Partial discharge detection network system based on adaptive attention mechanism

Abstract

The invention relates to the technical field of power system monitoring and fault diagnosis, and discloses a partial discharge detection network system based on an adaptive attention mechanism. Comprising a data acquisition module, a signal preprocessing and time-frequency image generation module, a deep learning network of an adaptive attention mechanism, a fault classification and diagnosis module and a real-time monitoring and alarm module, the data acquisition module acquires a multi-mode partial discharge signal in real time, and preprocessing is performed through the signal preprocessing and time-frequency image generation module. The deep learning network of the adaptive attention mechanism carries out feature extraction and classification on the time-frequency image, finally, the fault classification and diagnosis module outputs a diagnosis result, and the real-time monitoring and alarm module carries out real-time monitoring and automatic alarm according to the diagnosis result. According to the system, multi-mode partial discharge signals are collected in real time, feature extraction and classification are carried out through a deep learning network, the attention degree of different features can be dynamically adjusted through a self-adaptive attention mechanism, and the accuracy and robustness of detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system monitoring and fault diagnosis, and in particular relates to a partial discharge detection network system based on an adaptive attention mechanism. Background Art

[0002] Partial discharge (PD) is an early manifestation of insulation failure in power equipment, and its detection and diagnosis are crucial to ensuring the safe operation of power systems. Traditional PD detection methods, such as frequency domain analysis and time domain signal processing, perform well in some cases, but their accuracy and real-time performance are often insufficient when dealing with complex and changing power environments.

[0003] Traditional frequency domain analysis and time domain signal processing methods are sensitive to environmental noise and are prone to misjudgment and missed judgment. In complex electromagnetic environments, the accuracy and reliability of these methods may be affected. Traditional signal processing methods are often based on fixed algorithms and models, which are difficult to adapt to different types of partial discharge signals. For diverse discharge modes, these methods may not be able to effectively extract features, limiting the accuracy of detection. Some traditional detection methods require a long data processing time, resulting in insufficient real-time performance. In power systems that require rapid response, this may mean a delay in fault diagnosis, thereby increasing system risks. Traditional methods usually rely on manually designed features for feature extraction, which may not fully reflect the complexity and diversity of partial discharge signals. The models used in traditional methods may lack sufficient generalization capabilities, resulting in a decrease in model performance when encountering new or unseen partial discharge modes. Some traditional methods may require manual participation in feature selection and model tuning, which not only increases the workload but also increases the possibility of human error. Different detection equipment may require different signal processing methods, resulting in poor system compatibility and interoperability.

[0004] With the rapid development of deep learning technology, especially the introduction of adaptive attention mechanism, it provides more powerful technical support for the detection and feature extraction of partial discharge signals. In view of this, we propose a partial discharge detection network system based on adaptive attention mechanism to solve the above problems. Summary of the invention

[0005] The present invention aims to solve the technical problem that the local discharge detection methods in the above-mentioned prior art, such as frequency domain analysis and time domain signal processing, although they perform well in some cases, are often insufficient in accuracy and real-time performance when dealing with complex and changeable power environments.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A partial discharge detection network system based on an adaptive attention mechanism includes a data acquisition module, a signal preprocessing and time-frequency image generation module, a deep learning network of an adaptive attention mechanism, a fault classification and diagnosis module, and a real-time monitoring and alarm module; wherein the deep learning network of the adaptive attention mechanism includes a CNN module, an adaptive attention mechanism module, and a multi-head attention mechanism;

[0008] The data acquisition module is responsible for collecting multimodal partial discharge signals from power equipment in real time, which are then passed to the signal preprocessing and time-frequency image generation module to filter, denoise and normalize the signals. The processed time-frequency images are input into the deep learning network of the adaptive attention mechanism. The time-frequency images are used as the input data of the network. After feature extraction through the CNN module, the extracted feature maps are sent to the adaptive attention mechanism module to further enhance the key features. The output feature maps of the adaptive attention mechanism module are input into the multi-head attention mechanism. The final output of the deep learning network of the adaptive attention mechanism, that is, the probability distribution of partial discharge types, is passed to the fault classification and diagnosis module. The output fault diagnosis results of the fault classification and diagnosis module are passed to the real-time monitoring and alarm module. The real-time monitoring and alarm module is responsible for real-time monitoring of partial discharge signals and automatically triggers an alarm when an abnormality is detected or the set threshold is exceeded.

