Fault detection methods, devices, computer-readable storage media, and processors for overhead ground wires

Abstract

This invention discloses a method, apparatus, computer-readable storage medium, and processor for fault detection of overhead ground wires. The method includes: acquiring vibration signals from the overhead ground wire; preprocessing the vibration signals based on a target empirical wavelet transform algorithm to obtain multiple modal components, wherein different modal components represent different frequency components in the vibration signals; extracting features from the multiple modal components using a feature extraction model to obtain feature vectors, wherein the feature vectors characterize the deep-level features related to faults in the overhead ground wires within the multiple modal components; and detecting the feature vectors using a fault detection model to obtain fault detection results, wherein the fault detection results include at least the fault type and the fault location on the overhead ground wire. This invention solves the technical problem of low accuracy in fault detection of overhead ground wires in related technologies.

Description

Technical Field

[0001] This invention relates to the field of power engineering technology, and more specifically, to a method, apparatus, computer-readable storage medium, and processor for detecting faults in overhead ground wires. Background Technology

[0002] With the continuous development of power systems, overhead ground wires play a crucial role in ensuring the stability and safety of power transmission. However, because overhead ground wires are exposed to the natural environment for extended periods, they are susceptible to various factors that can lead to failure.

[0003] In related technologies, commonly used intelligent algorithms for fault diagnosis of overhead ground wires include decision trees, backpropagation neural networks, support vector machines, K-nearest neighbors classifiers, and deep learning. However, most of these algorithms suffer from low accuracy, poor generalization ability, or are prone to overfitting.

[0004] There is currently no effective solution to the aforementioned technical problem of low accuracy in detecting faults in overhead ground wires. Summary of the Invention

[0005] This invention provides a method, apparatus, computer-readable storage medium, and processor for detecting faults in overhead ground wires, thereby at least addressing the technical problem of low accuracy in detecting faults in overhead ground wires in related technologies.

[0006] According to one aspect of the present invention, a fault detection method for an overhead ground wire is provided. The method includes: acquiring a vibration signal of the overhead ground wire, wherein the vibration signal is used to at least characterize the operating state of the overhead ground wire; preprocessing the vibration signal based on a target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal, wherein the target empirical wavelet transform algorithm is an improved version of the empirical wavelet transform algorithm, and different modal components are used to represent different frequency components in the vibration signal; extracting features from the multiple modal components using a feature extraction model to obtain feature vectors, wherein the feature extraction model is obtained by training a convolutional autoencoder model, and the feature vectors are used to characterize the deep-level features related to the fault of the overhead ground wire in the multiple modal components; and detecting the feature vectors using a fault detection model to obtain a fault detection result, wherein the fault detection model is obtained by pre-training an ensemble learning algorithm model using a particle swarm optimization algorithm, and the fault detection result includes at least the fault type of the overhead ground wire and the fault location on the overhead ground wire.

[0007] Optionally, the vibration signal is preprocessed based on the target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal, including: transforming the vibration signal from the time domain to the frequency domain based on the fast Fourier transform algorithm to obtain the signal spectrum of the vibration signal; determining the spectral segmentation points of the signal spectrum based on the scale space method; dividing the signal spectrum into multiple spectral intervals based on the spectral segmentation points, wherein the multiple spectral intervals are used to represent different frequency components in the signal spectrum; fitting the frequency maxima points in each of the multiple spectral intervals based on the cubic Hermitian interpolation algorithm to generate a spectral envelope, wherein the spectral envelope is used to represent the trend of the spectrum; and using the local minima points on the spectral envelope as boundary points to perform a secondary division of the signal spectrum to obtain multiple modal components of the vibration signal.

[0008] Optionally, determining the spectral segmentation points of the signal spectrum based on the scale-space method includes: smoothing the signal spectrum using a Gaussian function to determine the spectral segmentation points.

[0009] Optionally, feature extraction is performed on multiple modal components using a feature extraction model to obtain feature vectors, including: extracting local features of multiple modal components using the convolutional layer of the feature extraction model, wherein the local features are features related to the fault of the overhead ground wire, and the local features are used to characterize the intensity and change trend of multiple modal components at a specified location; and performing dimensionality reduction processing on the local features of multiple modal components using the pooling layer of the feature extraction model to obtain feature vectors.

[0010] Optionally, the fault detection model is used to detect the feature vector to obtain the fault detection result, including: inputting the feature vector into the fault detection model for fault detection; receiving the output result of the fault detection model; and determining the fault detection result based on the output result.

