A grounding wire location method and system based on deep residual networks

Abstract

This invention relates to a grounding wire localization method and system based on a deep residual network. The method first acquires and preprocesses grounding wire signals to generate a dataset. Then, it uses an improved Teager-Kaiser energy operator to extract multi-scale features and outputs a time-frequency energy spectrum. Next, it constructs a localization model containing a deep residual network. Based on the time-frequency energy spectrum, the model is trained using a multi-task learning framework and a loss function. The trained model is then lightweighted and deployed to an edge terminal. Grounding wire signals are acquired in real time and input into the lightweight model on the edge terminal, thereby outputting the real-time spatial coordinates and position status of the grounding wire. Compared with existing technologies, this invention has advantages such as high positioning accuracy, strong anti-interference capability, and good real-time performance.

Description

Technical Field

[0001] This invention relates to the field of power transmission technology, and in particular to a grounding wire positioning method and system based on a deep residual network. Background Art

[0002] In power systems, transmission line grounding wires are critical protective equipment for ensuring the safety of maintenance personnel, and their correct installation and condition monitoring directly affect operational safety. Traditional grounding wire location mainly relies on manual inspection or simple sensor monitoring, which suffers from low efficiency, poor real-time performance, and high false alarm rates. Especially in high-voltage / ultra-high-voltage transmission scenarios, where lines are widely distributed and terrain is complex, manual inspection is costly and difficult to cover the entire area. In recent years, signal processing-based location methods (such as wavelet transform and Hilbert-Huang transform) have been gradually applied, but due to strong electromagnetic interference and multi-source noise coupling, feature extraction accuracy is insufficient, leading to a high false alarm rate. At the same time, deep learning technology has shown great potential in power equipment condition monitoring, but existing models (such as CNN and LSTM) have limited ability to learn features of weak fault signals and do not fully incorporate the physical characteristics of transmission line grounding wires.

[0003] Existing transmission line grounding wire location technologies have the following drawbacks: First, traditional manual inspections rely on subjective judgment, making it difficult to achieve all-weather, full-coverage monitoring. Second, signal processing methods based on wavelet transform or Hilbert-Huang transform are susceptible to noise pollution in environments with strong electromagnetic interference, causing weak features to be submerged. Third, existing deep learning models lack the ability to extract time-frequency domain coupled features and have not been optimized for the correlation between grounding wire vibration and current signals, resulting in low positioning accuracy and a high false alarm rate, making it difficult to meet the real-time monitoring needs of complex power grid scenarios.

[0004] In summary, current transmission line grounding wire positioning technology has low accuracy, poor anti-interference ability, and cannot achieve real-time monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a grounding wire positioning method and system based on deep residual networks, so as to achieve accurate positioning and real-time monitoring of the grounding wire status and improve the level of intelligent operation and maintenance of power grid.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to one aspect of the present invention, a grounding wire location method based on a deep residual network is provided, the method comprising the following steps:

[0008] S1. Acquire the grounding wire signal and preprocess it to generate a signal dataset;

[0009] S2. Use the improved Teager-Kaiser energy operator to extract features from the signal dataset and output the time-frequency energy spectrum.

[0010] S3. Construct a grounding wire location model that includes a deep residual network;

[0011] S4. Based on the time-frequency energy spectrum, the grounding wire location model is trained using a multi-task learning framework and loss function, and the trained grounding wire location model is output.

[0012] S5. Perform lightweight processing on the trained grounding wire positioning model and deploy the lightweight model to the edge terminal;

[0013] S6. Real-time acquisition of grounding wire signals and input of them into the lightweight model of the edge terminal, thereby outputting the real-time spatial coordinates and position status of the grounding wire.

[0014] As a preferred technical solution, the grounding wire signal in S1 includes a grounding current signal and a mechanical vibration signal. The preprocessing includes: aligning the time of each signal and recording the time stamp information of each signal; and then filtering, denoising and standardizing each signal.

[0015] As a preferred technical solution, the specific process of S2 includes:

[0016] S21. The improved Teager-Kaiser energy operator is used to capture the weak features in the grounding wire signal to obtain the processed signal;

[0017] S22. Perform time-frequency analysis on the processed signal to generate an initial time-frequency energy spectrum.

