Three-phase cable partial discharge detection method of deep residual network based on space-time information

Abstract

The invention provides a three-phase cable partial discharge detection method based on a deep residual network of spatio-temporal information, and the method comprises the steps: a voltage signal obtaining and preprocessing module collects a voltage signal sample of a cable, reduces the noise, and eliminates a low-frequency waveform; the local maximum value detection module is used for identifying a local maximum value index in each flattened signal; the signal splicing module flattens Np pulses with the length of w extracted from each sample into a form of (1, Np * w), and splices the Np pulses with the length of w with other two-phase signals into a form of (3, N * pw); and the partial discharge detection network module is used for outputting the probability of whether the cable has partial discharge or not by combining a deep separable convolution residual network with the channel mixing layer. According to the method, the waveform of the three-phase signal is sent into the partial discharge detection network to obtain a result whether the three-phase medium-voltage cable has partial discharge, spatial-temporal information coupling sensing is realized by using the channel mixing layer, and the accuracy and robustness of partial discharge detection of the three-phase medium-voltage cable are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge detection of cables, and in particular to a three-phase cable partial discharge detection method based on a deep residual network of spatiotemporal information. Background Art

[0002] Electricity plays a vital role in modern society, requiring strong power infrastructure and effective system maintenance to ensure stable operation; cable partial discharge detection is an important part of power system maintenance, because partial discharge may cause insulation material damage, which in turn causes cable failure and serious power outages. Accurate detection and assessment of cable partial discharge is essential to prevent these potential risks.

[0003] Existing technology and its shortcomings: 1. Complex preprocessing: Traditional methods usually use very computationally intensive preprocessing algorithms, such as wavelet transform, Hilbert transform, etc., to process raw data; these preprocessing steps not only increase the complexity and cost of the system, but also make it difficult to achieve fast processing, limiting the ability of real-time monitoring.

[0004] 2. Difficulty in designing relevant features: Traditional methods often fail to grasp the key point of partial discharge detection, that is, detecting partial discharge pulses. In the process of designing and extracting features, some key information may be ignored, resulting in difficulty in feature extraction and accompanied by information loss, thus affecting detection accuracy.

[0005] 3. Large number of model parameters: Traditional partial discharge detection models usually rely on complex machine learning or statistical methods, which contain a large number of parameters (such as weights and biases in neural networks). A large number of model parameters means high time cost for training and inference, and also increases the risk of overfitting.

[0006] 4. The models have poor generalization capabilities, which makes it difficult for them to detect partial discharge phenomena efficiently and accurately when faced with new data.

[0007] Therefore, the prior art has deficiencies and needs further improvement. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a three-phase cable partial discharge detection system and method based on a deep residual network of spatiotemporal information.

[0009] To achieve the above object, the specific scheme of the present invention is as follows: The present invention provides a three-phase cable partial discharge detection method based on a deep residual network of spatiotemporal information. The method directly extracts key features from data through automated learning, thereby avoiding the cumbersome manual feature extraction process in traditional methods. The method is designed to include the following modules: The voltage signal acquisition and preprocessing module is used to collect voltage signal samples of the cable and apply the exponentially weighted moving average method to each signal to reduce the noise of the original sequence and eliminate low-frequency waveforms; A local maximum detection module is used to identify the local maximum index in each signal after flattening, and select Np representative local maxima and their surrounding waveforms from these indexes; A signal splicing module is used to flatten the Np pulses of length w extracted from each sample into the form of (1, Np*w), and splice them with other two-phase signals into the form of (3, Np*w); Partial discharge detection network module, the network contains three residual blocks, each residual block consists of three layers of convolution, batch normalization layer and ReLU activation function, and shortcut connection is used to ensure the effective transmission of information and gradient; in addition, the network also integrates a channel mixing layer, all two-dimensional convolution layers are replaced by depthwise separable convolution to improve computational efficiency, and feature compression is performed through a global average pooling layer; finally, the network uses the Sigmoid activation function to output the probability of whether there is partial discharge in the cable.

[0010] Furthermore, the residual block in the partial discharge detection network module includes three convolutional layers, each of which is connected to a batch normalization layer and a ReLU activation function, and shortcut connections are used to ensure that the gradient flows directly through the connection, thereby improving the training efficiency of the deep network.