[0011] Optionally, the fault detection method for overhead ground wires further includes: acquiring vibration signal samples of the overhead ground wire; preprocessing the vibration signal samples to obtain multiple modal component samples of the vibration signal samples; dividing the multiple modal component samples into a first training set and a first test set according to a preset ratio; training a convolutional autoencoder feature extraction model using the first training set to obtain a trained convolutional autoencoder feature extraction model; validating the trained convolutional autoencoder feature extraction model using the first test set to obtain a first verification result, wherein the first verification result is used to evaluate the feature extraction capability of the trained convolutional autoencoder feature extraction model; and determining the trained convolutional autoencoder feature extraction model as a feature extraction model in response to the first verification result indicating that the feature reconstruction error of the trained convolutional autoencoder feature extraction model is less than a preset threshold, wherein the feature reconstruction error is used to characterize the degree of difference between the reconstructed signal output by the convolutional autoencoder feature extraction model and the input signal of the convolutional autoencoder feature model in the first test set.

[0012] Optionally, the fault detection method for overhead ground wires further includes: obtaining multiple feature vector samples by performing feature extraction on multiple modal component samples in a first test set using a feature extraction model; dividing the multiple feature vector samples into a second training set and a second test set; inputting the second training set into an ensemble learning algorithm model and using a particle swarm optimization algorithm to obtain a target parameter combination of the ensemble learning algorithm model, wherein the target parameter combination includes at least a learning rate parameter and a regularization parameter; adjusting the parameters in the ensemble learning algorithm model using the parameters in the target parameter combination to obtain an adjusted ensemble learning algorithm model; verifying the adjusted ensemble learning algorithm model for fault identification using the second test set to obtain a second verification result, wherein the second verification result is used to indicate the performance of the adjusted ensemble learning algorithm model on the second test set; and determining the adjusted ensemble learning algorithm model as a fault detection model in response to the second verification result indicating that the accuracy of the adjusted ensemble learning algorithm model in identifying fault types is greater than a second preset threshold.

[0013] According to another aspect of the present invention, a fault detection device for overhead ground wires is also provided. The device includes: an acquisition unit for acquiring vibration signals of the overhead ground wire, wherein the vibration signals are used to at least characterize the operating state of the overhead ground wire; a processing unit for preprocessing the vibration signals based on a target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signals, wherein the target empirical wavelet transform algorithm is an improved version of the empirical wavelet transform algorithm, and different modal components are used to represent different frequency components in the vibration signals; an extraction unit for extracting features from the multiple modal components using a feature extraction model to obtain feature vectors, wherein the feature extraction model is obtained by training a convolutional autoencoder model, and the feature vectors are used to characterize deep-level features related to faults in the multiple modal components; and a detection unit for detecting the feature vectors using a fault detection model to obtain fault detection results, wherein the fault detection model is obtained by pre-training an ensemble learning algorithm model using a particle swarm optimization algorithm, and the fault detection results include at least the fault type of the overhead ground wire and the fault location on the overhead ground wire.

[0014] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is run by a processor, it controls the device where the storage medium is located to execute the overhead ground wire fault detection method of the present invention.

[0015] According to another aspect of the present invention, a processor is also provided. The processor is used to run a program, wherein the program executes the overhead ground wire fault detection method according to the embodiments of the present invention.

[0016] According to another aspect of the present invention, a computer program product is also provided. The program product includes computer instructions that, when executed by a processor, implement the overhead ground wire fault detection method of the present invention.

[0017] In this embodiment of the invention, vibration signals of overhead ground wires are acquired, wherein the vibration signals are used to at least characterize the operating state of the overhead ground wires; the vibration signals are preprocessed based on a target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signals, wherein the target empirical wavelet transform algorithm is an improvement of the empirical wavelet transform algorithm, and different modal components are used to represent different frequency components in the vibration signals; feature extraction models are used to extract features from the multiple modal components to obtain feature vectors, wherein the feature extraction model is obtained by training a convolutional autoencoder model, and the feature vectors are used to characterize the deep-level features related to faults in the overhead ground wires among the multiple modal components; fault detection models are used to detect the feature vectors to obtain fault detection results, wherein the fault detection model is obtained by pre-training an ensemble learning algorithm model using a particle swarm optimization algorithm, and the fault detection results include at least the fault type of the overhead ground wire and the fault location on the overhead ground wire. In other words, in this embodiment of the invention, the vibration signal of the overhead ground wire is acquired, and the vibration signal is preprocessed using an improved empirical wavelet transform algorithm to eliminate redundant information and noise interference in the vibration signal, thereby obtaining multiple modal components in the vibration signal to improve the accuracy of subsequent feature extraction for fault diagnosis. After obtaining multiple modal components, the feature extraction model is used to further extract features from the multiple modal components to obtain feature vectors that characterize the deep-level features related to the fault of the overhead ground wire in the multiple modal components. Then, the fault detection model is used to detect the feature vectors to obtain the fault detection results. Since the fault detection model is pre-trained using the particle swarm optimization algorithm to train the ensemble learning algorithm, the fault detection model can more accurately identify faults in the overhead ground wire, improve the accuracy of the fault detection results, and thus achieve the technical effect of improving the fault detection accuracy of the overhead ground wire. This provides strong technical support for the safe operation and maintenance of the power system, thereby solving the technical problem of low fault detection accuracy of the overhead ground wire in related technologies. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of a fault detection method for an overhead ground wire according to an embodiment of the present invention;