[0009] Preferably, the data acquisition module collects partial discharge signals in real time through sensors installed on the power equipment. The collected partial discharge signals include but are not limited to TEV signals, ultrasonic signals and high-frequency electromagnetic wave signals and will be used synchronously as basic inputs in subsequent processing.

[0010] Preferably, the signal preprocessing and time-frequency image generation module converts the filtered, denoised and normalized preprocessed time domain signal into a time-frequency image using short-time Fourier transform or wavelet transform;

[0011] The time-frequency image can simultaneously reflect the time domain and frequency domain characteristics of the signal, providing more suitable input data for subsequent feature extraction.

[0012] Preferably, the CNN module captures the local spatial features of the signal; the adaptive attention mechanism module calculates the attention weight according to each area of ​​the input feature map, and the attention weight represents the network's evaluation of the importance of different feature areas, so as to dynamically adjust the degree of attention to different feature areas; the multi-head attention mechanism calculates the attention distribution of different subspaces in parallel through multiple independent attention heads, and fuses the features of different heads through splicing and linear transformation.

[0013] As a preferred option, the deep learning network design of the adaptive attention mechanism is as follows:

[0014] Input data preparation: The multi-scale feature maps processed by the convolutional neural network (CNN) are input into the adaptive attention module, which generates an adaptive attention weight map for each feature map to strengthen important feature areas and suppress minor features.

[0015] The specific design of the adaptive attention mechanism: generate query, key and value vectors, calculate the attention score, and perform Softmax operation to generate a weighted feature map;

[0016] Multi-head attention mechanism: Use multiple independent attention heads to parallelly calculate the attention distribution of different subspaces, concatenate and fuse the outputs of all attention heads, and obtain the multi-head attention feature map through linear transformation;

[0017] Output and feature fusion: The original CNN feature map is fused with the multi-head attention feature map through residual connection, and then the global feature representation is obtained through pooling or dimensionality reduction operation for subsequent fault classification.

[0018] As a preference, the specific steps of the adaptive attention mechanism are as follows:

[0019] Step 1: Calculation of Query, Key, and Value: For each input feature map, the corresponding Query, Key, and Value vectors are generated by multiplying three different trainable weight matrices. The formula is as follows:

[0020] Step 2: Calculate the attention score: Calculate the attention score by the dot product of Query and Key: Among them, d k is the dimension of the Key vector; is a scaling factor for numerical stability;

[0021] Step 3: Apply the Softmax function: Normalize the calculated attention score through the Softmax function to obtain the probability distribution, that is, the attention weight matrix A: A = Softmax (Attention_score); make the sum of all weights equal to 1 to achieve weighted summation;

[0022] Step 4: Calculate the weighted value: multiply the obtained attention weight A by the corresponding value vector V to generate a weighted feature map O, O = A × V;

[0023] Step 5: Output weighted feature map: The weighted feature map O will be used as the output of the self-attention module. This output is concatenated with the original input feature map to enhance the feature representation.

[0024] As a preferred embodiment, the content of the multi-head attention mechanism is as follows: input the generated weighted feature map O, and through the multi-head attention mechanism, use multiple independent attention heads to calculate the attention distribution of different subspaces in parallel: O i =Head i (Q, K, V), i = 1, 2, ..., h; concatenate the outputs of all attention heads and fuse the features of different heads through linear transformation: MultiHead_Output = Concat(O1, O2, ..., O h )×W o , where W o It is a trainable weight matrix used for linear transformation. The fused multi-head attention feature map contains the global dependency and multi-scale features of the local discharge signal.

[0025] Preferably, the content of output and feature fusion is as follows: input the feature map output by the multi-head attention mechanism, fuse the original CNN feature map with the multi-head attention feature map through residual connection to retain the original detail information: Final_Feature=CNN_output+MultiHead_Output, and the final feature map is formed into a global feature representation through pooling operation or dimensionality reduction, which is input into the classifier for partial discharge type classification, and outputs a complete feature vector for subsequent fault classification.