[0011] Furthermore, the partial discharge detection network module also includes a channel mixing layer, which uses a convolution operation with a kernel size of (3,1) to achieve effective coupling of time and space information of the three-phase cable signal.

[0012] Furthermore, all two-dimensional convolutional layers of the partial discharge detection network module are replaced with depthwise separable convolutional layers to reduce the number of model parameters and computational complexity while maintaining network performance.

[0013] Furthermore, the partial discharge detection network module is provided with a global average pooling layer before the output layer, which is used to further compress the feature dimension, reduce the risk of overfitting, and improve the generalization ability of the model.

[0014] The partial discharge detection method comprises the following steps: S1, obtains voltage signal samples of the cable and preprocesses each signal, including applying an exponentially weighted moving average method to reduce noise and remove low-frequency components; S2, perform local maximum detection on the preprocessed signal, determine a set of local maximum indexes, sort by absolute value and select the first Np local maxima, and then extract the waveforms around these maxima to form an array of (Np, w); S3, flattening and concatenating the pulse data extracted above into a (3, Np*w) format; S4, builds a partial discharge detection network, which contains multiple residual blocks, each of which consists of a depth-separable convolutional layer, a batch normalization layer, and a ReLU activation function, and introduces a channel mixing layer to enhance the spatiotemporal information coupling capability; S5, using binary cross entropy as a loss function to train the partial discharge detection network, and adopting a learning rate decay strategy and a 5-fold cross validation method to improve model stability and generalization ability; S6, inputting the cable voltage signal to be detected into the trained partial discharge detection network to obtain the probability result of whether partial discharge occurs in the cable.

[0015] Furthermore, in step S4, each residual block in the partial discharge detection network includes at least three depth-separable convolutional layers, each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and shortcut connections are used to ensure efficient information transfer.

[0016] Furthermore, in step S4, a channel mixing layer is introduced when constructing the partial discharge detection network. The channel mixing layer uses a convolution kernel of size (3, 1) to better capture the spatiotemporal relationship between the three-phase cable signals.

[0017] Furthermore, in step S4, during the training process, the partial discharge detection network is optimized using an ADAM optimizer and a learning rate decay strategy is implemented to accelerate the convergence speed and improve the sparse gradient problem.

[0018] Furthermore, in step S4, the partial discharge detection network uses a global average pooling layer before final classification to reduce the number of features and help prevent overfitting.

[0019] The technical solution of the present invention has the following beneficial effects: 1. Automatic feature extraction: The local discharge features are automatically extracted from the cable signals through deep learning methods, which reduces the dependence on manual feature engineering and improves the automation level of the detection process.

[0020] 2. Comprehensive feature capture: By utilizing multi-scale convolution and global average pooling techniques, comprehensive capture and aggregation of local discharge features can be achieved, thus enhancing the model's ability to detect local discharges of different modes.

[0021] 3. Spatiotemporal information coupling: A channel mixing layer is introduced, and a convolution operation with a kernel size of (3,1) is used to achieve effective coupling of the temporal and spatial information of the three-phase cable signal, which significantly improves the accuracy and robustness of partial discharge detection in three-phase medium voltage cables.

[0022] 4. Simplify the detection process: By flattening the cable voltage signal, extracting the pulses and splicing them into a unified format and then sending them to the partial discharge detection network, the entire detection process is simplified and the detection results are more intuitive and easy to understand.

[0023] 5. Reduce the number of parameters and computational complexity: Replacing all two-dimensional convolutional layers with depth-wise separable convolutions effectively reduces the number of parameters and computational complexity of the model.

[0024] While maintaining network performance, the model is made more suitable for deployment in resource-constrained environments, reducing hardware requirements and operating costs.

[0025] 6. Enhance model generalization ability: The global average pooling layer is used to further compress the feature dimension, reduce the risk of overfitting, and improve the model's ability to adapt to new data.

[0026] A learning rate decay strategy and a 5-fold cross-validation method were implemented to improve model stability and generalization ability and ensure reliable performance in different scenarios.