[0020] Figure 2 This is a flowchart of an overhead ground wire fault diagnosis method according to an embodiment of this application;

[0021] Figure 3This is a flowchart of a method for optimizing an XGBoost model using a particle swarm optimization algorithm according to an embodiment of this application;

[0022] Figure 4 This is a schematic diagram of a particle swarm optimization algorithm according to an embodiment of this application;

[0023] Figure 5 This is a schematic diagram of an overhead ground wire fault detection device according to an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, functional component, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, functional components, or devices.

[0026] According to an embodiment of the present invention, an embodiment of a fault detection method for overhead ground wire is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0027] Figure 1 This is a flowchart of a fault detection method for an overhead ground wire according to an embodiment of the present invention, such as... Figure 1 As shown, the method may include the following steps:

[0028] Step S101: Obtain the vibration signal of the overhead ground wire.

[0029] In the technical solution provided in step S101 of the present invention, the overhead ground wire, also known as the overhead lightning protection wire or ground wire, is a component of the transmission lines in the power system. It is mainly used in high-voltage transmission lines, erected in parallel with the overhead conductors (i.e., power transmission lines), to protect the transmission lines from direct lightning strikes and induced lightning damage. Vibration signals are used to at least characterize the operating status of the overhead ground wire.

[0030] In this embodiment, relevant operational data of the overhead ground wire can be collected, and the vibration signal of the overhead ground wire can be extracted from the collected operational data.

[0031] Alternatively, the vibration signal of the overhead ground wire can be collected by a sensor installed on the overhead ground wire, or the vibration signal of the overhead ground wire can be obtained from the monitoring equipment of the power system. The specific method of obtaining the overhead ground wire is not limited here.

[0032] Step S102: The vibration signal is preprocessed based on the target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal.

[0033] In the technical solution provided in step S102 of the present invention, the target empirical wavelet transform algorithm is an improvement on the empirical wavelet transform algorithm. That is, the target empirical wavelet transform algorithm can also be called the improved empirical wavelet transform algorithm (IEWT). Different modal components among multiple modal components are used to characterize different frequency components in the signal.

[0034] In this embodiment, the target empirical wavelet transform algorithm can accurately capture different frequency information in the vibration signal. Through adaptive filter design, the vibration signal is decomposed into multiple modal components, each of which corresponds to a certain frequency range of the vibration signal. Such decomposition helps to extract clear fault features from complex vibration signals, providing a high-quality data foundation for subsequent fault diagnosis.

[0035] For example, when preprocessing vibration signals using the target empirical wavelet transform algorithm, the acquired vibration signals can first be subjected to a Fast Fourier Transform (FFT) to convert them into frequency domain signals for analysis of their frequency components. Then, the scale-space method is used to determine the spectral segmentation boundaries. Scale-space analysis analyzes the signal characteristics at different scales through Gaussian smoothing to identify local minima as initial spectral segmentation points. By processing the segmented spectral intervals, the frequency maxima of each interval are found, and these points are fitted into a continuous trend line using cubic Hermitian interpolation, serving as the spectral trend. Based on the spectral trend, all local minima are found as new boundary points, re-dividing the Fourier spectrum into multiple finer parts, corresponding to different modes or frequency components of the signal. The boundary points are normalized, and the corresponding empirical wavelet functions and empirical scaling functions are calculated based on the Meyer wavelet basis. These functions are used to construct bandpass filters to extract single frequency components from the original signal.

[0036] Optionally, the original vibration signal can be decomposed into a series of empirical mode functions (EMFs) using the bandpass filter constructed above. Each EMF represents a mode component within a specific frequency range of the signal, thereby achieving adaptive decomposition of the signal.

[0037] Optionally, after decomposing multiple modal components from the vibration signal, the weighted kurtosis index of each modal component can be calculated, the components related to the fault can be screened out for reconstruction, and the components with noise interference can be eliminated to improve the signal-to-noise ratio and ensure the clarity of fault features in subsequent analysis.

[0038] In this step, the target empirical wavelet transform algorithm is used to accurately capture different frequency information in the vibration signal. Through adaptive filter design, the signal is decomposed into multiple modal components, each of which corresponds to a certain frequency range of the signal. This decomposition helps to extract clear fault features from complex vibration signals, providing a high-quality data foundation for subsequent fault diagnosis.

[0039] Step S103: Use the feature extraction model to extract features from multiple modal components to obtain feature vectors.