[0018] S23. Use dimensionality reduction techniques to process the initial time-frequency energy spectrum, reduce the number of features, and retain the preset important positioning information to obtain the time-frequency energy spectrum.

[0019] As a preferred technical solution, the improved Teager-Kaiser energy operator in S2 introduces an adaptive threshold filtering strategy and a multi-scale energy fusion strategy.

[0020] As a preferred technical solution, the specific formula for the adaptive threshold filtering strategy is as follows:

[0021] ;

[0022] in, For the improved Teager-Kaiser energy operator; The input signal is sampled at discrete time point n; This is the result after median filtering of the input signal; lThis is a noise suppression coefficient dynamically adjusted using a particle swarm optimization algorithm. This coefficient is automatically adapted based on the local signal-to-noise ratio, increasing in preset high-noise sections. l Enhanced filtering reduces noise in a preset low-noise range. l To preserve signal details.

[0023] As a preferred technical solution, the multi-scale energy fusion strategy includes a sliding window mechanism and a multi-band energy fusion mechanism. The sliding window length covers the typical time-frequency characteristic range of a ground fault. By calculating the instantaneous energy envelope within the window, a time-varying energy trajectory is generated. Furthermore, the energy trajectory is decomposed into a low-frequency band of 0-5kHz and a high-frequency band of 5-10kHz, and weighted and fused respectively to enhance the transient impact characteristics of weak grounding events.

[0024] As a preferred technical solution, the deep residual network in S3 uses an improved bottleneck structure as its core module, employs 1×1 convolution to achieve channel number compression and expansion, replaces traditional convolution with 3×3 depthwise separable convolution, and embeds a lightweight SE attention mechanism. The deep residual network as a whole adopts an 18-layer depth design, and alleviates the gradient vanishing problem through residual connections.

[0025] As a preferred technical solution, the multi-task learning framework in S4 has a shared feature extraction module at the bottom layer and a dual-branch structure at the top layer. The dual-branch structure includes a location regression branch and a state classification branch. The location regression branch outputs spatial coordinates through a fully connected layer, while the state classification branch predicts two states, normal or abnormal, through a Softmax layer.

[0026] As a preferred technical solution, when training the grounding wire location model in S4, the Adam algorithm is used, combined with a cosine annealing learning rate scheduling strategy, with the initial learning rate set to 1×10. -4 The input signal is decayed to 0.9 times its original value every 10 epochs, and MixUp is used for data augmentation during training to mix the input signal in the time and frequency domains, generating virtual samples to expand the boundary scene data. The loss function is a joint loss function, which is a weighted sum of the Smooth L1 loss function for the location regression task and the Focal Loss loss function for the state classification task. The specific formula is as follows:

[0027] ;

[0028] in, The joint loss function value is used to measure the overall difference between the model's predicted results and the true labels, and to guide the updating of model parameters; These are the loss weighting coefficients for the location regression task; The loss function is Smooth L1. For the location regression task, the true label; This represents the model's predicted value for the location regression task; The loss weight coefficients for the state classification task; Focal Loss function; True labels for state classification tasks; This represents the model's predicted value for the state classification task.

[0029] According to another aspect of the present invention, a grounding wire positioning system based on a deep residual network is provided. The system includes a data processing module, a feature extraction module, a grounding wire positioning model, a model training module, and a model application module.

[0030] The data processing module collects and preprocesses data to generate a dataset;

[0031] The feature extraction module uses the improved Teager-Kaiser energy operator to extract features from the dataset and outputs a time-frequency energy spectrum.

[0032] The grounding wire location model includes a deep residual network;

[0033] The model training module is based on the time-frequency energy spectrum and uses a multi-task learning framework and loss function to train the grounding wire location model, and outputs the trained grounding wire location model.

