Overhead transmission line fault detection and classification method based on discrete wavelet transform and time sequence convolutional network

Abstract

The invention discloses an overhead transmission line fault detection and classification method based on discrete wavelet transform and a time sequence convolutional network, and relates to the technical field of electronics. The method comprises the following steps: acquiring original signals of three-phase current and grounding current of the overhead transmission line, and carrying out normalization processing and moving average filtering preprocessing; discrete wavelet transform (DWT) of a Daubechies wavelet basis function Db4 is adopted to perform multi-scale decomposition on the preprocessed signal, a high-frequency detail coefficient is extracted, and a multi-channel time sequence input vector is constructed; and inputting the input vector into a time sequence convolutional neural network (TCN). According to the method, fault features are efficiently extracted through DWT, time sequence data are accurately modeled through TCN, the problems that a traditional method is low in accuracy, slow in response and weak in anti-interference capacity in a complex fault scene are solved, the fault classification accuracy is 99.9%, the parallel computing capacity is high, training is stable, the method can adapt to different fault conditions, and the method is suitable for large-scale popularization and application. And the operation reliability and the intelligent protection level of the power system are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of electronic technology, specifically to a method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks. Background Technology

[0002] With the continuous expansion of power system scale and the continuous improvement of operating voltage levels, the safe and stable operation of transmission lines is crucial to ensuring the reliability of the entire power grid system. In high-voltage and ultra-high-voltage transmission systems, faults on overhead transmission lines (such as single-phase grounding faults, phase-to-phase short circuits, three-phase short circuits, and two-phase grounding faults) not only affect power quality but may also lead to equipment damage or even large-scale power outages. Therefore, the rapid and accurate detection and classification of various faults in transmission lines is a key link in improving the stability and reliability of power supply. Traditional methods for detecting and classifying transmission line faults mainly rely on protection principles based on changes in electrical quantities, such as impedance methods, traveling wave methods, and Fourier analysis methods. While these methods can identify common short-circuit or grounding faults to some extent, they often suffer from low sensitivity, high false positive rates, and slow response speeds when facing high-resistance grounding faults, complex and varied fault types, and external noise interference. In addition, traditional methods rely on human experience in the feature extraction process, making it difficult to adapt to the increasingly complex fault modes in modern power grids.

[0003] In recent years, with the development of artificial intelligence and big data technologies, fault detection and classification methods based on machine learning and deep learning have gradually become research hotspots. For example, Support Vector Machines (SVM), fuzzy logic, decision trees, k-nearest neighbors (k-NN), artificial neural networks (ANN), and their improved models have been widely applied in power system fault diagnosis. These methods achieve automatic identification of fault types by training on historical data, improving classification accuracy and intelligence. However, these models still have certain limitations when processing fault signals with time-series characteristics. For example, although recurrent neural networks (RNN) and long short-term memory networks (LSTM) can capture long-term dependencies in time series, their training process is complex, computationally expensive, and difficult to parallelize; convolutional neural networks (CNN) perform well in image processing, but their fixed receptive field limits their ability to model long-term series.

[0004] To overcome the aforementioned problems, this invention proposes a fault detection and classification method for overhead transmission lines based on Discrete Wavelet Transform (DWT) and Temporal Convolutional Neural Network (TCN). TCN is a deep learning model specifically designed for processing time-series data. Through causal convolution and dilated convolution mechanisms, it effectively expands the receptive field of the model while ensuring the sequentiality of the time series and possesses good parallel computing capabilities. Compared with traditional RNN / LSTM structures, TCN reduces training complexity and improves real-time performance and generalization ability while maintaining high prediction accuracy. DWT is combined with DWT for fault signal feature extraction, utilizing its multi-resolution analysis capabilities to effectively separate high-frequency transient components in the fault current signal, thereby improving the input feature quality of the subsequent classification model. This method can achieve high-accuracy fault classification under various complex conditions. Summary of the Invention

[0005] This invention aims to provide a fault detection and classification method for overhead transmission lines based on Discrete Wavelet Transform (DWT) and Temporal Convolutional Neural Networks (TCNs). It primarily addresses the problems of low accuracy, slow response speed, reliance on manual feature extraction, and difficulty in adapting to high-noise environments inherent in traditional fault detection methods under complex fault scenarios. By combining DWT for fault signal feature extraction and utilizing TCNs for efficient modeling of time-series data, this method achieves rapid and accurate identification and classification of various common transmission line fault types. Furthermore, it exhibits good robustness and generalization ability, maintaining stable performance under complex conditions such as different fault locations, transition resistances, and initial phase angles. This improves the intelligence level and operational reliability of power system protection, providing strong support for rapid fault isolation and repair.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A fault detection and classification method for overhead transmission lines based on discrete wavelet transform and temporal convolutional networks includes the following steps:

[0008] S1. High-precision current transformers (CTs) and voltage transformers (VTs) are installed along the transmission line to collect three-phase current signals Ia, Ib, Ic and grounding current signal Ig in real time; the sampling frequency is set to 20kHz to ensure that transient signals at the moment of fault occurrence can be captured, and each signal group has a length of 100,000 sampling points, corresponding to a time window of 0.1 seconds;

[0009] The acquired raw signal is preprocessed. First, a moving average filter is used to remove high-frequency noise interference, and then normalization is performed. The normalization formula is as follows: ,in I represents the normalized current signal, and I represents the original current signal. This represents the minimum value of the original current signal. The maximum value of the original current signal is used to improve the accuracy of subsequent feature extraction through preprocessing.

[0010] S2. Feature extraction based on discrete wavelet transform (DWT): The preprocessed fault current signal is decomposed into multiple scales using discrete wavelet transform (DWT), and the Daubechies wavelet basis function Db4 is selected as the mother wavelet for multi-resolution analysis (MRA).

[0011] The definition of discrete wavelet transform is: Where x(k) represents the input signal and ψ represents the mother wavelet function, Represents the scale factor. The translation parameters are represented by m, n, and k, which represent integer indices.

[0012] The signal is decomposed into multiple frequency bands by DWT to obtain approximation coefficients and detail coefficients. High-frequency detail coefficients containing fault information are extracted to form time series data.

[0013] S3. Time series input construction: The high-frequency detail coefficients extracted in step 2 are used to form a multi-channel time series input vector with an input dimension of 1×T×4, where T is the time step (T=100,000 in this embodiment), and 4 corresponds to the four channels of three-phase current Ia, Ib, Ic and ground current Ig. This input vector is used as the input data for the subsequent TCN model.

[0014] S4. Fault classification based on temporal convolutional neural network (TCN):

[0015] TCN Model Structure Construction: A TCN model with multiple residual blocks is constructed. Each residual block includes two dilated convolutional layers, a ReLU activation function, a weight normalization layer, and a Dropout layer, and each residual block is configured with residual connections. The TCN model structure consists of an input layer, multiple TCN residual blocks, a fully connected layer, and an output layer. The output layer uses a Softmax classifier to output the probability distribution of 12 fault types.

[0016] TCN core mechanism:

[0017] Causal convolution: To preserve the causal relationships in time series, one-dimensional causal convolution is used, meaning that only the input information from the current time step and earlier is used during the convolution process to avoid leaking future information; the output sequence y of causal convolution satisfies: , where x is the input sequence, f is the filter, K is the filter size, and t represents the time step index.

[0018] Dilated convolution: To capture dependencies over a longer time span, dilated convolution is used in each layer. The basic idea is to insert holes between the convolution kernels, causing the receptive field to grow exponentially. When the dilation factor is d, the convolution operation is defined as follows: , where d is the expansion factor, which controls the interval of skip sampling.

[0019] Receptive field calculation: The receptive field of a TCN determines the length of historical information the model can capture; the receptive field of a single-layer TCN Where K is the kernel size and d is the dilation factor of the current layer; if L layers of TCN residual blocks are stacked, and the dilation factor of each layer increases exponentially (e.g., ... ), then the overall sensory field .

[0020] Zero padding: To maintain consistency in the time dimension between input and output, left zero padding (Causal Padding) is performed at the input. The padding length is determined by... Decide.

[0021] Residual Connections: The output of each residual block is defined as follows: ,in This represents the transformation function after operations such as convolution and activation. As input to the residual block, this structure helps alleviate the vanishing gradient problem and accelerates convergence.

[0022] Model Training and Optimization: The TCN model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and a total of 30 epochs. The cross-entropy loss function was used. ,in It's a real label. It is the predicted probability, and N is the total number of categories; an early stopping mechanism is introduced during training to prevent overfitting, and the model performance is evaluated using metrics such as confusion matrix, precision, recall, and F1 score.