[0026] As a preferred method, the training process of the deep learning network model of the adaptive attention mechanism is as follows: input the preprocessed time-frequency image dataset, the local discharge type label annotated by L, use the cross entropy loss function, Adam optimizer and batch training strategy, during the training process, the network automatically adjusts the attention weight through back propagation, and the trained local discharge detection network model accurately classifies the local discharge type on a given test dataset;

[0027] Among them, the specific contents of the cross entropy loss function, Adam optimizer and batch training strategy are as follows:

[0028] The cross entropy loss function is used to calculate the difference between the predicted value and the true label: Among them, C is the number of categories, yi is the true label, is the probability predicted by the model;

[0029] Use the Adam optimizer to train the model and adjust the trainable parameters in the network, including the CNN convolution kernel and the weight matrix in the adaptive attention mechanism;

[0030] Batch training is adopted, and it is iterated in multiple epochs. In each epoch, the model calculates the prediction results through forward propagation and updates the weights through back propagation.

[0031] Adaptive attention weight adjustment: In each iteration, the network automatically adjusts the attention weight matrix through the back-propagation algorithm, so that the model can better capture the key information areas.

[0032] Preferably, the reasoning process of the deep learning network model of the adaptive attention mechanism is as follows: input new partial discharge signal data, pass the input image through the trained network, and then pass through CNN, adaptive attention mechanism, multi-head attention mechanism and fully connected layer in sequence, and finally output the classification result, output the probability distribution of the partial discharge type and the final classification result, which are used to generate a fault diagnosis report.

[0033] Preferably, the fault classification and diagnosis module inputs the feature map generated by the adaptive attention mechanism into the classifier to classify the partial discharge type and generate a corresponding fault diagnosis report, and the types of classification results include internal discharge, surface discharge and corona discharge;

[0034] Preferably, the real-time monitoring and alarm module monitors the partial discharge signal of the power equipment in real time and automatically determines whether it exceeds the preset threshold. When an abnormal signal is detected, the system will immediately trigger an alarm to notify the operation and maintenance personnel to conduct inspection and maintenance.

[0035] Compared with the prior art, the technical effects and advantages of the present invention are:

[0036] The partial discharge detection network system based on the adaptive attention mechanism collects multimodal partial discharge signals from power equipment, such as TEV (transverse electromagnetic wave), ultrasonic signals and high-frequency electromagnetic wave signals. Then, these signals are preprocessed, including filtering, denoising and normalization, and converted into time-frequency images using short-time Fourier transform (STFT) or wavelet transform (WT). Next, the system uses a convolutional neural network (CNN) to extract feature maps, and dynamically adjusts the network's attention to different feature areas through an adaptive attention mechanism to strengthen key features and suppress secondary features. Thereafter, the attention distribution of different subspaces is calculated in parallel through a multi-head attention mechanism, and the outputs of all attention heads are fused. Finally, the fused feature vector is input into the classifier to output the probability distribution of the partial discharge type and the final classification result.

[0037] By combining a convolutional neural network (CNN) with an adaptive attention module, the network's attention distribution is dynamically adjusted to achieve accurate classification and fault diagnosis of partial discharge signals. Unlike the traditional fixed attention mechanism, the adaptive attention module can dynamically adjust the attention distribution according to changes in input features, thereby accurately capturing the key information of partial discharge signals even in complex environments.

[0038] The adaptive attention mechanism can dynamically adjust the network's attention to different feature areas, improving the accuracy and robustness of partial discharge signal detection, especially in complex power environments. This mechanism enhances the model's adaptability to diverse discharge patterns in complex power environments, allowing the system to more effectively handle different types of partial discharge signals. The system has real-time monitoring and alarm functions, and can automatically trigger an alarm when the partial discharge signal exceeds the set threshold, improving the real-time nature of power grid monitoring and the reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a system principle block diagram of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] The following is combined with Figure 1 To further explain this application in detail,

[0042] The embodiment of the present application discloses a partial discharge detection network system based on an adaptive attention mechanism, comprising: a data acquisition module, a signal preprocessing and time-frequency image generation module, a deep learning network of an adaptive attention mechanism, a fault classification and diagnosis module, and a real-time monitoring and alarm module; wherein the deep learning network of the adaptive attention mechanism comprises a CNN module, an adaptive attention mechanism module, and a multi-head attention mechanism;

[0043] The data acquisition module is responsible for real-time acquisition of multi-modal partial discharge signals from power equipment, such as TEV (transverse electromagnetic wave), ultrasonic signals and high-frequency electromagnetic wave signals. Data is collected through sensors, and the signals will serve as the basic input for subsequent processing.

[0044] Signal preprocessing and time-frequency image generation module: The collected partial discharge signal is filtered, denoised and normalized to remove noise and enhance signal quality. Subsequently, the signal is converted into a time-frequency image using short-time Fourier transform (STFT) or wavelet transform (WT) for input into the deep learning network for feature extraction.