[0040] In the technical solution provided in step S103 of the present invention, the feature extraction model is obtained by training a Convolutional Autoencoder (CAE) model. This training process includes unsupervised training and supervised training. In the unsupervised training phase, the CAE model learns the intrinsic representation of the data by minimizing the reconstruction error between the input and output. The training objective is to make the output of the decoder in the CAE model as close as possible to the input of the encoder in the CAE model; this similarity is typically measured using the mean squared error loss function. Through the backpropagation algorithm, the CAE model continuously adjusts its weights to optimize this reconstruction performance, thereby learning features useful for data compression and recovery. In supervised training, the CAE model is trained by adding a fully connected layer and a... The softmax classifier is trained as a multi-class deep learning model. At this stage, the CAE model not only needs to learn feature representations but also needs to correctly classify samples based on these features. By minimizing the cross-entropy loss function, the CAE model learns how to map extracted features to specific fault categories, a step that enhances the model's classification ability. After unsupervised and supervised training of the CAE model, the encoder part of the trained CAE model is used as a feature extraction model to extract features from multiple modal components. The feature vectors are used to represent the deep-level features related to overhead ground wire faults extracted from multiple modal components.

[0041] In this embodiment, since the encoder part of the CAE model includes convolutional layers and pooling layers, after the encoder in the trained CAE model is determined as a feature extraction model, the feature extraction model also includes convolutional layers and pooling layers. Therefore, when using the feature extraction model to extract features from multiple modal components, multiple modal components can be input into the feature extraction model, and the convolutional layers and pooling layers in the feature extraction model will convert the multiple modal components into compact feature representations.

[0042] Optionally, the convolutional layer performs convolution operations with multiple modal components using convolution kernels to extract local features from these modal components; the pooling layer, on the other hand, removes redundant information from the multiple modal components through dimensionality reduction operations, retaining key features to obtain a feature vector. This feature vector not only contains the intrinsic attributes of each modal component but has also been trained for classification, thus possessing high discriminative power and judgment value.

[0043] In this step, the feature extraction model can automatically learn deep and discriminative feature vectors from multiple modal components, providing accurate data for subsequent fault identification, thereby improving the performance and reliability of the entire fault diagnosis of overhead ground wires.

[0044] Step S104: Use the fault detection model to detect the feature vector and obtain the fault detection result.

[0045] In the technical solution provided in step S104 of the present invention, the fault detection model is obtained by training the Extreme Gradient Boosting (XGBoost) tree in advance using the Particle Swarm Optimization (PSO) algorithm. The fault detection result includes at least the fault type of the overhead ground wire and the fault location on the overhead ground wire.

[0046] In this embodiment, the feature vector is input into the XGBoost fault detection model optimized based on the PSO algorithm. The fault detection model can detect the feature vector according to the learned fault modes and classification rules, and output the fault detection results.

[0047] The following section describes the process of training XGBoost using the PSO algorithm to obtain a fault detection model.

[0048] Optionally, the PSO algorithm mimics the process of birds flying and foraging to find the optimal solution. Assume there is a target location in the region for the birds to find, and define m birds distributed at arbitrary locations in the feasible space. Each bird can be considered as a particle Z = {z1, z2, ... z...}. m Each particle's position represents a possible target position Z. i ={z i1 z i2 ...z in Particles determine their fitness for their current and target positions using a fitness function. Each time the flock iterates through its flight path, it records its best position (Pbest) and the best position found by the entire flock (Gbest). Each bird combines these two extreme positions to determine its next flight direction and speed (V). i ={v i1 v i2 ...v in They then forage again. This iterative process gradually shifts the foraging center of the entire flock towards the globally optimal direction, until they find the best overall position.

[0049] Optionally, when training the XGBoost algorithm using the PSO algorithm, the relevant parameters of the XGBoost model are first set, such as the learning rate, the penalty term coefficient (gamma), and the iteration parameters, inertia weights, and learning rate of the particle swarm algorithm.

[0050] Optionally, after setting the parameters of the XGBoost model and the particle swarm optimization algorithm, the PSO algorithm can be used to optimize key XGBoost parameters (such as learning_rate and gamma). For example, the PSO algorithm can update Pbest and Gbest, as well as the particle's velocity and position, based on the fitness of each particle (parameter combination) during XGBoost model training.

[0051] The velocity vector iteration formula can be expressed by the following formula:

[0052]

[0053] The position vector iteration formula can be expressed by the following formula:

[0054]

[0055] Where ω represents the inertia weight, which is between [0, 1]; c1 and c2 represent the learning factors; r1 and r2 represent random probability values ​​between [0, 1]; and k represents the number of iterations of PSO. Used to indicate speed Used to indicate location.

[0056] Optionally, the above iterative process can be repeated until the velocity vector and position vector reach the preset number of iterations or the fitness meets the termination condition.