[0034] The model application module performs lightweight processing on the trained grounding wire positioning model and deploys the lightweight model to the edge terminal. It collects grounding wire signals in real time and inputs them into the lightweight model of the edge terminal, thereby outputting the real-time spatial coordinates and position status of the grounding wire to achieve real-time positioning of the grounding wire.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. This invention encompasses the entire process of signal acquisition, feature extraction, model construction and training, lightweight deployment, and real-time application, forming a complete technical closed loop. By enhancing the time-frequency domain feature extraction of weak grounding signals through an improved Teager-Kaiser energy operator, and by constructing a deep residual network architecture and training it using a multi-task learning framework, combined with edge terminal deployment, the real-time spatial coordinates and position status output of the grounding wire are achieved. This method offers high positioning accuracy and strong anti-interference capabilities, meeting the rapid response requirements of transmission line operation and maintenance, promoting the development of transmission line operation and maintenance towards intelligence and reduced manpower, and ensuring the safe and stable operation of the power grid.

[0037] 2. This invention improves upon existing operators to capture weak features, uses time-frequency analysis to generate spectra, and employs dimensionality reduction techniques to simplify features. Layered processing preserves the core features of the grounding signal (such as transient impulses and frequency abrupt changes) while eliminating redundant information. Compared to directly using the original signal, the time-frequency energy spectrum more intuitively reflects the signal's time-frequency characteristics, and dimensionality reduction further improves the model's computational efficiency.

[0038] 3. This invention introduces a Teager-Kaiser energy operator with an adaptive threshold filtering and multi-scale energy fusion strategy. This enhances the feature extraction capability of weak grounding signals, suppresses strong electromagnetic noise interference, and can accurately capture weak signal features in environments with strong interference. The adaptive filtering dynamically balances noise reduction and detail preservation, while multi-scale fusion covers features across different frequency bands, solving the problem of low sensitivity of traditional operators to weak signals and providing richer feature inputs for the localization model. The adaptive filtering strategy dynamically adjusts the parameters using a particle swarm optimization algorithm. l The coefficients are used to adapt the filtering intensity to the local signal-to-noise ratio of the signal. Strong noise segments are filtered more intensely to suppress interference, while weak noise segments are filtered less intensely to preserve details. This avoids feature loss or noise residue caused by traditional fixed-threshold filtering, significantly improving the signal-to-noise ratio and feature integrity of the preprocessed signal. In the multi-scale energy fusion strategy, a sliding window covers the typical time-frequency range of the fault, and combined with weighted fusion of high and low frequency bands, it highlights the transient impact characteristics of weak grounding events. Compared to single-scale analysis, multi-scale fusion can more comprehensively capture the time-frequency distribution pattern of the fault signal and improve the correlation between features and the location of the grounding wire.

[0039] 4. The deep residual network used in this invention reduces computational cost through 1×1 convolutions and 3×3 depthwise separable convolutions, constructing a lightweight residual network. An attention mechanism is embedded to strengthen key feature channels, and the SE attention mechanism enhances the weights of key features. Eighteen layers of depth and residual connections ensure feature representation capabilities. This architecture maintains high-precision positioning despite its lightweight design, solving the problems of computational complexity and deployment difficulties inherent in traditional deep networks.

[0040] 5. In the multi-task learning framework of this invention, spatial coordinates and states are output simultaneously through a shared feature extraction module combined with a dual-branch approach of location regression and state classification. Compared to a single-task model, the dual-branch design can acquire multi-dimensional information in a single inference, reducing the number of model calls and meeting the integrated needs of multiple tasks in operation and maintenance. It simultaneously realizes grounding wire location regression and state classification. The model integrates energy features and spatial features, and processes multi-source signals through a dual-branch network structure, significantly improving positioning accuracy and anti-interference capability, meeting the real-time monitoring needs of complex power grid scenarios, and improving monitoring robustness in complex environments.

[0041] 6. In this invention, the Adam algorithm combined with cosine annealing learning rate scheduling is used to achieve rapid convergence of model parameters; and the MixUp data augmentation is used to expand the boundary scene samples to alleviate the sample imbalance problem; the joint loss function balances the weights of regression and classification tasks. The trained model has stronger generalization ability and can maintain stable positioning accuracy and state judgment accuracy in complex transmission line environments. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the steps of the grounding wire positioning method based on deep residual network in this invention;

[0043] Figure 2 This is a schematic diagram illustrating the steps of feature extraction and output of time-frequency energy spectrum in this invention. Detailed Implementation

[0044] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0045] This solution aims to address the low accuracy and poor anti-interference capabilities of existing grounding wire location technologies by integrating and improving the Teager-Kaiser energy operator with a deep residual network. It enhances the feature extraction capability of weak grounding signals and suppresses strong electromagnetic noise interference; constructs a lightweight deep learning model to achieve real-time and accurate location of grounding wire status; reduces false alarm rates and improves monitoring robustness in complex environments; and promotes the intelligent and less-manned development of transmission line operation and maintenance, ensuring the safe and stable operation of the power grid.