[0027] 7. High detection accuracy: Compared with other existing methods (such as the Kaggle VSB competition winner model, CNN+LSTM hybrid model, etc.), the invented algorithm shows higher detection accuracy on MCC (Matthews correlation coefficient), proving its superiority in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a partial discharge inspection flow chart of the present invention; Figure 2 is a pulse extraction flow chart of the present invention; Figure 3 It is a network structure diagram of the deep residual network model based on spatiotemporal information of the present invention; Figure 4 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments; it can be understood that the specific embodiments described herein are only used to explain the present invention rather than to limit the present invention; it should also be noted that, for ease of description, only the parts related to the present invention rather than all are shown in the accompanying drawings.

[0030] Combination Figure 1-Figure 4 As shown, the present invention provides a three-phase cable discharge detection method based on a deep residual network of spatiotemporal information. The method directly extracts key features from data through automated learning, thereby avoiding the cumbersome manual feature extraction process in traditional methods. The method is designed to include the following modules: The voltage signal acquisition and preprocessing module is used to collect voltage signal samples of the cable and apply the exponentially weighted moving average method to each signal to reduce the noise of the original sequence and eliminate low-frequency waveforms; A local maximum detection module is used to identify the local maximum index in each signal after flattening, and select Np representative local maxima and their surrounding waveforms from these indexes; A signal splicing module is used to flatten the Np pulses of length w extracted from each sample into the form of (1, Npw), and splice them with other two-phase signals into the form of (3, Npw); Partial discharge detection network module, the network contains three residual blocks, each residual block consists of three layers of convolution, batch normalization layer and ReLU activation function, and shortcut connection is used to ensure the effective transmission of information and gradient; in addition, the network also integrates a channel mixing layer, all two-dimensional convolution layers are replaced by depthwise separable convolution to improve computational efficiency, and feature compression is performed through a global average pooling layer; finally, the network uses the Sigmoid activation function to output the probability of whether there is partial discharge in the cable.

[0031] The residual block in the partial discharge detection network module includes at least three convolutional layers, each of which is connected with a batch normalization layer and a ReLU activation function, and a shortcut connection is used to ensure that the gradient flows directly through the connection, thereby improving the training efficiency of the deep network.

[0032] The partial discharge detection network module also includes a channel mixing layer, which uses a convolution operation with a kernel size of (3,1) to achieve effective coupling of the temporal and spatial information of the three-phase cable signal.

[0033] All two-dimensional convolutional layers of the partial discharge detection network module are replaced with depthwise separable convolutional layers to reduce the number of model parameters and computational complexity while maintaining network performance.

[0034] The local discharge detection network module is provided with a global average pooling layer before the output layer, which is used to further compress the feature dimension, reduce the risk of overfitting, and improve the generalization ability of the model.

[0035] A three-phase cable discharge detection method based on a deep residual network of spatiotemporal information, the method comprising the following steps: S1, obtains voltage signal samples of the cable and preprocesses each signal, including applying an exponentially weighted moving average method to reduce noise and remove low-frequency components; S2, perform local maximum detection on the preprocessed signal, determine a set of local maximum indexes, sort by absolute value and select the first Np local maxima, and then extract the waveforms around these maxima to form an array of (Np, w); S3, flattening and concatenating the pulse data extracted above into a (3, Np*w) format; S4, builds a partial discharge detection network, which contains multiple residual blocks, each of which consists of a depth-separable convolutional layer, a batch normalization layer, and a ReLU activation function, and introduces a channel mixing layer to enhance the spatiotemporal information coupling capability; S5, using binary cross entropy as a loss function to train the partial discharge detection network, and adopting a learning rate decay strategy and a 5-fold cross validation method to improve model stability and generalization ability; S6, inputting the cable voltage signal to be detected into the trained partial discharge detection network to obtain the probability result of whether partial discharge occurs in the cable.

[0036] In step S4, each residual block in the partial discharge detection network includes at least three depth-separable convolutional layers, each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and shortcut connections are used to ensure efficient information transmission.

[0037] In step S4, a channel mixing layer is introduced when constructing the partial discharge detection network. The channel mixing layer uses a convolution kernel of size (3, 1) to better capture the spatiotemporal relationship between the three-phase cable signals.

[0038] In step S4, during the training process, the partial discharge detection network is optimized using the ADAM optimizer and a learning rate decay strategy is implemented to accelerate the convergence speed and improve the sparse gradient problem.

[0039] In step S4, the partial discharge detection network uses a global average pooling layer before final classification to reduce the number of features and help prevent overfitting.