[0045] Deep Learning Network with Adaptive Attention Mechanism:

[0046] Convolutional Neural Network (CNN) module: used to extract multi-scale local features from time-frequency images. These features will be refined step by step through convolution and pooling operations.

[0047] Adaptive attention mechanism module: The adaptive attention mechanism dynamically adjusts the attention weights on the feature maps extracted by CNN to enhance the expression of important feature areas while suppressing minor features. This module can automatically identify key areas in the signal and adapt to different types of partial discharge signals.

[0048] Multi-head attention mechanism: A multi-head attention mechanism is used to process different subspaces of the feature map in parallel, improve the perception of global dependencies, and finally fuse them to form a more comprehensive feature representation.

[0049] Fault classification and diagnosis module: The feature map generated by the adaptive attention mechanism is input into the classifier to classify the partial discharge type and generate the corresponding fault diagnosis report. The classification results include internal discharge, surface discharge, corona discharge and other types.

[0050] The real-time monitoring and alarm module can monitor the partial discharge signals of power equipment in real time and automatically determine whether they exceed the preset threshold. When an abnormal signal is detected, the system will immediately trigger an alarm to notify the operation and maintenance personnel to conduct inspection and maintenance.

[0051] The specific implementation steps of the partial discharge detection network system based on the adaptive attention mechanism are as follows:

[0052] 1. Data collection and preprocessing

[0053] 1.1 Data acquisition: The system collects partial discharge signals in real time through sensors installed on power equipment. These signals include but are not limited to TEV signals, ultrasonic signals and high-frequency electromagnetic wave signals. These signals will be used synchronously in subsequent processing.

[0054] 1.2 Signal preprocessing: The collected signal is subjected to noise filtering and normalization to enhance the effective components of the signal and remove unnecessary noise. A bandpass filter is used to remove high-frequency interference and low-frequency noise, and wavelet denoising technology is used to further clean up the signal.

[0055] 1.3 Time-frequency image generation: The preprocessed signal is transformed into a time-frequency image through short-time Fourier transform (STFT) or wavelet transform (WT). The purpose of generating a time-frequency image is to visualize the time domain characteristics and frequency domain characteristics of the signal for subsequent feature extraction.

[0056] 2. Adaptive Attention Mechanism Design

[0057] In the partial discharge detection network of the present invention, the adaptive attention mechanism module dynamically adjusts the attention distribution of the network, concentrates resources on key feature areas, and improves the classification ability of signals. The following are detailed design steps, including the input, output and training process of each step.

[0058] 2.2.1 Input Data Preparation

[0059] enter:

[0060] Feature maps extracted by CNN: Multi-scale feature maps (usually three-dimensional tensors with dimensions [batch_size, height, width, channels]) processed by convolutional neural networks (CNN). These feature maps contain the local spatial features of local discharge signals.

[0061] Output:

[0062] Feature Map: The adaptive attention module will generate an adaptive attention weight map for each feature map, strengthening important feature areas and suppressing minor features.

[0063] 2.2.2 Specific Design of Adaptive Attention Mechanism

[0064] Step 1: Calculate self-attention weights

[0065] Input: Multi-scale feature map output by CNN (dimension is [batch_size, height, width, channels]).

[0066] A set of trainable weight matrices (Query, Key, Value), usually of dimension [channels, d_k], where dk is the dimension of the feature vector.

[0067] operate:

[0068] 1. Generate Query, Key, Value: Multiply the input feature map by the weight matrix to generate Query (Q), Key (K) and Value (V) vectors:

[0069] Q=CNN output ×W q

[0070] K = CNN output ×W k

[0071] V=CNN output ×W v ;

[0072] These vectors are used to calculate the attention weights.

[0073] 2. Calculate the attention score: Calculate the attention score by the dot product of Query and Key: Here, d kis the dimension of the Key vector, and the dot product result is scaled to maintain numerical stability.

[0074] 3. Softmax operation: Perform a Softmax operation on the attention score to obtain the attention weight matrix (A). Each item in the matrix represents the relative importance of the feature to other features:

[0075] A=Softmax(Attention_score); This makes the sum of all weights equal to 1, thus achieving weighted summation.