[0046] Example 1

[0047] In this embodiment, a grounding wire location method based on a deep residual network is adopted. An adaptive threshold filtering ITKEO operator is designed to enhance the extraction of time-frequency domain features of weak grounding signals and suppress strong electromagnetic noise. A lightweight residual network is constructed, embedding an attention mechanism to strengthen key feature channels, simultaneously achieving grounding wire position regression and state classification. A space-time joint location model is proposed, fusing ITKEO energy features and DRN spatial features. A dual-stream network structure is used to process multi-source signals (current harmonics and vibration), significantly improving location accuracy and anti-interference capability, meeting the real-time monitoring needs of complex power grid scenarios. The method steps are as follows: Figure 1 As shown, it specifically includes:

[0048] S1. Acquire the grounding wire signal and preprocess it to generate a signal dataset;

[0049] S2. Use the improved Teager-Kaiser energy operator to extract features from the signal dataset and output the time-frequency energy spectrum.

[0050] S3. Construct a grounding wire location model that includes a deep residual network;

[0051] S4. Based on the time-frequency energy spectrum, the grounding wire location model is trained using a multi-task learning framework and loss function, and the trained grounding wire location model is output.

[0052] S5. Perform lightweight processing on the trained grounding wire positioning model and deploy the lightweight model to the edge terminal;

[0053] S6. Real-time acquisition of grounding wire signals and input of them into the lightweight model of the edge terminal, thereby outputting the real-time spatial coordinates and position status of the grounding wire.

[0054] The grounding wire signals in S1 include grounding current signals and mechanical vibration signals. Preprocessing includes: aligning the time of each signal and recording its time stamp information; then filtering, denoising, and standardizing each signal. Targeted acquisition of grounding current and mechanical vibration signals, through time alignment, filtering, denoising, and standardization, eliminates signal time stamp deviations and noise interference, unifying the data scale. The preprocessed data better matches the model input requirements, providing a high-quality foundation for subsequent feature extraction and model training, and reducing the impact of invalid data on the positioning results.

[0055] The specific process of S2 is as follows: Figure 2 As shown, it includes:

[0056] S21. The improved Teager-Kaiser energy operator is used to capture the weak features in the grounding wire signal to obtain the processed signal;

[0057] S22. Perform time-frequency analysis on the processed signal to generate an initial time-frequency energy spectrum.

[0058] S23. Use dimensionality reduction techniques to process the initial time-frequency energy spectrum, reduce the number of features, and retain the preset important positioning information to obtain the time-frequency energy spectrum.

[0059] The improved Teager-Kaiser energy operator in S2 introduces an adaptive threshold filtering strategy and a multi-scale energy fusion strategy.

[0060] The specific formula for the adaptive threshold filtering strategy is as follows:

[0061] ;

[0062] in, For the improved Teager-Kaiser energy operator; The input signal is sampled at discrete time point n; This is the result after median filtering of the input signal; l This is a noise suppression coefficient dynamically adjusted using a particle swarm optimization algorithm. This coefficient is automatically adapted based on the local signal-to-noise ratio, increasing in preset high-noise sections. l Enhanced filtering reduces noise in a preset low-noise range. l To preserve signal details.

[0063] The multi-scale energy fusion strategy includes a sliding window mechanism and a multi-band energy fusion mechanism. The sliding window length covers the typical time-frequency characteristic range of a ground fault. By calculating the instantaneous energy envelope within the window, a time-varying energy trajectory is generated. The energy trajectory is further decomposed into a low-frequency band of 0-5kHz and a high-frequency band of 5-10kHz, and weighted and fused respectively to enhance the transient impact characteristics of weak grounding events.