[0023] Fault classification reasoning: Input the time series input vector constructed in step 3 into the trained TCN model for forward propagation, and output the fault category prediction result at the current time through model reasoning.

[0024] S5. Result Output and Protection Action Trigger: Based on the fault classification results of the TCN model, if a fault is detected, the corresponding protection device is triggered; at the same time, fault-related information, including fault type, fault location, fault timestamp, etc., is recorded for subsequent analysis.

[0025] Working principle and beneficial effects of the present invention:

[0026] 1. Working Principle: The system collects the raw signals of three-phase currents (Ia, Ib, Ic) and grounding current (Ig) of overhead transmission lines through current transformers (CTs) and voltage transformers (VTs). After normalization and moving average filtering to eliminate noise interference and amplitude differences, the preprocessed signals are decomposed into multiple scales using the Daubechies wavelet basis function Db4 for discrete wavelet transform (DWT). High-frequency detail coefficients containing core fault information are extracted and a multi-channel time series input vector in "1×T×4" format is constructed. This vector is input into a temporal convolutional neural network (TCN) composed of residual blocks with multiple causal convolution, dilated convolution, ReLU activation function, weight normalization, and dropout layers. Leveraging the advantages of TCN in expanding the receptive field and supporting parallel computing while ensuring the causality of the time series, the fault classification results are output through model training and inference. Finally, the corresponding protection device is triggered based on the classification results, and relevant information such as fault type, location, and timestamp is recorded. The entire process achieves efficient and accurate detection and classification of faults in overhead transmission lines.

[0027] 2. Beneficial effects: (1) This invention extracts the high-frequency transient components of the fault current signal by combining discrete wavelet transform (DWT), effectively separating fault features from noise and improving the quality of input features; at the same time, it utilizes the efficient modeling capability of the TCN model for time series data to achieve high-precision identification of 12 types of faults (including fault-free state). In the simulation model of a 500kV actual transmission line, the fault classification accuracy under various fault types, different fault locations, and different transition resistance conditions is as high as 99.9%, which is significantly better than existing models such as BiLSTM and GRU, and has high-precision fault classification capability.

[0028] (2) This invention employs causal convolution and dilated convolution structures through TCN, which effectively expands the receptive field of the model while ensuring the sequential nature of the time series, enabling it to capture dependencies over a longer time range. Compared to traditional RNN / LSTM, TCN can better model long-distance dependencies, improve the ability to identify complex fault modes, and is suitable for data processing with strong time series characteristics, such as transmission line fault signals, thus possessing stronger time series modeling capabilities.

[0029] (3) Since TCN is a deep learning model built on convolutional operations, it naturally supports parallel computing, avoiding the serialization bottleneck caused by the recursive structure of RNN / LSTM when processing long sequences. This makes the method of the present invention have higher real-time performance in actual deployment, can quickly respond to fault detection requirements, and is suitable for protection devices in power systems with high response speed requirements, resulting in strong parallel computing capabilities and fast inference speed.

[0030] (4) This invention introduces residual connections through TCN, which effectively alleviates the gradient vanishing problem during deep network training and improves the training stability of the model. In addition, the combination of ReLU activation function and weight normalization further enhances the generalization ability and convergence speed of the model, reduces the training difficulty, enables the model to quickly achieve the ideal training effect, reduces the preparation time before model deployment, and makes the model structure stable and training convergence fast.

[0031] (5) The method of the present invention enhances the resistance to noise interference through data preprocessing and DWT feature extraction; the parallel structure and dilated convolution mechanism of the TCN model enable it to adapt to complex conditions such as different fault locations, transition resistance, and fault initial phase angle, and can maintain stable classification performance in a variety of complex fault scenarios. It has good robustness and generalization ability, and is suitable for the complex operating environment of actual power systems, resulting in excellent robustness and generalization ability.

[0032] Preferably, in step 1, the sensor uses high-precision current transformers (CTs) and voltage transformers (VTs), the signal sampling frequency is set to 20kHz, each signal group has a length of 100,000 sampling points, corresponding to a time window of 0.1 seconds.

[0033] Preferably, in step 1, the normalization formula is: ,in I represents the normalized current signal, and I represents the original current signal. This represents the minimum value of the original current signal. This represents the maximum value of the original current signal.

[0034] Preferably, in step 2, the discrete wavelet transform is defined as follows: Where x(k) represents the input signal and ψ represents the mother wavelet function, Represents the scale factor. The translation parameters are represented by m, n, and k, which represent integer indices.

