A lightweight identification method for arc discharge modes in oil based on DSCNN
By using a lightweight, deep separable convolutional neural network model to process arc discharge signals, the problems of large computational scale and slow recognition speed in existing technologies are solved, and rapid arc discharge pattern recognition is achieved in resource-constrained environments.
Patent Information
- Application Number
- CN202411106251.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2025-10-31
- Estimated Expiration
- 2044-08-13
AI Technical Summary
In existing technologies, arc discharge fault diagnosis models have large computational scales, cannot quickly identify short-gap arc discharge modes, and are not applicable to embedded systems or mobile maintenance equipment with limited computing resources.
A lightweight deep separable convolutional neural network (DSCNN) model is used to mine the time-frequency spectrum of the discharge signal. The time-frequency spectrum is generated by wavelet transform, and regularization and ReLU6 activation function are added after depthwise convolution and pointwise convolution to form a DSCNN module, which realizes fast and simple discharge pattern recognition.
While ensuring high recognition accuracy, it reduces computational load and model complexity, improves inference speed, is suitable for resource-constrained environments, and meets online monitoring requirements.
Smart Images
Figure CN119044690B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of arc discharge technology for oil-immersed power transformers, and in particular to a lightweight identification method for arc discharge modes in oil based on DSCNN. Background Technology
[0002] Arc discharge is one of the most serious faults in oil-immersed power transformers, posing a significant threat to national economic development and the safe operation of the power system. When arc discharge develops into a long-gap arc discharge stage, the accumulated discharge is already substantial, and serious equipment failure is almost inevitable. To reduce safety hazards and equipment damage, prevention is necessary when the discharge develops into a short-gap arc discharge or even an electric spark discharge stage. After an arc discharge fault occurs in a transformer, the violent reaction generates various physical and chemical signals, including acoustic, optical, thermal, and other signals. Utilizing online monitoring technology to acquire these signals in real time and explore their generation mechanisms and evolution patterns is a crucial approach for preventative maintenance.
[0003] Furthermore, discharge pattern recognition can quickly and accurately locate defect locations and optimize maintenance strategies in a timely manner. In the field of pattern recognition, artificial intelligence algorithms based on machine learning / deep learning have been extensively studied, such as k-means clustering, support vector machines, Markov models, k-nearest neighbors, and random decision forests. Considering the high danger of arc discharge, rapid identification of its fault types is crucial. However, in some application scenarios, such as embedded systems or mobile maintenance equipment, available computing resources and storage space are very limited. This necessitates that arc discharge fault diagnosis models ensure high accuracy while also minimizing model complexity and data size to operate within limited resources.
[0004] Current research on arc discharge diagnosis mainly focuses on long-gap arc discharge, while research on pattern recognition for short-gap arc discharge and arc discharge caused by different defects is scarce. Existing discharge pattern recognition methods are mostly aimed at partial discharge, which are slow in detection speed, require large computational scale, and have high computer configuration requirements, making them unable to effectively address the challenges of arc discharge faults. There is an urgent need to propose a new lightweight short-gap arc discharge pattern recognition method to resolve the contradiction between the high speed required for arc discharge pattern recognition in online monitoring and the slow speed of existing recognition methods, which are not applicable to arc discharge faults.
[0005] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0006] This invention provides a lightweight identification method for arc discharge patterns in oil based on DSCNN. It uses a lightweight deep separable convolutional neural network model to mine the time-frequency spectrum of the discharge signal, achieving fast and simple discharge pattern identification. This solves the contradiction between the requirement for fast arc discharge pattern identification in online monitoring and the slow speed of existing identification methods, which are not applicable to arc discharge faults.
[0007] A lightweight method for identifying arc discharge patterns in oil based on DSCNN includes:
[0008] Step 1: Conduct short-gap oil arc discharge test experiments on oil-immersed power transformers;
[0009] Step 2: Use sensors to collect the discharge time-domain data of the arc discharge test experiment, perform wavelet transform on it, and generate a time-frequency spectrum;
[0010] Step 3: Divide the time-frequency spectrum into a training set and a test set. Use the training set to train a lightweight deep separable convolutional neural network model. The lightweight deep separable convolutional neural network model decomposes the standard convolution operation into two parts: depthwise convolution and pointwise convolution. Regularization and activation functions are added after both depthwise convolution and pointwise convolution to form a DSCNN module. Use the test set to test the trained model and record the first training parameters.