[0064] The deep residual network in S3 uses an improved bottleneck structure as its core module. It employs 1×1 convolutions to compress and expand the number of channels, replaces traditional convolutions with 3×3 depthwise separable convolutions, and embeds a lightweight SE attention mechanism. The deep residual network adopts an 18-layer depth design and alleviates the gradient vanishing problem through residual connections.

[0065] The multi-task learning framework in S4 has a shared feature extraction module at the bottom layer and a dual-branch structure at the top layer. The dual-branch structure includes a location regression branch and a state classification branch. The location regression branch outputs spatial coordinates through a fully connected layer, while the state classification branch predicts two states, normal or abnormal, through a Softmax layer.

[0066] When training the grounding wire location model in S4, the Adam algorithm is used, combined with a cosine annealing learning rate scheduling strategy, with the initial learning rate set to 1×10. -4 The loss function decays to 0.9 times its original value every 10 epochs, and MixUp is used for data augmentation during training to mix the input signal in the time and frequency domains, generating virtual samples to expand the boundary scene data. The joint loss function is a weighted sum of the Smooth L1 loss function for location regression and the Focal Loss loss function for state classification, and its specific formula is as follows:

[0067] ;

[0068] in, The joint loss function value is used to measure the overall difference between the model's predicted results and the true labels, and to guide the updating of model parameters; These are the loss weighting coefficients for the location regression task; The loss function is Smooth L1. For the location regression task, the true label; This represents the model's predicted value for the location regression task; The loss weight coefficients for the state classification task; Focal Loss function; True labels for state classification tasks; This represents the model's predicted value for the state classification task.

[0069] In summary, this method covers the entire process from signal acquisition, feature extraction, model building and training, lightweight deployment, and real-time application, forming a complete technical closed loop. By enhancing the time-frequency domain feature extraction of weak grounding signals through an improved Teager-Kaiser energy operator, and by constructing a deep residual network architecture and training it using a multi-task learning framework, combined with edge terminal deployment, it achieves real-time spatial coordinates and position status output of the grounding wire. This solves the problems of low accuracy and poor anti-interference in existing grounding wire positioning technologies, meets the rapid response requirements of transmission line operation and maintenance, promotes the development of transmission line operation and maintenance towards intelligence and reduced manpower, and ensures the safe and stable operation of the power grid.

[0070] Example 2

[0071] In this embodiment, a grounding wire positioning system based on a deep residual network is adopted. The system includes a data processing module, a feature extraction module, a grounding wire positioning model, a model training module, and a model application module.

[0072] The data processing module collects and preprocesses data to generate a dataset;

[0073] The feature extraction module uses the improved Teager-Kaiser energy operator to extract features from the dataset and outputs a time-frequency energy spectrum.

[0074] The grounding wire location model includes a deep residual network;

[0075] The model training module is based on the time-frequency energy spectrum and uses a multi-task learning framework and loss function to train the grounding wire location model, and outputs the trained grounding wire location model.

[0076] The model application module performs lightweight processing on the trained grounding wire positioning model and deploys the lightweight model to the edge terminal. It collects grounding wire signals in real time and inputs them into the lightweight model of the edge terminal, thereby outputting the real-time spatial coordinates and position status of the grounding wire to achieve real-time positioning of the grounding wire.

[0077] The specific implementation steps of this system are as follows:

[0078] Step S1: Data Acquisition and Preprocessing

[0079] Step S2: Improved Teager-Kaiser energy operator feature extraction

[0080] Step S3: Deep Residual Network Modeling and Training

[0081] Step S4: Spatiotemporal Feature Fusion and Localization Output

[0082] Step S1 includes the following steps:

[0083] Step S11: Install high-precision current sensors and vibration sensors at key locations on the grounding wire of the transmission line to collect grounding current and mechanical vibration data in real time;

[0084] Step S12: Ensure time alignment of multi-sensor data through the time synchronization module and record the time stamp information of the signal;

[0085] Step S13: Filter, denoise, and standardize the acquired signals to generate a high-quality dataset for subsequent analysis.