[0035] Preferably, in step 4, the output sequence y of the causal convolution satisfies: , where x is the input sequence, f is the filter, K is the filter size, and t represents the time step index.

[0036] Preferably, in step 4, the output sequence y of the dilated convolution satisfies: , where d is the expansion factor, K is the filter size, f is the filter, and t represents the time step index.

[0037] Preferably, in step 4, the receptive field calculation of the TCN model satisfies: the receptive field of a single-layer TCN... When stacking L-layer TCN residual blocks, the total receptive field , where K is the kernel size and d is the expansion factor of the current layer, and the expansion factor increases exponentially.

[0038] Preferably, in step 4, the TCN model is trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, 30 training epochs, and the cross-entropy loss function. ,in It's a real label. It represents the predicted probability, and N is the total number of categories.

[0039] Preferably, in step 4, the residual block output of the TCN model satisfies: ,in This represents the transformation function after operations such as convolution and activation. is the input to the residual block, and ReLU is the activation function.

[0040] Preferably, in step 5, the fault-related information includes fault type, fault location, and fault timestamp. The fault type includes single-phase grounding fault, phase-to-phase short-circuit fault, three-phase short-circuit fault, two-phase grounding fault, and no-fault state, totaling 12 categories. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the overall workflow of the overhead transmission line fault detection and classification method based on discrete wavelet transform and temporal convolutional network as described in this embodiment of the invention.

[0042] Figure 2 This is a data preprocessing flowchart of the overhead transmission line fault detection and classification method based on discrete wavelet transform and temporal convolutional network described in an embodiment of the present invention;

[0043] Figure 3 This is a flowchart of the DWT feature extraction process for the overhead transmission line fault detection and classification method based on discrete wavelet transform and temporal convolutional network as described in this embodiment of the invention.

[0044] Figure 4 This is a flowchart of the TCN model inference for the overhead transmission line fault detection and classification method based on discrete wavelet transform and temporal convolutional network as described in an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example

[0047] like Figures 1-4 As shown, this embodiment provides a fault detection and classification method for overhead transmission lines based on discrete wavelet transform and temporal convolutional networks, including the following steps:

[0048] S1. Fault signal acquisition and preprocessing: The three-phase current signals Ia, Ib, Ic and grounding current signal Ig of the overhead transmission line are acquired by sensors. The acquired raw signals are normalized and filtered by moving average to eliminate noise interference.

[0049] S2. Feature extraction based on discrete wavelet transform (DWT): The Daubechies wavelet basis function Db4 is used as the mother wavelet to perform multi-scale decomposition on the preprocessed fault current signal to obtain high-frequency detail coefficients containing fault information and form time series data.

[0050] S3. Time series input construction: The high-frequency detail coefficients extracted in step 2 are used to form a multi-channel time series input vector with an input dimension of 1×T×4, where T is the time step and 4 corresponds to the four channels of three-phase current and grounding current.

[0051] S4. Fault classification based on temporal convolutional neural network (TCN): Construct a TCN model containing multiple residual blocks, wherein the residual blocks include causal convolutional layers, dilated convolutional layers, ReLU activation functions, weight normalization layers and Dropout layers. Input the time series input vector constructed in step 3 into the TCN model, and output the fault classification result through model training and inference.

[0052] S5. Result Output and Protection Action Trigger: Based on the fault classification results of the TCN model, trigger the corresponding protection device action and record fault-related information.

[0053] It should be specifically noted that the detection and classification of faults in overhead transmission lines are achieved through the collaborative work of the sensor and signal acquisition module, the data preprocessing and feature extraction module, the TCN fault classification model construction and training module, the model deployment and online inference module, and the result output and protection action control module.

[0054] Sensors and signal acquisition modules:

[0055] Sensor configuration: High-precision current transformers (CTs) and voltage transformers (VTs) are installed at key locations along the 500kV overhead transmission line. CTs are used to collect three-phase current signals Ia, Ib, Ic and grounding current signal Ig, while VTs are used to assist in monitoring voltage parameters to ensure the accuracy and comprehensiveness of signal acquisition.

[0056] Data Acquisition System: A high-speed data acquisition system (DAQ) is used to digitize the analog signals output by the sensors. The sampling frequency is strictly set to 20kHz, with 100,000 sampling points per signal group, corresponding to a 0.1-second time window, ensuring complete capture of transient signal characteristics at the moment of fault occurrence. The acquired digital signals are uploaded to edge computing devices within the substation via Ethernet or fiber optic communication interfaces to ensure real-time and reliable data transmission.