[0057] Optionally, after the iteration stops, the final model parameters can be determined as the optimal model parameters, and these model parameters can be substituted into XGBoost to obtain the fault detection model.

[0058] Alternatively, the XGBoost model is an ensemble learning algorithm optimized based on Gradient Boosting Decision Tree (GBDT). It performs well in classification and regression tasks. Compared with other machine learning algorithms, XGBoost has significant advantages in handling small sample data, fast training, easy parameter tuning, and good model interpretability.

[0059] Alternatively, when training XGBoost, the objective function (loss function) of the XGBoost model can be defined as follows:

[0060]

[0061] Among them, y i , The labels and predicted values ​​of the i-th sample are used to represent the label and the predicted value of the sample; γ and λ are used to represent the weighting coefficients.

[0062] Optionally, the optimization effect of the later step depends on the result of the previous step. When finding the minimum value of the objective function, the objective function needs to be approximated and rewritten as follows:

[0063]

[0064] The s-th iteration model is based on the s-1th iteration model with an added sub-model f(x). i ).

[0065] Optionally, XGBoost simplifies the above equation by introducing a second-order Taylor expansion:

[0066]

[0067] In the above formula, Used to represent the first-order gradient of the objective function; Used to represent the second gradient of the objective function.

[0068] Optionally, constant terms that do not affect the optimization result are removed, and f is replaced with leaf node scores. s (x i The objective function can be rewritten as:

[0069]

[0070] The above equation can be viewed as having ω as the independent variable. i The dependent variable is Obj. (s) A quadratic function in one variable.

[0071] Among them, the optimal value of leaf node j And the optimal objective function value at this point is:

[0072]

[0073]

[0074] Optionally, after calculating the optimal objective function, the gradient of the error with respect to the parameters is calculated using the backpropagation algorithm, and then the model parameters are updated using the gradient descent algorithm until the set number of training iterations is reached. The model with the best performance is then selected as the final prediction model.

[0075] In steps S101 to S104 above, the vibration signal of the overhead ground wire is acquired, and the vibration signal is preprocessed using an improved empirical wavelet transform algorithm to remove redundant information and noise interference from the vibration signal, thereby obtaining multiple modal components in the vibration signal to improve the accuracy of subsequent feature extraction for fault diagnosis. After obtaining multiple modal components, the feature extraction model is used to further extract features from the multiple modal components to obtain feature vectors that characterize the deep-level features related to the fault of the overhead ground wire in the multiple modal components. Then, the fault detection model is used to detect the feature vectors to obtain the fault detection results. Since the fault detection model is pre-trained using the particle swarm optimization algorithm to train the ensemble learning algorithm, the fault detection model can more accurately identify faults in the overhead ground wire, improve the accuracy of the fault detection results, and thus achieve the technical effect of improving the fault detection accuracy of the overhead ground wire. This provides strong technical support for the safe operation and maintenance of the power system, thereby solving the technical problem of low fault detection accuracy of the overhead ground wire in related technologies.

[0076] The method described in this embodiment will be further described below.

[0077] As an optional embodiment, step S102 involves preprocessing the vibration signal based on the target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal. This includes: converting the vibration signal from the time domain to the frequency domain based on the fast Fourier transform algorithm to obtain the signal spectrum of the vibration signal; determining the spectral segmentation points of the signal spectrum based on the scale space method; dividing the signal spectrum into multiple spectral intervals based on the spectral segmentation points, wherein the multiple spectral intervals are used to represent different frequency components in the signal spectrum; fitting the frequency maxima points in each of the multiple spectral intervals based on the cubic Hermitian interpolation algorithm to generate a spectral envelope, wherein the spectral envelope is used to represent the trend of the spectrum; and using the local minima points on the spectral envelope as boundary points to perform a secondary division of the signal spectrum to obtain multiple modal components of the vibration signal.

[0078] In this embodiment, the vibration signal is converted from the time domain to the frequency domain using the Fast Fourier Transform (FFT) algorithm, which decomposes the vibration signal into the amplitude and phase of different frequency components to form a signal spectrum.

[0079] For example, the signal spectrum X(f) can be obtained by performing a fast Fourier transform on the vibration signal x(t) using the following formula.

[0080]

[0081] -i2πft

[0082] Here, X(f) represents the signal amplitude corresponding to frequency f, that is, the energy distribution of the frequency domain signal. e is used to convert the signal (x(t)) in the time domain to the frequency domain.

[0083] Optionally, after converting the vibration signal from the time domain to the frequency domain and obtaining the corresponding signal spectrum, the scale-space method can be used to determine the segmentation points of the signal spectrum. The scale-space method smooths the signal spectrum at different scales to find local minima. These local minima mark the boundaries of different frequency components, providing a basis for subsequently dividing the spectrum into multiple spectral intervals. Using the scale-space method, the segmentation boundaries of the spectrum can be adaptively determined, avoiding the limitations of pre-setting frequency bandwidth and improving the accuracy of modal component extraction.