[0086] Step S2 includes the following steps:

[0087] Step S21: An improved Teager-Kaiser energy operator is proposed, which captures weak features in the grounding wire signal more accurately through adaptive filtering and multi-scale energy fusion;

[0088] Step S22: Perform time-frequency analysis on the processed signal to generate a time-frequency energy spectrum, so as to intuitively show the changes of signal energy with time and frequency;

[0089] Step S23: Use dimensionality reduction techniques to reduce the number of features while retaining the information most important to the localization task, thereby improving computational efficiency.

[0090] Step S21 includes the following:

[0091] To address the issues of traditional Teager-Kaiser Energy Operator (TKEO) being susceptible to noise contamination and lacking sufficient capability for extracting weak grounding signal features under strong electromagnetic interference environments, this invention proposes an improved Teager-Kaiser Energy Operator (ITKEO). This operator significantly enhances the accuracy and robustness of signal feature extraction by introducing an adaptive threshold filtering mechanism and a multi-scale energy fusion strategy.

[0092] First, based on the classic TKEO formula, this invention designs a dynamic adjustment term to suppress noise interference. The traditional TKEO operator formula is:

[0093]

[0094] Furthermore, the ITKEO operator incorporates an adaptive threshold filtering module, resulting in an optimized formula:

[0095]

[0096] in, This is a noise suppression coefficient dynamically adjusted using the Particle Swarm Optimization (PSO) algorithm. This coefficient is automatically adapted based on the signal's local signal-to-noise ratio (SNR), increasing in areas of high noise. Enhanced filtering reduces noise in the weak noise range. To preserve signal details.

[0097] Secondly, to capture the multi-scale characteristics of grounding wire vibration and current signals, the ITKEO operator employs a sliding window mechanism and a multi-band energy fusion strategy. The sliding window length is set to 512 sampling points, covering the typical time-frequency characteristics of grounding faults. By calculating the instantaneous energy envelope within the window, a time-varying energy trajectory is generated. The energy trajectory is further decomposed into a low-frequency band of 0-5kHz and a high-frequency band of 5-10kHz, and weighted with 0.6 and 0.4 respectively before weighted fusion, enhancing the transient impact characteristics of weak grounding events (such as loose connections).

[0098] Step S3 includes the following steps:

[0099] Step S31: A deep residual network was designed to reduce the number of parameters and improve inference speed while maintaining localization accuracy through bottleneck structure and attention mechanism;

[0100] Step S32: Employ a multi-task learning framework that combines location regression and state classification tasks, and achieve efficient model training through a joint loss function and optimization algorithm.

[0101] Step S31 includes the following:

[0102] To address the dual requirements of lightweight models and strong feature representation capabilities in power transmission line grounding wire location tasks, this invention designs a deep separable residual network architecture. The network uses an improved bottleneck structure as its core module, employing 1×1 convolutions to compress and expand the number of channels, and replacing traditional convolutions with 3×3 deep separable convolutions to reduce computational cost. To enhance the network's sensitivity to key features, a lightweight Squeeze-and-Excitation (SE) attention mechanism is embedded in the residual module. This mechanism generates channel-level statistics through global average pooling, learns nonlinear mapping relationships through two fully connected layers, and finally generates channel weight coefficients in the range of 0-1, dynamically weighting the feature map. The network as a whole adopts an 18-layer deep design, mitigating the gradient vanishing problem through residual connections.

[0103] Step S32 includes the following:

[0104] To achieve coordinated optimization of accurate grounding wire location regression and intelligent state classification, this invention constructs a multi-task learning framework and designs a joint loss function with dynamically adjusted weights. Addressing the coupling characteristics of location regression and state classification in the grounding wire localization task, the network's lower layers share a feature extraction module, while the higher layers employ a dual-branch structure to handle the two tasks separately. The location regression branch outputs spatial coordinates through a fully connected layer, while the state classification branch outputs spatial coordinates through a fully connected layer. The layer predicts two states: normal and abnormal.

[0105] The joint loss function consists of a weighted average of the Smooth L1 loss for location regression and the Focal Loss for classification tasks.

[0106]

[0107] in, This is the dynamic balance coefficient. The design addresses the sparse sample size of abnormal grounding wire states through Focal Loss, and its modulation factor... It effectively suppresses the weight of easily classified samples, allowing the model to focus on difficult cases such as virtual connections.