[0057] Data preprocessing and feature extraction module:

[0058] Data normalization and filtering: The raw current signal uploaded to the edge computing device is preprocessed. First, a sliding window mean filtering algorithm is used to remove high-frequency noise interference from the signal, with the sliding window size set to 5 sampling points; then, the normalization formula is applied. The filtered current signal is normalized to map the signal amplitude to the [0,1] interval, thus eliminating the influence of differences in signal amplitude.

[0059] Feature extraction based on DWT: The wavelet analysis toolkit in Matlab was used to select the Daubechies wavelet basis function Db4 to perform multi-scale decomposition on the preprocessed current signal. The number of decomposition layers was set to 4. The approximation coefficients and detail coefficients of the signal were separated by multi-resolution analysis (MRA). The high-frequency detail coefficients after 4-layer decomposition were extracted as fault features. These coefficients contain transient feature information when the fault occurs.

[0060] Time series input construction: The extracted high-frequency detail coefficients are arranged in chronological order to form a time series input vector with a dimension of 1×100,000×4, where 1 represents a single batch input, 100,000 is the time step (corresponding to the number of sampling points collected), and 4 corresponds to the signals of the four channels Ia, Ib, Ic, and Ig.

[0061] TCN Fault Classification Model Construction and Training Module:

[0062] TCN Model Structure Design: The TCN model is built based on the PyTorch deep learning framework, and its structure is as follows:

[0063] Input layer: Receives time-series input vectors with dimensions of 1×100,000×4;

[0064] TCN Residual Blocks: A total of 6 TCN residual blocks are stacked. Each residual block contains two dilated convolutional layers (kernel size K=3), a ReLU activation function, a weight normalization layer, and a Dropout layer (dropout probability set to 0.2). The dilation factor d increases exponentially, i.e., d=1,2,4,8,16,32.

[0065] Residual connections: Each residual block is set with a residual connection, which adds the input of the residual block to the output after convolution, activation and other operations, and then outputs it through the ReLU activation function to alleviate the gradient vanishing problem;

[0066] Fully connected layer: maps the high-dimensional features output by the TCN residual block to a 12-dimensional low-dimensional space, corresponding to 12 types of faults;

[0067] Output layer: The Softmax activation function is used to output the probability distribution of 12 fault types.

[0068] Model training process:

[0069] Dataset preparation: A 500kV transmission line simulation model was built in the MATLAB / Simulink environment to simulate 12 types of faults (including single-phase grounding AG, AB, AC; two-phase short circuit AB, BC, CA; two-phase grounding ABG, BCG, CAG; three-phase short circuit ABC; three-phase grounding ABCG; and no fault state). Fault data were generated under different fault locations (line start, middle, and end), different transition resistances (0Ω, 100Ω, 500Ω, 1000Ω), and different initial phase angles of the fault (0°, 90°, 180°, 270°). A total of 100,000 sets of sample data were collected and divided into training set, validation set, and test set in a ratio of 7:2:1.

[0070] Model Training: The TCN model was trained using the training set with the Adam optimizer, an initial learning rate of 0.001, a batch size of 64, and a total of 30 epochs. The cross-entropy loss function was used to measure the difference between the model's predicted values ​​and the true labels. During training, the validation set was used to monitor model performance, and an early stopping mechanism was introduced. Training was stopped when the validation set loss did not decrease for 5 consecutive epochs to prevent overfitting. After training, the model performance was evaluated using the test set, and the evaluation metrics included accuracy, precision, recall, F1 score, and confusion matrix.

[0071] Model evaluation results: The test set evaluation results show that the model achieves a classification accuracy of 99.9% for 12 types of faults, and the precision, recall and F1 score for each type of fault are all higher than 99.5%, indicating that the model has excellent classification performance.

[0072] Model Deployment and Online Inference Module:

[0073] Deployment environment: The trained TCN model is exported as ONNX format and deployed on an edge computing device (configured with an embedded GPU, model NVIDIA Jetson Xavier NX) in the substation. The edge computing device communicates with the data acquisition system in real time via Ethernet to ensure the real-time performance of data transmission and inference.