[0040] Example: The present invention proposes a three-phase cable partial discharge detection method based on a deep residual network of spatiotemporal information, and the specific steps are as follows: 1. Voltage signal acquisition and preprocessing: Get the voltage signal samples of the cable, each sample includes three signals, each signal contains m sampling points. Apply the exponentially weighted moving average method to each signal to reduce the noise in the original sequence and eliminate the low-frequency waveform.

[0041] 2. Local maximum detection: Perform local maximum detection on each flattened signal to obtain a set of local maximum indexes. Sort the voltage absolute values ​​corresponding to these indexes from large to small, and filter out Np local maxima. Then extract the waveforms around these Np local maxima with a window size of w to form an array of (Np, w).

[0042] 3. Signal splicing: Flatten the Np pulses of length w extracted from each sample into the form of (1, Np*w) and splice them with other two-phase signals into the form of (3, Np*w).

[0043] 4. Establishment and training of partial discharge detection network: When constructing the partial discharge detection network, an innovative architecture of residual neural network combined with channel mixing layer is adopted to optimize the feature extraction of three-phase cable signals. In the network design, the input signal passes through three residual blocks, each of which contains three convolutional layers, batch normalization layer and ReLU activation function, and shortcut connection is used to ensure the effective transmission of information and gradient. The key improvement is the introduction of channel mixing layer, which realizes the coupling of spatiotemporal information through the kernel size of (3,1), significantly improving the feature expression ability. In addition, all two-dimensional convolutional layers are replaced with depthwise separable convolution. This optimization reduces the number of parameters and computational complexity while maintaining network performance, making the network more suitable for deployment in resource-constrained environments. Finally, the output of the channel mixing layer is feature compressed through the global average pooling layer, which further reduces the number of parameters, effectively prevents overfitting, and enhances the generalization ability of the model.

[0044] 5. Model training: In order to deal with the stagnation of the validation set performance, a learning rate decay strategy was implemented, and the stability of the model was enhanced through 5-fold cross validation. After determining the optimal threshold for each model, its average value was calculated to form the final decision basis. Finally, the probability of the network output was compared with this average threshold to obtain the final classification result.

[0045] 6. Loss function: The partial discharge detection network uses binary cross entropy as the loss function, and its calculation formula is as follows: Loss=−(ylog(p)+(1−y)log(1−p)) Where log represents the natural logarithm; y is a binary label (0 or 1), indicating the sample labels without partial discharge and with partial discharge; p represents the probability prediction of the PD network for samples with PD.

[0046] The present invention has the following beneficial effects: Automatic feature extraction: The local discharge features in the cable signal are automatically extracted through deep learning methods, reducing the dependence on manual feature extraction.

[0047] Comprehensive feature capture: Through multi-scale convolution and global average pooling, comprehensive capture and aggregation of local discharge features are achieved, enhancing the detection capability of the model.

[0048] Spatiotemporal information coupling: The channel mixing layer is used to realize spatiotemporal information coupling perception, which significantly improves the accuracy and robustness of partial discharge detection in three-phase medium-voltage cables.

[0049] Simplify the detection process: by flattening the voltage signal of the cable, extracting the pulses and splicing them in turn, the waveform of the three-phase signal is sent to the partial discharge detection network to obtain the result of whether the three-phase medium voltage cable has partial discharge.

[0050] Reduced parameters and computational complexity: All two-dimensional convolutional layers are replaced with depth-wise separable convolutions. This optimization reduces the number of parameters and computational complexity while maintaining network performance, making the network more suitable for deployment in resource-constrained environments.

[0051] The specific steps are as follows: Step 1: Get a cable voltage signal sample The voltage signal samples of the cable are obtained, each sample contains m sampling points. The Exponential Weighted Moving Average (EWMA) method is applied to each sample to reduce the noise in the original sequence and eliminate the low-frequency waveform. In this embodiment, the parameters of the EWMA method are set to ɑ=100, β=1, and the formula is as follows: The adjustment formula for the predicted value is:

[0052] Residual calculation formula:

[0053] The process is to initialize an all-zero array N with the same shape as the input signal i , Where p is the first element of the input signal X p i is the adjusted forecast value; α and β are tuning parameters; x i is the actual observed value; x i represents the residual, which is the difference between the actual observed value and the predicted value; For each element of X (starting from the second element), the smoothing coefficient between it and the previous element is calculated, and then the current element is updated to generate a smoothing sequence N. In this embodiment, each sample contains 800,000 points, that is, m=800,000, and the specific number of sampling points can be set according to actual needs.