[0108] Network optimization employs the Adam algorithm, combined with a cosine annealing learning rate scheduling strategy, with the initial learning rate set to... The value decays to 0.9 times its original value every 10 epochs. To enhance the model's generalization ability, MixUp data augmentation is introduced during training to mix the input signal in the time and frequency domains, generating virtual samples to expand the boundary scene data, which is significantly better than the single-task learning baseline model.

[0109] Step S4 also includes the following steps:

[0110] Step S41: Deploy the lightweight model to the edge terminal, adapt the hardware and design electromagnetic shielding to ensure detection accuracy under strong electric fields;

[0111] Step S42: Verify the stability and accuracy of the system under extreme conditions through laboratory and field multi-scenario testing.

[0112] Step S41 also includes the following:

[0113] To meet the real-time requirements of power transmission line grounding wire status monitoring, an edge computing terminal deployment architecture based on the NVIDIA Jetson TX2 development board was designed. This platform features a 256-core Pascal architecture GPU and a dual-core Denver 2 64-bit CPU, equipped with 8GB LPDDR4 memory and 32GB eMMC storage, providing 15W low-power computing power. The lightweight residual network model is optimized using the TensorRT engine with FP16 / INT8 hybrid quantization, compressing the model size to 12.7MB, increasing inference speed by 2.3 times, and achieving a single-frame response time of 12ms. For complex electromagnetic environments, a dual-layer electromagnetic protection design is adopted: the outer layer uses an aluminum alloy shielding shell, achieving -80dB@1GHz electromagnetic shielding effectiveness based on the Faraday cage principle; the inner layer integrates a three-stage filtering circuit, including a common-mode inductor, X / Y capacitors, and a feedthrough filter, implementing differential-mode / common-mode hybrid filtering for power and signal lines, ensuring a 97.6% detection accuracy under strong electric field interference of 10kV / m. The hardware interface uses military-grade aviation connectors, with an operating temperature range of -40℃ to 85℃, and is IP67 certified. System integration utilizes Docker containerization to decouple model inference, data preprocessing, and communication protocols, supporting OTA remote updates.

[0114] Step S42 also includes the following:

[0115] To comprehensively verify system performance, multi-dimensional testing and verification were conducted. In the laboratory phase, a test set of typical grounding faults such as loose connections and open circuits was constructed to simulate different fault characteristics, focusing on verifying the model's ability to identify weak signals and its response speed. Field testing selected UHVDC and AC lines, with equipment operating under extreme temperature and humidity conditions for extended periods to verify hardware weather resistance and software stability. For mountainous microclimate zones, simulations of freezing rain and strong winds were conducted to verify the equipment's reliability under combined stresses such as mechanical vibration and low-temperature condensation. Simultaneously, electromagnetic compatibility testing ensured stable communication and data acquisition accuracy under strong electromagnetic interference. Comprehensive testing shows that the system meets design goals in key indicators such as positioning accuracy, response speed, and environmental adaptability. In particular, the fault identification rate under complex operating conditions is significantly improved compared to traditional methods, providing reliable technical support for intelligent operation and maintenance of transmission lines.

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A grounding wire location method based on a deep residual network, characterized in that, The method steps include: S1. Acquire and preprocess grounding wire signals to generate a signal dataset; S2. Extract features from the signal dataset using an improved Teager-Kaiser energy operator to output a time-frequency energy spectrum; S3. Construct a grounding wire location model containing a deep residual network; S4. Train the grounding wire location model based on the time-frequency energy spectrum using a multi-task learning framework and loss function, and output the trained grounding wire location model; S5. Lighten up the trained grounding wire location model and deploy it to the edge terminal; S6. Acquire grounding wire signals in real time and input them into the lightweight model on the edge terminal to output the real-time spatial coordinates and position status of the grounding wire. The improved Teager-Kaiser energy operator in S2 introduces an adaptive threshold filtering strategy and a multi-scale energy fusion strategy; The specific formula for the adaptive threshold filtering strategy is as follows: ； in, For the improved Teager-Kaiser energy operator; The input signal is sampled at discrete time point n; This is the result after median filtering of the input signal; λ This is a noise suppression coefficient dynamically adjusted using a particle swarm optimization algorithm. This coefficient is automatically adapted based on the local signal-to-noise ratio, increasing in preset high-noise sections. λ Enhanced filtering reduces noise in a preset low-noise range. λ To preserve signal details; The multi-scale energy fusion strategy includes a sliding window mechanism and a multi-band energy fusion mechanism. The sliding window length covers the typical time-frequency characteristic range of a ground fault. By calculating the instantaneous energy envelope within the window, a time-varying energy trajectory is generated. The energy trajectory is further decomposed into a low-frequency band of 0-5kHz and a high-frequency band of 5-10kHz, and weighted and fused respectively to enhance the transient impact characteristics of weak grounding events.