[0076] Generating weighted feature maps

[0077] Input: Attention weight matrix A (dimension is [batch_size, height, width, height\timeswidth]). Value vector V (same dimension as the original feature map).

[0078] Operation: Multiply the attention weight matrix A by the Value vector V to generate the weighted feature map O:

[0079] O = A × V; This process is similar to applying an adaptive filter to the feature map to enhance those feature areas considered important by the attention mechanism.

[0080] Output: The weighted feature map O (with the same dimensions as the input feature map [batch_size, height, width, channels]), which contains the key features amplified by the adaptive attention mechanism.

[0081] 2.2.3 Multi-Head Attention Mechanism

[0082] Input: The weighted feature map O generated in the previous step.

[0083] Operation: Through the multi-head attention mechanism, use multiple independent attention heads (for example, 8 attention heads) to parallelly calculate the attention distribution of different subspaces: i =Head i (Q, K, V), i=1, 2, ...., h;

[0084] Concatenate the outputs of all attention heads and fuse the features of different heads through linear transformation: MultiHead_Output = Concat(O1, O2, ..., O h )×W o ; Here W o is a trainable weight matrix for linear transformation.

[0085] Output: The fused multi-head attention feature map (with the same dimension as the input feature map), which also contains the global dependency and multi-scale features of the PD signal.

[0086] 2.2.4 Output and feature fusion

[0087] Feature fusion and global features are expressed as follows:

[0088] Input: Feature map output by the multi-head attention mechanism.

[0089] Operation: Through residual connection, the original CNN feature map is fused with the multi-head attention feature map to retain the original detail information:

[0090] Final_Feature=CNN_output+MultiHead_Output;The final feature map is formed into a global feature representation through pooling operation or dimensionality reduction, and is input into the classifier for partial discharge type classification.

[0091] Output: A complete feature vector (of dimension [batch_size, feature_dim]) for subsequent fault classification.

[0092] 2.2.5 Training Process

[0093] The model training is as follows:

[0094] Input: preprocessed time-frequency image dataset, annotated partial discharge type labels.

[0095] operate:

[0096] Loss function: The cross-entropy loss function is used to calculate the difference between the predicted value and the true label:

[0097]

[0098] Where C is the number of categories, yi is the true label, is the probability predicted by the model.

[0099] Optimization algorithm: Use the Adam optimizer to train the model and adjust the trainable parameters in the network (including CNN convolution kernels, weight matrices in the adaptive attention mechanism, etc.).

[0100] Training strategy: Mini-batch training is used, and it is iterated in multiple epochs. In each epoch, the model calculates the prediction results through forward propagation and updates the weights through back propagation.

[0101] Adaptive attention weight adjustment: In each iteration, the network automatically adjusts the attention weight matrix through the back-propagation algorithm, so that the model can better capture the key information areas.

[0102] Output: A trained partial discharge detection network model that can accurately classify partial discharge types on a given test dataset.

[0103] 2.2.6 Reasoning Process

[0104] The model reasoning content is as follows:

[0105] Input: New partial discharge signal data (time-frequency image obtained through preprocessing and time-frequency image generation module).

[0106] Operation: The input image is passed through the trained network, through CNN, adaptive attention mechanism, multi-head attention mechanism and fully connected layer in sequence, and finally the classification result is output.

[0107] Output: probability distribution of partial discharge types and final classification results, which are used to generate fault diagnosis reports.

[0108] Through the above detailed design, the adaptive attention mechanism of the present invention can dynamically adjust the network's attention to different feature areas, thereby significantly improving the accuracy and robustness of partial discharge signal detection. This mechanism not only enhances the model's adaptability to diverse discharge modes in complex power environments, but also improves the real-time performance and reliability of the system in practical applications.

[0109] 2.3 Feature fusion and classification: The feature map processed by the adaptive attention mechanism is fused with other global features to form a complete feature representation. The fused feature vector is input into the classifier:

[0110] Fully connected layer and classification: The multi-layer fully connected neural network maps the fused feature vector to the classification output and calculates the probability distribution of various types of faults through the Softmax function.

[0111] Fault type identification: Classification results include internal discharge, surface discharge, corona discharge, etc. The system outputs the corresponding fault type based on the highest probability.

[0112] 3. Fault diagnosis and report generation

[0113] 3.1 Generation of diagnostic report: Based on the classification results, the system automatically generates a fault diagnosis report for the power equipment. The report includes the fault type, possible causes, and recommended maintenance measures.