[0011] Step 4: Train a conventional convolutional neural network model using the training set, test the trained model using the test set, and record the second training parameters. The convolutional neural network model includes convolution, pooling, and fully connected operations. Step 5: By comparing the first and second training parameters, the lightweight depthwise separable convolutional neural network model has smaller parameter size, less computation, and shorter inference time when performing short-gap arc discharge pattern recognition in oil. This confirms the lightweight advantage of the DSCNN-based short-gap arc discharge pattern recognition method, meeting the requirements of online monitoring.
[0012] The lightweight identification method for arc discharge modes in oil based on DSCNN includes conducting short-gap arc discharge test experiments in oil-immersed power transformers. The test platform is constructed by setting up an arc discharge test platform in oil. An AC power supply is applied to the sealed discharge platform by a voltage regulator. The discharge platform has electrodes for discharge and voltage sensors, high-frequency coils, ultrasonic sensors, pressure sensors, and ultra-high frequency sensors for measuring discharge parameters.
[0013] In the lightweight identification method for arc discharge modes in oil based on DSCNN, the short-gap arc fault models tested by the discharge platform are divided into needle-plate discharge, surface discharge, and inter-turn discharge.
[0014] In the aforementioned lightweight identification method for arc discharge modes in oil based on DSCNN, step 2 involves collecting voltage, current, ultrasonic, oil pressure, and ultra-high frequency time-domain signals from the arc discharge test experiment, and generating corresponding time-frequency spectra through wavelet transform.
[0015] In the lightweight recognition method for arc discharge patterns in oil based on DSCNN, a lightweight deep separable convolutional neural network model is trained using a training set. The input layer compresses the time-frequency spectrum to 64 pixels × 64 pixels before feeding it to the convolutional layer. The convolutional layer first captures the global structure and preliminary features of the time-frequency spectrum through convolutional kernels, then extracts deep features through four DSCNN modules, performs downsampling through pooling, and concatenates tensors using the torch.cat() method. The obtained features are then fed to a fully connected layer. Two fully connected layers classify the extracted features, and the classification results are fed to the output layer. The output layer uses the Softmax function to output the predicted probability, thus achieving the recognition task.
[0016] In the lightweight identification method for arc discharge patterns in oil based on DSCNN, the training set trains a conventional convolutional neural network model. The network structure consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected to each other. The first training parameter and the second training parameter each include the recognition accuracy, parameter size, computational cost, model size, and inference time, respectively.
[0017] In the lightweight identification method for arc discharge patterns in oil based on DSCNN, after training, if the identification accuracy of the test set is greater than 80%, the training parameters are recorded; if the identification accuracy of the test set is less than 80%, the learning rate is manually modified and retraining is performed.
[0018] In the lightweight identification method for arc discharge modes in oil based on DSCNN, the algorithm for generating time-frequency spectra from time-domain signals through wavelet transform is as follows:
[0019] ,in, Represents time-domain signal In time and scale wavelet transform value at, It is the mother wavelet function. For the complex conjugate of the mother wavelet function, It is a scale parameter. It is the time shift parameter.
[0020] In the lightweight identification method for arc discharge modes in oil based on DSCNN, in step 3, the regularization process selects the batch normalization module to perform normalization and regularization processing.
[0021] In the lightweight identification method for arc discharge modes in oil based on DSCNN, in step 3, the activation function is the ReLU6 function, and the expression of the ReLU6 function is: , where x is the output value of the activation function, and max() and min() functions are used to find the maximum and minimum values of the sequence, respectively.
[0022] In the lightweight identification method for arc discharge modes in oil based on DSCNN, the ReLU6 function restricts the output value by truncating values greater than 6.
[0023] Compared with existing technologies, the present invention has the following advantages: By adopting a deep separable convolutional neural network, introducing regularization, and improving activation functions, the present invention achieves lightweight data processing, improves inference speed while ensuring high recognition accuracy, and reduces the requirements for equipment computing power, and is expected to be promoted in the field of arc discharge fault identification and online monitoring. Attached Figure Description
[0024] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.
[0025] In the attached diagram:
[0026] Figure 1 This is a flowchart illustrating a lightweight identification method for short-gap arc discharge modes in oil based on DSCNN.