2. The grounding wire positioning method based on a deep residual network according to claim 1, characterized in that, The grounding wire signal in S1 includes a grounding current signal and a mechanical vibration signal. The preprocessing includes: aligning the time of each signal and recording the time stamp information of each signal; and then filtering, denoising and standardizing each signal.

3. The grounding wire positioning method based on a deep residual network according to claim 1, characterized in that, The specific process of S2 includes: S21. The improved Teager-Kaiser energy operator is used to capture the weak features in the grounding wire signal to obtain the processed signal; S22. Perform time-frequency analysis on the processed signal to generate an initial time-frequency energy spectrum. S23. Use dimensionality reduction techniques to process the initial time-frequency energy spectrum, reduce the number of features, and retain the preset important positioning information to obtain the time-frequency energy spectrum.

4. The grounding wire positioning method based on a deep residual network according to claim 1, characterized in that, The deep residual network in S3 uses an improved bottleneck structure as its core module, employs 1×1 convolutions to compress and expand the number of channels, replaces traditional convolutions with 3×3 depthwise separable convolutions, and embeds a lightweight SE attention mechanism. The deep residual network as a whole adopts an 18-layer depth design and alleviates the gradient vanishing problem through residual connections.

5. The grounding wire positioning method based on a deep residual network according to claim 1, characterized in that, The multi-task learning framework in S4 has a shared feature extraction module at the bottom layer and a dual-branch structure at the top layer. The dual-branch structure includes a position regression branch and a state classification branch. The position regression branch outputs spatial coordinates through a fully connected layer, while the state classification branch predicts two states, normal or abnormal, through a Softmax layer.

6. The grounding wire positioning method based on a deep residual network according to claim 5, characterized in that, In step S4, when training the grounding wire location model, the Adam algorithm is used, combined with a cosine annealing learning rate scheduling strategy, and the initial learning rate is set to 1×10. -4 The loss function decays to 0.9 times its original value every 10 epochs, and MixUp is used for data augmentation during training to mix the input signal in the time and frequency domains, generating virtual samples to expand the boundary scene data. The loss function is a joint loss function, which is a weighted sum of the Smooth L1 loss function for the location regression task and the Focal Loss loss function for the state classification task. Its specific formula is as follows: ； in, The joint loss function value is used to measure the overall difference between the model's predicted results and the true labels, and to guide the updating of model parameters; These are the loss weighting coefficients for the location regression task; The loss function is Smooth L1. For the location regression task, the true label; This represents the model's predicted value for the location regression task; The loss weight coefficients for the state classification task; Focal Loss function; True labels for state classification tasks; This represents the model's predicted value for the state classification task.

7. A grounding wire positioning system based on a deep residual network, characterized in that, The system operates using a grounding wire location method based on a deep residual network as described in any one of claims 1-6. The system includes a data processing module, a feature extraction module, a grounding wire location model, a model training module, and a model application module. The data processing module collects data and preprocesses the data to generate a dataset; The feature extraction module uses the improved Teager-Kaiser energy operator to extract features from the dataset and outputs a time-frequency energy spectrum. The grounding wire positioning model includes a deep residual network; The model training module is based on the time-frequency energy spectrum and uses a multi-task learning framework and loss function to train the grounding wire location model, and outputs the trained grounding wire location model. The model application module performs lightweight processing on the trained grounding wire positioning model and deploys the lightweight model to the edge terminal. It collects grounding wire signals in real time and inputs them into the lightweight model of the edge terminal, thereby outputting the real-time spatial coordinates and position status of the grounding wire to achieve real-time positioning of the grounding wire.

Images