge signals through a sensor network. After preprocessing, the signals are input into a trained fault identification model. The model outputs the fault type as "internal floating discharge" with a severity score of 0.72 (belonging to the severe level).

[0129] Implementation process:

[0130] Severity classification:

[0131] The model outputs a severity score (range 0-1) through a Sigmoid layer and automatically grades it according to a preset threshold.

[0132] Mild: 0-0.33;

[0133] Moderate: 0.33-0.66;

[0134] Severity: 0.66-1;

[0135] In this example, the severity score of 0.72 > 0.66, so it is judged as a "critical" fault.

[0136] Differentiated operation and maintenance suggestions are generated:

[0137] Based on the classification results, the system automatically matches the corresponding operation and maintenance strategies from the knowledge base and generates specific suggestions:

[0138] Minor faults (0-0.33): Quarterly monitoring is recommended, that is, partial discharge retesting every 3 months, combined with routine inspections, without the need for immediate shutdown.

[0139] Moderate fault (0.33-0.66): It is recommended to "monitor monthly + prohibit overload operation", that is, to conduct a detailed inspection once a month and limit the transformer load to within 80% of the rated capacity to prevent the fault from worsening.

[0140] Severe fault (0.66-1): It is recommended to "immediately shut down and carry out emergency repairs", that is, immediately arrange for the transformer to be taken out of operation, and carry out insulation repair or component replacement within 24 hours, while activating the emergency plan.

[0141] In this example, for a "critical" fault, the system outputs the following suggestions:

[0142] "Fault type: internal floating discharge; severity: severe; recommended action: immediately shut down the power supply and contact the maintenance team to conduct insulation inspection and repair within 24 hours. At the same time, activate the backup transformer power supply plan to ensure continuous operation of the power grid."

[0143] Multi-terminal output and decision execution:

[0144] Operations and maintenance personnel recommend using edge LCD displays, mobile app push notifications, and synchronized output via cloud platforms to achieve multi-terminal collaboration.

[0145] After receiving the advice, the on-site maintenance personnel immediately carried out the shutdown operation. The maintenance team prioritized the transformer based on the advice. The entire process from fault identification to decision execution took only 15 minutes (traditional manual diagnosis usually takes more than 60 minutes).

[0146] Validity Proof

[0147] Reduced fault response time: In this embodiment, the entire process from signal input to suggestion generation is automated, reducing the response time by more than 30% and avoiding insulation breakdown accidents caused by delayed processing.

[0148] Accuracy of Operation and Maintenance Decisions: Based on the model's grading results (laboratory validation grading accuracy of 92.5%), the recommendations are highly targeted, avoiding the problems of "over-maintenance" or "under-maintenance". In this case, serious faults were identified and handled in a timely manner, preventing potential power outage losses (estimated to have avoided economic losses of approximately 500,000 yuan).

[0149] Adapting to complex scenarios: In this example, the transformer operating environment has harmonic interference (harmonic distortion rate of 6%), but through the new energy harmonic correction module built into the model, the severity score error is less than 2%, ensuring the reliability of the recommendations.

[0150] This embodiment demonstrates that the present invention not only achieves accurate fault identification, but also significantly improves the efficiency and scientific nature of operation and maintenance decisions through automated classification and suggestion generation, providing a complete solution for the intelligent operation and maintenance of power transformers.

[0151] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0153] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0154] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for identifying and classifying the severity of discharge faults in power transformers based on wavelet denoising and SE attention mechanism, characterized in that, include: Step 1: Collect partial discharge signals in real time through a sensor network deployed at key locations on the power transformer; Step 2: Preprocess the collected partial discharge signals to construct a dataset; Step 3: Construct a fault identification model based on an attention-enhanced lightweight convolutional neural network; Step 4: Train the fault identification model based on the dataset; Step 5: Input the partial discharge signal to be detected into the trained fault identification model, and output the fault identification result and severity classification; Step 6: Generate differentiated operation and maintenance recommendations based on severity level.

2. The method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 1, characterized in that, In step 1, partial discharge signals are collected in real time through a sensor network deployed at key locations on the power transformer. Specifically: A high-frequency pulse current sensor is installed at the bottom of the transformer tank where the power is connected. The current sensor transmits the signal to the partial discharge instrument via a shielded twisted pair cable, and transmits the signal along with the synchronous voltage signal to an oscilloscope to acquire the partial discharge signal.

3. The method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 2, characterized in that, The partial discharge signal includes PRPD data and PRPS data.

4. The method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 3, characterized in that, In step 2, the acquired partial discharge signals are preprocessed, specifically as follows: The acquired partial discharge signals were sequentially subjected to wavelet threshold denoising, data standardization, and adaptive weighted fusion.

5. A method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 4, characterized in that, The fault identification model includes an input module, a feature extraction module, an attention optimization module, a feature dimensionality reduction and encoding module, an overfitting suppression module, and a classification and grading module.

6. The method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 5, characterized in that, The feature extraction module includes three convolutional units, each of which consists of a convolutional layer, a batch normalization layer, and a ReLU activation layer.

7. A method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 6, characterized in that, The convolutional layer has 3×3 convolutional kernels, with the number of kernels being 16, 32, and 64 respectively, a stride of 1, and a padding method of "same".

8. A method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 7, characterized in that, The attention optimization module adopts the SE channel attention mechanism, which dynamically adjusts the feature channel weights through squeeze-excitation operations.

9. A method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 7, characterized in that, The feature reduction and encoding module includes two hybrid pooling layers and two fully connected layers.

10. A method for identifying and classifying the severity of power transformer discharge faults based on wavelet denoising and SE attention mechanism according to claim 9, characterized in that, The classification and grading module adopts a dual-output structure, including a Softmax layer and a Sigmoid layer. The Softmax layer is used to output the probability of the fault type, and the Sigmoid layer is used to output the severity score.

Images