[0114] 3.2 Real-time monitoring and alarm: The system supports real-time monitoring and automatically triggers an alarm when the partial discharge signal exceeds the set threshold, notifying the operation and maintenance personnel to handle the fault in time. The monitoring data can be recorded and archived for subsequent analysis.

[0115] 4. System application scenarios

[0116] 4.1 Online monitoring of substations and high voltage cables

[0117] The system is suitable for online monitoring of substation equipment and high-voltage cables. Through sensors deployed on the equipment, partial discharge signals are collected and analyzed in real time. The system can effectively identify early insulation failures in power equipment and provide early warnings to avoid major accidents caused by partial discharge.

[0118] 4.2 Remote fault diagnosis

[0119] The system can be integrated with the smart grid big data platform to provide remote fault diagnosis functions. By analyzing the data of remote devices in real time, operation and maintenance personnel can promptly identify problems and take measures to reduce the impact of equipment failures on the power grid.

[0120] The adaptive attention mechanism of the present invention can dynamically adjust the network's attention to different feature areas, which means that it can identify and strengthen key features related to partial discharge signals while suppressing irrelevant or noise features. This mechanism enables the model to focus resources on important information, thereby improving the accuracy of detection.

[0121] By introducing the attention weight map, the adaptive attention mechanism can generate a weight map for each feature map, so that important areas and minor areas can be distinguished in the feature map. This weight distribution helps to improve the model's ability to express the characteristics of partial discharge signals.

[0122] In partial discharge signal detection, we may encounter situations where the number of samples of certain discharge types is far less than that of other types. The adaptive attention mechanism can give more attention to minority samples in a weighted manner, thereby alleviating the impact of sample imbalance.

[0123] By dynamically adjusting the attention weights, the adaptive attention mechanism can make the model more flexible and able to adapt to different local discharge patterns and complex environmental conditions, thereby improving the generalization ability and robustness of the model.

[0124] The adaptive attention mechanism not only focuses on local features, but also combines global information through means such as multi-head attention mechanism, while considering the global dependencies between features, to further improve the overall understanding of the discharge signal.

[0125] In power equipment, partial discharge signals may be disturbed by various background noises. The adaptive attention mechanism can identify and reduce these interferences, allowing the model to focus more on the discharge signal itself.

[0126] The adaptive attention mechanism can quickly adapt to new signal patterns and environmental changes, which makes the partial discharge signal detection system have better real-time and adaptability in practical applications.

[0127] In summary, the adaptive attention mechanism enhances the expression of partial discharge signal characteristics by optimizing the weight distribution of feature maps, improves the accuracy and robustness of the detection algorithm, enables the model to better adapt to the complex and changeable power system environment, and has broad application prospects.

[0128] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A partial discharge detection network system based on an adaptive attention mechanism, characterized in that: include: Data acquisition module, signal preprocessing and time-frequency image generation module, deep learning network with adaptive attention mechanism, fault classification and diagnosis module, and real-time monitoring and alarm module; among them, the deep learning network with adaptive attention mechanism includes CNN module, adaptive attention mechanism module and multi-head attention mechanism; The data acquisition module is responsible for collecting multimodal partial discharge signals from power equipment in real time, which are then passed to the signal preprocessing and time-frequency image generation module to filter, denoise and normalize the signals. The processed time-frequency images are input into the deep learning network of the adaptive attention mechanism. The time-frequency images are used as the input data of the network. After feature extraction through the CNN module, the extracted feature maps are sent to the adaptive attention mechanism module to further enhance the key features. The output feature maps of the adaptive attention mechanism module are input into the multi-head attention mechanism. The final output of the deep learning network of the adaptive attention mechanism, that is, the probability distribution of partial discharge types, is passed to the fault classification and diagnosis module. The output fault diagnosis results of the fault classification and diagnosis module are passed to the real-time monitoring and alarm module. The real-time monitoring and alarm module is responsible for real-time monitoring of partial discharge signals and automatically triggers an alarm when an abnormality is detected or the set threshold is exceeded.

2. A partial discharge detection network system based on an adaptive attention mechanism according to claim 1, characterized in that: The data acquisition module collects partial discharge signals in real time through sensors installed on the power equipment. The collected partial discharge signals include but are not limited to TEV signals, ultrasonic signals and high-frequency electromagnetic wave signals, and will be used synchronously as basic inputs in subsequent processing.