[0074] Inference process: The edge computing device receives the current signal uploaded by the data acquisition system in real time; the real-time signal is preprocessed and DWT features are extracted to construct a time series input vector; the input vector is input into the deployed TCN model for forward propagation inference, and the inference time is controlled within 10ms; the model outputs the fault category prediction result and the corresponding probability value.

[0075] Result output and protection action control module:

[0076] Results Judgment and Output: The edge computing device judges the inference results of the TCN model. If the predicted probability is greater than 0.95, it is determined to be the corresponding fault type; if the predicted probability of all fault types is less than 0.95, it is determined to be a fault-free state. At the same time, the edge computing device uploads information such as fault type, fault timestamp, and fault location (determined by the fault location mapping relationship during model training) to the power system monitoring center in real time, and stores the fault data locally for subsequent analysis.

[0077] Protection action triggering: If a fault is detected, the edge computing device sends an action command to the corresponding protection device (such as a circuit breaker) through the communication interface, triggering the circuit breaker to trip, thereby achieving rapid isolation of the faulty line and preventing the fault from escalating; at the same time, after receiving the fault information, the monitoring center notifies the operation and maintenance personnel to repair the fault, thereby improving the power supply reliability of the power system.

[0078] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A fault detection and classification method for overhead transmission lines based on discrete wavelet transform and temporal convolutional networks, characterized in that, Includes the following steps: S1. Fault signal acquisition and preprocessing: The three-phase current signals Ia, Ib, Ic and grounding current signal Ig of the overhead transmission line are acquired by sensors. The acquired raw signals are normalized and filtered by moving average to eliminate noise interference. S2. Feature extraction based on discrete wavelet transform (DWT): The Daubechies wavelet basis function Db4 is used as the mother wavelet to perform multi-scale decomposition on the preprocessed fault current signal to obtain high-frequency detail coefficients containing fault information and form time series data. S3. Time series input construction: The high-frequency detail coefficients extracted in step 2 are used to form a multi-channel time series input vector with an input dimension of 1×T×4, where T is the time step and 4 corresponds to the four channels of three-phase current and grounding current. S4. Fault classification based on temporal convolutional neural network (TCN): Construct a TCN model containing multiple residual blocks, wherein the residual blocks include causal convolutional layers, dilated convolutional layers, ReLU activation functions, weight normalization layers and Dropout layers. Input the time series input vector constructed in step 3 into the TCN model, and output the fault classification result through model training and inference. S5. Result Output and Protection Action Trigger: Based on the fault classification results of the TCN model, trigger the corresponding protection device action and record fault-related information.

2. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 1, the sensor uses high-precision current transformers (CTs) and voltage transformers (VTs), the signal sampling frequency is set to 20kHz, each signal group has a length of 100,000 sampling points, corresponding to a time window of 0.1 seconds.

3. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 1, the normalization formula is: ,in I represents the normalized current signal, and I represents the original current signal. This represents the minimum value of the original current signal. This represents the maximum value of the original current signal.

4. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 2, the discrete wavelet transform is defined as follows: Where x(k) represents the input signal and ψ represents the mother wavelet function, Represents the scale factor. The translation parameters are represented by m, n, and k, which represent integer indices.

5. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 4, the output sequence y of the causal convolution satisfies: , where x is the input sequence, f is the filter, K is the filter size, and t represents the time step index.

6. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 4, the output sequence y of the dilated convolution satisfies: , where d is the expansion factor, K is the filter size, f is the filter, and t represents the time step index.

7. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 4, the receptive field calculation of the TCN model satisfies the following: the receptive field of a single-layer TCN When stacking L-layer TCN residual blocks, the total receptive field , where K is the kernel size and d is the expansion factor of the current layer, and the expansion factor increases exponentially.

8. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 4, the TCN model is trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, 30 training epochs, and the cross-entropy loss function. ,in It's a real label. It represents the predicted probability, and N is the total number of categories.

9. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 4, the residual block output of the TCN model satisfies: ,in This represents the transformation function after operations such as convolution and activation. is the input to the residual block, and ReLU is the activation function.

10. The method for fault detection and classification of overhead transmission lines based on discrete wavelet transform and temporal convolutional networks according to claim 1, characterized in that, In step 5, the fault-related information includes fault type, fault location, and fault timestamp. The fault types include single-phase grounding fault, phase-to-phase short-circuit fault, three-phase short-circuit fault, two-phase grounding fault, and no-fault state, totaling 12 categories.

Images