[0027] Figure 2 This invention provides an experimental platform for testing arc discharge in oil-paper insulation.
[0028] Figure 3 This is a schematic diagram illustrating the convolution process of a depth-separable convolutional neural network according to the present invention.
[0029] Figure 4 This is a schematic diagram of a DSCNN module of the present invention;
[0030] Figure 5 This is a schematic diagram of the lightweight depthwise separable convolutional neural network model of the present invention.
[0031] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation
[0032] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
[0033] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.
[0034] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.
[0035] like Figures 1 to 5 As shown, the lightweight identification method for arc discharge modes in oil based on DSCNN includes the following steps:
[0036] Step 1: Conduct short-gap oil arc discharge test experiments on oil-immersed power transformers;
[0037] Step 2: Use sensors to collect the discharge time-domain data of the arc discharge test experiment, perform wavelet transform on it, and generate a time-frequency spectrum;
[0038] Step 3: Divide the time-frequency spectrum into a training set and a test set. Use the training set to train a lightweight deep separable convolutional neural network model. The lightweight deep separable convolutional neural network model decomposes the standard convolution operation into two parts: depthwise convolution and pointwise convolution. Regularization and activation functions are added after both depthwise convolution and pointwise convolution to form a DSCNN module. Use the test set to test the trained model and record the first training parameters.
[0039] Step 4: Train a conventional convolutional neural network model using the training set, test the trained model using the test set, and record the second training parameters. The convolutional neural network model includes convolution, pooling, and fully connected operations. Step 5: By comparing the first and second training parameters, the lightweight depthwise separable convolutional neural network model has smaller parameter size, less computation, and shorter inference time when performing short-gap arc discharge pattern recognition in oil. This confirms the lightweight advantage of the DSCNN-based short-gap arc discharge pattern recognition method, meeting the requirements of online monitoring.
[0040] In a preferred embodiment of the lightweight identification method for arc discharge modes in oil based on DSCNN, the short-gap arc discharge test experiment of an oil-immersed power transformer includes building an arc discharge test platform in oil. The AC power supply is regulated by a voltage regulator to apply a controllable voltage to the sealed discharge platform. The discharge platform has electrodes for discharge and voltage sensors, high-frequency coils, ultrasonic sensors, pressure sensors, and ultra-high frequency sensors for measuring discharge parameters.
[0041] In a preferred embodiment of the lightweight identification method for arc discharge modes in oil based on DSCNN, the short-gap arc fault models tested by the discharge platform are divided into needle-plate discharge, surface discharge, and inter-turn discharge.
[0042] In a preferred embodiment of the lightweight identification method for arc discharge modes in oil based on DSCNN, in step 2, voltage, current, ultrasound, oil pressure, and ultra-high frequency time-domain signals from the arc discharge test experiment are collected, and corresponding time-frequency spectra are generated through wavelet transform.
[0043] In a preferred embodiment of the lightweight identification method for arc discharge patterns in oil based on DSCNN, a lightweight deep separable convolutional neural network model is trained using a training set. The input layer compresses the time-frequency spectrum to 64 pixels × 64 pixels before feeding it to the convolutional layer. The convolutional layer first captures the global structure and preliminary features of the time-frequency spectrum through convolutional kernels, then extracts deep features through four DSCNN modules, performs downsampling through pooling, and concatenates tensors using the torch.cat() method. The obtained features are then fed to a fully connected layer. Two fully connected layers classify the extracted features, and the classification results are fed to the output layer. The output layer uses the Softmax function to output the predicted probability, thus achieving the identification task.
[0044] In a preferred embodiment of the lightweight identification method for arc discharge patterns in oil based on DSCNN, the network structure of the conventional convolutional neural network model trained on the training set consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer interconnected with each other.
[0045] In a preferred embodiment of the lightweight identification method for arc discharge patterns in oil based on DSCNN, the first training parameter and the second training parameter each include identification accuracy, parameter size, computational load, model size, and inference time.
[0046] In a preferred embodiment of the lightweight identification method for arc discharge patterns in oil based on DSCNN, after training, if the identification accuracy of the test set is greater than 80%, the training parameters are recorded; if the identification accuracy of the test set is less than 80%, the learning rate is manually modified and retraining is performed.