[0054] Step 2: Local maximum detection Figure 2 The pulse extraction process performs local maximum detection on each sample signal after smoothing to obtain a set of local maximum indexes, and sorts the sample values ​​corresponding to these indexes from large to small to screen out Np local maxima. Then extract the waveforms around these local maxima, with a window size of w, to form an array of (Np, w). In this embodiment, the local maximum detection method extracts 200 local maxima for each sample, i.e., Np=200, and extracts the 5 time points before and 35 time points after the index with the index of these local maxima as the center, i.e., [tw / 8, t+7w / 8], where t is the index of the local maximum, and the length of the extracted pulse is 40, i.e., w=40. Finally, each sample extracts a pulse array of (200, 40), and the specific number and length of the extracted pulses can be adjusted according to actual conditions.

[0055] Step 3: Signal flattening and splicing Since the acquired signal samples are in three-phase form, that is, 3 signals are one sample, the Np pulses of length w extracted from each signal are flattened into the form of (1, Np*w) and spliced ​​with other two-phase signals into the form of (3, Np*w). In this embodiment, Np=200, w=40, and each group of samples becomes (3, 8000) after processing.

[0056] Step 4: Establishing the Partial Discharge Detection Network A partial discharge detection network is established, and the data of (3, Np*w) is used as network input and input into the detection network for training. The partial discharge detection network includes a residual neural network, a channel mixing module, and a global average pooling layer.

[0057] Network structure Figure 3The network structure of the model is shown in Figure 1. The residual neural network includes three residual modules, namely the first residual module, the second residual module and the third residual module. Each residual module includes a two-dimensional convolution layer, a BN layer and a ReLU activation function arranged in sequence. The channel mixing module includes a two-dimensional convolution layer, a BN layer and a ReLU activation function layer.

[0058] The input tensor is (k, 3, Np*w), where k represents the batch size, 3 represents the three-phase signal, and Np*w represents the length of the pulse waveform. The input tensor passes through the first residual module, the second residual module, the third residual module, and the channel mixing layer in sequence, and then is input to the global average pooling layer, and finally outputs the predicted probability through the single neuron layer.

[0059] The specific process is as follows: First, the input tensor (1,3,8000) is input to the first residual module, whose input tensor is (1, 3,8000,1) and output tensor is (1,3,8000,8). The convolution kernel of the first layer is (1,8), padded with 1, and passes through the BN layer and ReLU activation function. The convolution kernels of the second and third layers are (1,5) and (1,3), respectively, and the rest of the structure is the same as the first layer. The kernel size of the residual connection layer is (1,1), and the input tensor form is (1,3,8000,1), which is added to the output result of the third convolution layer and input to the next residual module. Residual connection allows the gradient to flow directly through the connection, reducing the gradient vanishing effect, making the network deeper, and improving feature extraction capabilities.

[0060] The structures of the second and third residual modules are the same as the first residual module, except for the number of convolution kernels. The first residual module has 8 convolution kernels, and the second and third residual modules have 16 convolution kernels respectively. The output of the third residual module is (1,3,8000, 16), which is the input of the channel mixing layer. The shape of the tensor after passing through the channel mixing layer is (1,1,8000,8), and the shape of the convolution kernel is (3,1), and the number is 8. The channel mixing module can fully integrate the coupling information of the three-phase signal and improve the detection robustness.

[0061] The output of the channel mixing layer (1,1,8000,8) is input into the global average pooling layer, and the output is (1,8). The global average pooling layer can replace the fully connected layer, reduce the number of model parameters, reduce computational complexity, and effectively prevent overfitting.

[0062] Finally, the output of the global average pooling layer (1, 8) is input into the single neuron layer to obtain the classification probability.

[0063] Training and Validation This embodiment uses 5-fold cross validation to determine the classification threshold. Specifically, the 2904 groups of samples are divided into 5 parts, 4 of which are used as training sets and 1 as validation sets each time, and the cycle is repeated 5 times to obtain 5 models. After determining the optimal threshold of each model, the average threshold is calculated.