3. A partial discharge detection network system based on an adaptive attention mechanism according to claim 1, characterized in that: The signal preprocessing and time-frequency image generation module converts the time domain signal after filtering, denoising and normalization preprocessing into a time-frequency image using short-time Fourier transform or wavelet transform.

4. A partial discharge detection network system based on an adaptive attention mechanism according to claim 1, characterized in that: The CNN module captures the local spatial features of the signal; The adaptive attention mechanism module calculates the attention weight according to each area of ​​the input feature map. The attention weight represents the network's evaluation of the importance of different feature areas, so as to dynamically adjust the degree of attention to different feature areas. The multi-head attention mechanism uses multiple independent attention heads to parallelly calculate the attention distribution of different subspaces, and fuses the features of different heads through splicing and linear transformation.

5. A partial discharge detection network system based on an adaptive attention mechanism according to claim 4, characterized in that: The design of the deep learning network with adaptive attention mechanism is as follows: Input data preparation: The multi-scale feature maps processed by the convolutional neural network (CNN) are input into the adaptive attention module, which generates an adaptive attention weight map for each feature map to strengthen important feature areas and suppress minor features. The specific design of the adaptive attention mechanism: generate query, key and value vectors, calculate the attention score, and perform Softmax operation to generate a weighted feature map; Multi-head attention mechanism: Use multiple independent attention heads to parallelly calculate the attention distribution of different subspaces, concatenate and fuse the outputs of all attention heads, and obtain the multi-head attention feature map through linear transformation; Output and feature fusion: The original CNN feature map is fused with the multi-head attention feature map through residual connection, and then the global feature representation is obtained through pooling or dimensionality reduction operation for subsequent fault classification.

6. A partial discharge detection network system based on an adaptive attention mechanism according to claim 5, characterized in that: The specific steps of the adaptive attention mechanism are as follows: Step 1: Calculation of Query, Key, and Value: For each input feature map, the corresponding Query, Key, and Value vectors are generated by multiplying three different trainable weight matrices. The formula is as follows: Step 2: Calculate the attention score: Calculate the attention score by the dot product of Query and Key: Among them, d k is the dimension of the Key vector; is a scaling factor for numerical stability; Step 3: Apply the Softmax function: Normalize the calculated attention score through the Softmax function to obtain the probability distribution, that is, the attention weight matrix A: A = Softmax (Attention_score); make the sum of all weights equal to 1 to achieve weighted summation; Step 4: Calculate the weighted value: multiply the obtained attention weight A by the corresponding value vector V to generate a weighted feature map O, O = A × V; Step 5: Output weighted feature map: The weighted feature map O will be used as the output of the self-attention module. This output is concatenated with the original input feature map to enhance the feature representation.

7. A partial discharge detection network system based on an adaptive attention mechanism according to claim 6, characterized in that: The content of the multi-head attention mechanism is as follows: Input the generated weighted feature map O, through the multi-head attention mechanism, use multiple independent attention heads to calculate the attention distribution of different subspaces in parallel: O i =Head i (Q, K, V), i = 1, 2, ..., h; concatenate the outputs of all attention heads and fuse the features of different heads through linear transformation: MultiHead_Output = Concat(O1, O2, ..., O h )×W o , where W o It is a trainable weight matrix used for linear transformation. The fused multi-head attention feature map contains the global dependency and multi-scale features of the local discharge signal.

8. A partial discharge detection network system based on an adaptive attention mechanism according to claim 7, characterized in that: The content of output and feature fusion is as follows: input the feature map output by the multi-head attention mechanism, and fuse the original CNN feature map with the multi-head attention feature map through residual connection to retain the original detail information: Final_Feature=CNN_output+MultiHead_Output. The final feature map is formed into a global feature representation through pooling operation or dimensionality reduction, which is input into the classifier for partial discharge type classification, and the complete feature vector is output for subsequent fault classification.

9. The partial discharge detection network system based on the adaptive attention mechanism according to claim 1, characterized in that: The fault classification and diagnosis module inputs the feature map generated by the adaptive attention mechanism into the classifier to classify the local discharge type and generate a corresponding fault diagnosis report. The types of classification results include internal discharge, surface discharge and corona discharge.

10. A partial discharge detection network system based on an adaptive attention mechanism according to claim 1, characterized in that: The real-time monitoring and alarm module monitors the partial discharge signal of the power equipment in real time and automatically determines whether it exceeds the preset threshold.

Images