[0047] In a preferred embodiment of the lightweight identification method for arc discharge modes in oil based on DSCNN, the algorithm for generating time-frequency spectra from time-domain signals through wavelet transform is as follows:
[0048] ,in, Represents time-domain signal In time and scale wavelet transform value at, It is the mother wavelet function. For the complex conjugate of the mother wavelet function, It is a scale parameter. It is the time shift parameter.
[0049] In a preferred embodiment of the lightweight identification method for arc discharge modes in oil based on DSCNN, in step 3, the regularization process selects the batch normalization module to perform normalization and regularization processing.
[0050] In a preferred embodiment of the lightweight identification method for arc discharge modes in oil based on DSCNN, in step 3, the activation function is the ReLU6 function, and the expression of the ReLU6 function is: .
[0051] Where x is the output value of the activation function, and the max() and min() functions are used to find the maximum and minimum values of the sequence, respectively.
[0052] In a preferred embodiment of the lightweight identification method for arc discharge modes in oil based on DSCNN, the ReLU6 function restricts the output value by truncating values greater than 6.
[0053] In one embodiment, such as Figure 1 As shown, the lightweight identification method for arc discharge modes in oil based on DSCNN includes,
[0054] Step 1: As Figure 2 As shown, a short-gap oil arc discharge test experiment was conducted.
[0055] Step 2: Use sensors to collect discharge time-domain data, perform wavelet transform on the time-domain signal, and generate a time-frequency spectrum;
[0056] Step 3: Divide the time-frequency spectrum into a training set and a test set. Use the training set to train a lightweight deep separable convolutional neural network model, and use the test set to test the trained model and record the training parameters.
[0057] Step 4: Train a regular convolutional neural network model using the same training set as in Step 3, test the trained model using the same test set as in Step 3, and record the training parameters;
[0058] Step 5: By comparing the training parameters obtained in Step 3 and Step 4, the lightweight advantage of the short-gap arc discharge pattern recognition method based on DSCNN is confirmed.
[0059] In one embodiment, the short gap is in the millimeter range in length.
[0060] In one embodiment, the weight parameters with the minimum loss value are solved using stochastic gradient descent, and the optimal values of the learning rate, iteration rounds, and network hyperparameters are automatically selected through cross-validation and grid search, ensuring that the model converges without consuming excessive computational resources.
[0061] In one embodiment, the expression for the ReLU6 function is: The ReLU6 function restricts the output value, truncating values greater than 6. This helps prevent excessively large activation values and facilitates model deployment on low-precision platforms. Figure 5 As shown, the lightweight deep separable convolutional neural network model has the following learning process: compressing the time-frequency spectrum size to 64 pixels × 64 pixels; quickly capturing the global structure and preliminary features of the time-frequency spectrum through a regular convolutional kernel; further extracting deep features through four DSCNN modules; downsampling through pooling layers; tensor concatenation using the torch.cat() method; classification of the extracted features through two fully connected layers; and outputting the predicted probability using the Softmax function to achieve the recognition task.
[0062] Specifically, the network structure parameters are shown in the table.
[0063]
[0064] The first recorded training parameters mentioned in step 3 include recognition accuracy, parameter size, computational cost, model size, and inference time. In step 5, by comparing the training parameters of the lightweight deep separable convolutional neural network model and the conventional convolutional neural network model, it is confirmed that the lightweight deep separable convolutional neural network model has a smaller parameter size, less computational cost, and shorter inference time when performing short-gap oil arc discharge pattern recognition, which can meet the requirements of online monitoring.
[0065] Specifically, assume the input feature map has M channels and the convolution kernel size is D. k If the number of channels in the output feature map is N, then the formula for calculating the number of parameters in the depthwise convolution is:
[0066]
[0067] The formula for calculating the number of parameters in the pointwise convolution is:
[0068]
[0069] The total number of parameters in the lightweight, deepest separable convolutional neural network model is:
[0070]
[0071] While maintaining the same receptive field and kernel size, the number of parameters required for a conventional convolutional neural network model is:
[0072]
[0073] The ratio of the total number of parameters in the lightweight depthwise separable convolutional neural network model to the number of parameters required by the conventional convolutional neural network model can be expressed as:
[0074]
[0075] Due to N, D k All are positive integers, and the ratio Q is a number that is always less than 1.
[0076] Specifically, the values of the remaining training parameters are shown in the table.