[0064] Loss Function and Optimization The partial discharge detection network uses binary cross entropy as the loss function, and the calculation formula is as follows:

[0065] Among them, y is a binary label, 0 indicates no partial discharge sample, 1 indicates a partial discharge sample, and p is the probability of a partial discharge sample output by the detection network.

[0066] This embodiment uses the ADAM optimizer to optimize the loss function. ADAM (Adaptive Moment Estimation) combines the ideas of momentum method and adaptive learning rate to accelerate the convergence of the model and effectively handle sparse gradients.

[0067] Finally, the tensor corresponding to the voltage signal of the cable to be detected is input into the trained partial discharge detection network, and two probabilities of whether the three-phase medium-voltage cable to be detected has a partial discharge fault are output. The final detection result is determined by combining the average threshold.

[0068] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0069] The above-described embodiments only express the more specific and detailed embodiments described in this application, but they cannot be understood as limiting the scope of the patent application. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of this application, which all belong to the protection scope of this patent application. Therefore, the protection scope of the patent application shall be based on the attached claims.

[0070] Comparison chart:

[0071] (MCC is a correlation coefficient that shows the difference between the true category of the sample and the category predicted by the algorithm. Its value is between -1 and 1: the larger the value, the more accurate the algorithm prediction; when the value is 0, the prediction effect is equivalent to random guessing; when the value is -1, the prediction is completely opposite to the true category. The reason why this algorithm achieves such good results is mainly due to comprehensive feature capture: through multi-scale convolution and global average pooling, comprehensive capture and aggregation of partial discharge features are achieved, and the detection ability of the model is enhanced. Spatiotemporal information coupling: The channel mixing layer is used to realize spatiotemporal information coupling perception, which significantly improves the accuracy and robustness of partial discharge detection of three-phase medium voltage cables.) How it works 1. Voltage signal acquisition and preprocessing: Steps: The system first obtains three-phase voltage signal samples of the cable, each sample includes three signals, and each signal contains m sampling points.

[0072] Preprocessing: Apply the Exponential Weighted Moving Average (EWMA) method to each signal to reduce the noise in the original sequence and eliminate the low-frequency waveform. In the embodiment, the parameters of the EWMA method are set to α=100 and β=1.

[0073] 2. Local maximum detection: Steps: Perform local maximum detection on each smoothed signal to obtain a set of local maximum indexes.

[0074] Screening and extraction: Sort the absolute voltage values ​​corresponding to these indexes from large to small, and screen out Np local maximum values. Then, with these local maximum values ​​as the center, extract the waveforms with a window size of w around them to form an array of (Np, w).

[0075] 3. Signal splicing: Flattening and splicing: Flatten the Np pulses of length w extracted from each sample into the form of (1, Np*w) and splice them with the other two-phase signals into the form of (3, Np*w). This step enables the three-phase signal to be input as a whole into the subsequent deep learning model.

[0076] 4. Establishment and training of partial discharge detection network: Network architecture: The innovative architecture of residual neural network combined with channel mixing layer is used to optimize the feature extraction of three-phase cable signals. In the network design, the input signal passes through three residual blocks, each of which contains three convolutional layers, batch normalization layers and ReLU activation functions, and uses shortcut connections to ensure the effective transmission of information and gradients.

[0077] Key improvements: The channel mixing layer is introduced to achieve the coupling of spatiotemporal information through the (3,1) kernel size, which improves the feature expression ability. In addition, all two-dimensional convolutional layers are replaced with depthwise separable convolutions, reducing the number of parameters and computational complexity.

[0078] Feature compression: Finally, the output of the channel mixing layer is feature compressed through the global average pooling layer, which further reduces the number of parameters and effectively prevents overfitting.

[0079] 5. Model training: Strategy: In order to improve model performance, a learning rate decay strategy was implemented, and 5-fold cross validation was used to enhance the stability of the model.

[0080] Threshold determination: After determining the optimal threshold for each model, calculate its average value to form the final decision basis. The probability of the network output is compared with this average threshold to obtain the final classification result.

[0081] 6. Loss Function: Definition: The partial discharge detection network uses binary cross entropy as the loss function to measure the difference between the model prediction value and the true label, thereby guiding model optimization.