[0084] Optionally, after determining the segmentation point, the signal spectrum can be divided into multiple spectral intervals, where each spectral interval is used to represent the components within a specific frequency range of a specific signal.

[0085] Optionally, after dividing the signal spectrum corresponding to the vibration signal into multiple spectral intervals, a cubic Hermite interpolation algorithm can be used to fit the frequency maxima in each spectral interval to generate a spectral envelope. The spectral envelope can clearly depict the trend of the signal spectrum, which helps to identify the main frequency components and potential fault modes in the signal.

[0086] Optionally, after obtaining the spectral envelope, the local minima on the spectral envelope can be considered as boundary points to perform a secondary division of the signal spectrum. This secondary division process can more accurately identify the boundaries of different frequency components in the signal, thereby effectively separating multiple modal components of the vibration signal. Each modal component represents a pure component in the signal that is not disturbed by other frequency components within a specific frequency range, which is helpful for subsequent feature extraction and fault diagnosis.

[0087] In this step, the preprocessing method based on target empirical wavelet transform can extract modal components carrying clear frequency features from complex vibration signals. These pure modal components not only facilitate subsequent feature extraction, such as deep learning feature learning using convolutional autoencoders (CAEs), but also significantly improve the accuracy and efficiency of fault diagnosis, such as pattern recognition using machine learning models like XGBoost. The entire preprocessing process fully utilizes signal spectral information, combining scale-space methods and smoothing filtering techniques, enabling adaptive processing of different types of vibration signals and extracting the most effective feature components for fault diagnosis.

[0088] As an optional implementation method, determining the spectrum segmentation points of a signal spectrum based on the scale space method includes: smoothing the signal spectrum using a Gaussian function to determine the spectrum segmentation points.

[0089] In this embodiment, the core of the scale-space method is to smooth the spectrum of the acquired vibration signal using a Gaussian kernel function. Each smoothing operation records the corresponding boundary points (local minima) on the processed spectrum. With each smoothing iteration, some spurious frequency band boundaries gradually disappear, while true boundaries are retained. After smoothing, a two-dimensional image of the scale and spectral boundaries can be plotted based on the recorded locations of local minima at different scales. In scale space, each initial minima corresponds to a scale-space curve. As the scale increases, the curves of spurious boundaries are interrupted, while those of true boundaries continue to extend. Using the IEWT method to find Fourier spectrum segmentation points involves finding meaningful modes in scale space. By setting a threshold, curves longer than the threshold are retained, and their corresponding boundary points are the spectrum segmentation points.

[0090] For example, optionally, the specific process of determining the segmentation points of the signal spectrum using the scale-space method is as follows: Assuming f(x) is a continuous signal, The kernel function, where t is set as the scale parameter. The scale space is defined as f(x). In this case, the scale space becomes smoother as t increases. For the acquired vibration signal, the scale space is defined as f(x). s Let x(n), (n=0,1,2....,N-1), be a discrete signal sequence with a sampling frequency, and let X(k) be the corresponding Fourier transform amplitude spectrum.

[0091]

[0092] in, X(k) is used to represent the signal spectrum after Fourier transform; X(k) is used to represent the amplitude spectrum after Fourier transform.

[0093] The discrete scale space representation of the Fourier spectrum X(k) is as follows:

[0094]

[0095] in: It is a discrete kernel function; This is the discretization scale parameter.

[0096] Alternatively, a truncation filter can be used to obtain a finite impulse response filter:

[0097]

[0098] Optionally, in the discrete scale space of the Fourier spectrum, L(k,t) is used to represent the function indicating that variable k is a local minimum. Spurious local minima gradually disappear as t increases. If n represents the number of times a local minimum occurs at different scales t, then C... i (i∈[1,n]) indicates that the length corresponding to this local minimum is L. i The “scale space curve”.

[0099] As an optional implementation, step S103 involves using a feature extraction model to extract features from multiple modal components to obtain feature vectors. This includes: using the convolutional layer of the feature extraction model to extract local features from multiple modal components, wherein the local features are features related to the fault of the overhead ground wire, and the local features are used to characterize the intensity and variation trend of multiple modal components at a specified location; and using the pooling layer of the feature extraction model to perform dimensionality reduction processing on the local features of multiple modal components to obtain feature vectors.

[0100] In this embodiment, the convolutional layer is a key network layer in the feature extraction model used to detect local patterns in the input data. When diagnosing faults in overhead ground wires, the convolutional layer can scan each modal component, identify the signal intensity and variation trend at specific locations, and thus obtain local features directly related to the fault type and severity of the overhead ground wire.

[0101] For example, taking any modal component as an example, the local features of that modal component can be extracted through a convolutional layer using the following formula.