[0077]
[0078] This invention significantly reduces the number of parameters and model complexity while maintaining performance, making it more suitable for resource-constrained environments such as mobile devices or embedded systems. By employing deep separable convolutional neural networks and improved activation functions, the data processing is lightweighted, improving inference speed and reducing the computational requirements of devices while ensuring high recognition accuracy. It holds promise for applications in arc discharge fault identification and online monitoring.
[0079] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.
Claims
1. A lightweight method for identifying arc discharge modes in oil based on DSCNN, characterized in that, Includes the following steps: Step 1: Conduct short-gap oil arc discharge test experiments on oil-immersed power transformers; Step 2: Use sensors to collect the discharge time-domain data of the arc discharge test experiment, perform wavelet transform on it, and generate a time-frequency spectrum; Step 3: Divide the time-frequency spectrum into a training set and a test set. Use the training set to train a lightweight deep separable convolutional neural network model. The lightweight deep separable convolutional neural network model decomposes the standard convolution operation into two parts: depthwise convolution and pointwise convolution. Regularization and activation functions are added after both depthwise convolution and pointwise convolution to form a DSCNN module. Use the test set to test the trained model and record the first training parameters. Step 4: Train a regular convolutional neural network model using the training set, test the trained model using the test set, and record the second training parameters. The convolutional neural network model includes convolution, pooling, and fully connected operations. Step 5: By comparing the first training parameters and the second training parameters, the lightweight deep separable convolutional neural network model has smaller parameter size, less computation, and shorter inference time when carrying out short-gap arc discharge pattern recognition in oil. This confirms the lightweight advantage of the DSCNN-based short-gap arc discharge pattern recognition method and meets the requirements of online monitoring.
2. The lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, Preferably, conducting short-gap arc discharge tests on oil-immersed power transformers involves building an oil-immersed arc discharge test platform. An AC power supply is regulated by a voltage regulator to apply a controllable voltage to the sealed discharge platform. The discharge platform has electrodes for discharge and voltage sensors, high-frequency coils, ultrasonic sensors, pressure sensors, and ultra-high-frequency sensors to measure discharge parameters. The short-gap arc fault models tested by the discharge platform are categorized into needle-plate discharge, surface discharge, and inter-turn discharge.
3. The lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, In step 2, voltage, current, ultrasonic, oil pressure, and ultra-high frequency time-domain signals of the arc discharge test experiment are collected, and corresponding time-frequency spectra are generated through wavelet transform.
4. The lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, In training a lightweight deep separable convolutional neural network model using a training set, the input layer compresses the time-frequency spectrum to 64 pixels × 64 pixels before feeding it to the convolutional layer. The convolutional layer first captures the global structure and preliminary features of the time-frequency spectrum through convolutional kernels, then extracts deep features through four DSCNN modules, performs downsampling through pooling operations, and concatenates tensors using the torch.cat() method. The obtained features are then fed to fully connected layers. Two fully connected layers classify the extracted features, and the classification results are fed to the output layer. The output layer uses the Softmax function to output the predicted probability, thus achieving the recognition task.
5. A lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, In training a conventional convolutional neural network model using a training set, the network structure consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer that are interconnected. The first training parameter and the second training parameter each include recognition accuracy, parameter size, computational cost, model size, and inference time, respectively.
6. The lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, After training, if the accuracy of the test set recognition is greater than 80%, the training parameters are recorded; if the accuracy of the test set recognition is less than 80%, the learning rate is manually adjusted and retraining is performed.
7. A lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, The algorithm for generating time-frequency spectra from time-domain signals using wavelet transform is as follows: ,in, Represents time-domain signal In time and scale wavelet transform value at, It is the mother wavelet function. For the complex conjugate of the mother wavelet function, It is a scale parameter. It is the time shift parameter.
8. A lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, In step 3, the regularization process selects the batch normalization module to perform normalization and regularization.
9. A lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 1, characterized in that, In step 3, the activation function is the ReLU6 function, and the expression for the ReLU6 function is: , where x is the output value of the activation function, and max() and min() functions are used to find the maximum and minimum values of the sequence, respectively.
10. A lightweight identification method for arc discharge modes in oil based on DSCNN according to claim 9, characterized in that, The ReLU6 function restricts the output value, truncating values greater than 6.