[0082] Summarize This invention makes full use of the temporal and spatial information of cable partial discharge signals by combining deep learning technology, especially the deep separable convolutional residual network. The entire process from signal acquisition, preprocessing, feature extraction to the final classification decision forms a complete closed loop, which significantly improves the accuracy and robustness of partial discharge detection of three-phase medium voltage cables.

[0083] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. All equivalent structural changes made by using the contents of the present invention specification and drawings under the inventive concept of the present invention, or directly / indirectly applied in other related technical fields are included in the protection scope of the present invention.

Claims

1. A three-phase cable partial discharge detection method based on a deep residual network of spatiotemporal information, characterized in that: This method directly extracts key features from data through automated learning, thus avoiding the tedious manual feature extraction process in traditional methods. The modules involved in this method include: The voltage signal acquisition and preprocessing module is used to collect voltage signal samples of the cable and apply the exponentially weighted moving average method to each signal to reduce the noise of the original sequence and eliminate low-frequency waveforms; A local maximum detection module is used to identify the local maximum index in each signal after flattening, and select Np representative local maxima and their surrounding waveforms from these indexes; A signal splicing module is used to flatten the Np pulses of length w extracted from each sample into the form of (1, Np*w), and splice them with other two-phase signals into the form of (3, Np*w); Partial discharge detection network module, the network contains three residual blocks, each residual block consists of three layers of convolution, batch normalization layer and ReLU activation function, and shortcut connection is used to ensure the effective transmission of information and gradient; in addition, the network also integrates a channel mixing layer, all two-dimensional convolution layers are replaced by depthwise separable convolution to improve computational efficiency, and feature compression is performed through a global average pooling layer; finally, the network uses the Sigmoid activation function to output the probability of whether there is partial discharge in the cable.

2. The detection method according to claim 1, characterized in that: The residual block in the partial discharge detection network module includes three convolutional layers, each of which is connected to a batch normalization layer and a ReLU activation function, and shortcut connections are used to ensure that the gradient flows directly through the connection, thereby improving the training efficiency of the deep network.

3. The detection method according to claim 1, characterized in that: The partial discharge detection network module also includes a channel mixing layer, which uses a convolution operation with a kernel size of (3,1) to achieve effective coupling of the temporal and spatial information of the three-phase cable signal.

4. The detection method according to claim 1, characterized in that: All two-dimensional convolutional layers of the partial discharge detection network module are replaced with depthwise separable convolutional layers to reduce the number of model parameters and computational complexity while maintaining network performance.

5. The detection method according to claim 1, characterized in that: The method comprises the following steps: S1, obtains voltage signal samples of the cable and preprocesses each signal, including applying an exponentially weighted moving average method to reduce noise and remove low-frequency components; S2, perform local maximum detection on the preprocessed signal, determine a set of local maximum indexes, sort by absolute value and select the first Np local maxima, and then extract the waveforms around these maxima to form an array of (Np, w); S3, flattening and concatenating the pulse data extracted above into a (3, Np*w) format; S4, builds a partial discharge detection network, which contains multiple residual blocks, each of which consists of a depth-separable convolutional layer, a batch normalization layer, and a ReLU activation function, and introduces a channel mixing layer to enhance the spatiotemporal information coupling capability; S5, using binary cross entropy as a loss function to train the partial discharge detection network, and adopting a learning rate decay strategy and a 5-fold cross validation method to improve model stability and generalization ability; S6, inputting the cable voltage signal to be detected into the trained partial discharge detection network to obtain the probability result of whether partial discharge occurs in the cable.

6. The detection method according to claim 5, characterized in that: In step S4, each residual block in the partial discharge detection network includes at least three depth-separable convolutional layers, each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and shortcut connections are used to ensure efficient information transmission.

7. The detection method according to claim 5, characterized in that: In step S4, a channel mixing layer is introduced when constructing the partial discharge detection network. The channel mixing layer uses a convolution kernel of size (3, 1) to better capture the spatiotemporal relationship between the three-phase cable signals.

8. The detection method according to claim 5, characterized in that: In step S4, during the training process, the partial discharge detection network is optimized using the ADAM optimizer and a learning rate decay strategy is implemented to accelerate the convergence speed and improve the sparse gradient problem.

9. The detection method according to claim 5, characterized in that: In step S4, the partial discharge detection network uses a global average pooling layer before final classification to reduce the number of features and help prevent overfitting.

Images