[0102]

[0103] Where σ represents the activation function; The vector is used to represent the i-th feature vector of the l-th convolutional layer; N is used to represent the data dimension of the convolutional layer. Used to represent the weights of the convolution kernel; Used to represent bias.

[0104] Optionally, the activation function added to the convolutional layer as a non-linear transformation in the above formula serves to improve the network's expressive power and normalize the data to prevent data overflow. In this embodiment, the activation function used is the Softmax function, which can convert the output values ​​of multi-class classification into a probability distribution ranging from [0,1] with a sum of 1. The expression for Softmax is: Among them, z i The value of the i-th node is used to represent the output value of the i-th node; N is used to represent the number of output nodes.

[0105] Optionally, after extracting local features of multiple modal components using the convolutional layers of the feature extraction model, the pooling layers of the feature extraction model are used to reduce the dimensionality of the local features of multiple modal components to obtain feature vectors. That is, the pooling layer is located after the convolutional layer, and its main function is to perform feature dimensionality reduction while keeping the most salient features unchanged. For overhead ground wire fault diagnosis, the pooling layer helps to remove redundant information and focus on the key parts of the extracted local features.

[0106] For example, pooling layers reduce the dimensionality of feature vectors by applying pooling operations, such as max pooling or average pooling, to a single value over a large region on each feature map. This dimensionality reduction reduces computational cost and the risk of overfitting. Pooling operations typically select the most significant feature value (such as the maximum or average value) within a region for preservation, which helps highlight important features in the signal, such as peaks or anomalous trends in fault signals, while ignoring background noise or irrelevant details.

[0107] In this step, the combination of convolutional and pooling layers not only extracts fault-related local features from multiple modal components but also removes redundant information through dimensionality reduction, ultimately generating feature vectors that accurately describe signal characteristics. This significantly improves the efficiency and accuracy of fault diagnosis. This feature extraction method is a typical example of the application of deep learning in signal processing, demonstrating the powerful ability of deep learning models to automatically learn and extract complex signal features.

[0108] As an optional embodiment, step S104, which uses a fault detection model to detect feature vectors and obtain fault detection results, includes: inputting feature vectors into the fault detection model for fault detection; receiving the output results of the fault detection model; and determining the fault detection results based on the output results.

[0109] In this embodiment, the feature vector contains key information extracted from the vibration signal. This information, after dimensionality reduction and optimization, can effectively characterize the vibration signal characteristics of the overhead ground wire and potential fault modes. For example, the feature vector may include the signal intensity at different frequencies, its variation trend, and some statistical features such as kurtosis and skewness. The fault detection model can be an XGBoost classifier, which has been trained on a large amount of training data and can understand the correlation between different feature vectors and specific fault types. Based on this, the feature vector is input into the fault detection model, which analyzes the feature vector through its internal decision tree structure. Using the weights and thresholds learned during training, it gradually narrows down the range of fault types and finally provides a prediction result for one or more fault types. At each layer of the decision tree, the model judges specific feature values ​​in the feature vector and divides the data into different child nodes according to the judgment results, until the leaf node of the tree is reached. The classification result of the leaf node is the final fault detection result.

[0110] Optionally, after completing fault detection, the fault detection model will output one or more possible fault types and the probability of these fault types occurring. The output may be a direct fault type label or a probability distribution, the latter being more common in multi-class classification problems. For example, for fault diagnosis of overhead ground wires, the model may output the probability of fault types such as "loose conductor," "broken strand," and "lightning damage," or directly give the most likely fault type.

[0111] As an optional embodiment, the fault detection method for overhead ground wires further includes: acquiring vibration signal samples of the overhead ground wire; preprocessing the vibration signal samples to obtain multiple modal component samples of the vibration signal samples; dividing the multiple modal component samples into a first training set and a first test set according to a preset ratio; training a convolutional autoencoder feature extraction model using the first training set to obtain a trained convolutional autoencoder feature extraction model; validating the trained convolutional autoencoder feature extraction model using the first test set to obtain a first verification result, wherein the first verification result is used to evaluate the feature extraction capability of the trained convolutional autoencoder feature extraction model; and determining the trained convolutional autoencoder feature extraction model as a feature extraction model in response to the first verification result indicating that the feature reconstruction error of the trained convolutional autoencoder feature extraction model is less than a preset threshold, wherein the feature reconstruction error is used to characterize the degree of difference between the reconstructed signal output by the convolutional autoencoder feature extraction model and the input signal of the convolutional autoencoder feature model in the first test set.

[0112] In this embodiment, a series of vibration signal samples reflecting the operating status of the overhead ground wire can be obtained by sensors installed on the overhead ground wire or by signals collected by power system monitoring equipment. These samples contain vibration information under normal operation and different fault conditions. After obtaining the vibration signal samples of the overhead ground wire, the vibration signal samples can be preprocessed to eliminate noise and irrelevant interference in the signal, so as to extract fault features more accurately. The original vibration signal samples are decomposed into multiple modal component samples by an improved empirical wavelet transform (IEWT), and each component sample represents the characteristics of the vibration signal sample in a specific frequency band.

[0113] Optionally, after obtaining multiple modal component samples, the samples are divided into a first training set and a first test set according to a preset ratio. For example, 70% of the modal component samples can be designated as the first training set, and 30% as the first test set. Then, the convolutional autoencoder feature extraction model is trained using the first training set to obtain the trained convolutional autoencoder feature extraction model.

[0114] Optionally, after obtaining the trained convolutional autoencoder feature extraction model, it can be validated using a first test set to obtain a first validation result. This first validation result is used to evaluate the feature extraction and signal reconstruction capabilities of the trained convolutional autoencoder model. This first validation result typically includes the feature reconstruction error, i.e., the degree of difference between the reconstructed signal output by the trained convolutional autoencoder model and the original input signal in the test set. If the feature reconstruction error is small, it indicates that the trained convolutional autoencoder model can accurately extract features and reconstruct signals on the test data, demonstrating that the trained convolutional autoencoder model has good generalization ability.

[0115] Optionally, if the feature reconstruction error is determined to be less than a preset threshold based on the first validation result, this indicates that the performance of the trained convolutional autoencoder model has reached the expected standard and can effectively extract and reconstruct features from vibration signals. In this case, the trained convolutional autoencoder model is determined as the feature extraction model for subsequent fault detection tasks. If the first validation result does not meet the requirements, the model parameters of the convolutional autoencoder model need to be adjusted, retrained, or even the feature extraction method needs to be changed until the trained convolutional autoencoder model can meet the performance indicators.

[0116] As an optional implementation, the fault detection method for overhead ground wires further includes: obtaining multiple feature vector samples by performing feature extraction on multiple modal component samples in a first test set using a feature extraction model; dividing the multiple feature vector samples into a second training set and a second test set; inputting the second training set into an ensemble learning algorithm model and using a particle swarm optimization algorithm to obtain a target parameter combination of the ensemble learning algorithm model, wherein the target parameter combination includes at least a learning rate parameter and a regularization parameter; adjusting the parameters in the ensemble learning algorithm model using the parameters in the target parameter combination to obtain an adjusted ensemble learning algorithm model; verifying the adjusted ensemble learning algorithm model for fault identification using the second test set to obtain a second verification result, wherein the second verification result is used to indicate the performance of the adjusted ensemble learning algorithm model on the second test set; and determining the adjusted ensemble learning algorithm model as a fault detection model in response to the second verification result indicating that the accuracy of the adjusted ensemble learning algorithm model in identifying fault types is greater than a second preset threshold.

[0117] In this embodiment, after training and validating the convolutional autoencoder and confirming that its feature extraction capability meets the requirements, the convolutional autoencoder model is used to extract features from each modal component sample in the first test set, resulting in multiple feature vector samples. These feature vectors are compact representations of local features in the vibration signal samples that are highly correlated with the fault, providing higher-quality input data for subsequent model training.

[0118] Optionally, from the multiple feature vector samples, a second training set and a second test set are further divided according to a preset ratio (usually the same as the division ratio of the first training set and the first test set). The second training set is used to adjust and optimize the parameters of the ensemble learning algorithm model, while the second test set is used to verify the performance of the optimized model.

[0119] Optionally, after obtaining the second training set and the second test set, the second training set can be input into the ensemble learning algorithm model, and the particle swarm optimization algorithm can be used to obtain the target parameter combination of the ensemble learning algorithm model, wherein the target parameter combination includes at least the learning rate parameter and the regularization parameter.

[0120] Optionally, after obtaining the target parameter combination of the ensemble learning algorithm model using the particle swarm optimization algorithm, the ensemble learning algorithm model can be adjusted using the target parameter combination. Then, the adjusted ensemble learning algorithm model is validated for fault identification using a second test set, yielding a second validation result. This second validation result reflects the performance of the adjusted ensemble learning algorithm model on unseen data, including key indicators such as recognition accuracy. If the fault type recognition accuracy is greater than a preset second threshold, it indicates that the adjusted ensemble learning algorithm model performs well in identifying and distinguishing different fault types, meeting the needs of practical applications. Therefore, the adjusted ensemble learning algorithm model can be determined as the final fault detection model.

[0121] The technical solutions of the embodiments of the present invention will be illustrated below with reference to preferred embodiments.

[0122] With the continuous development of power systems, overhead ground wires play a crucial role in ensuring the stability and safety of power transmission. However, because overhead ground wires are exposed to the natural environment for extended periods, they are susceptible to various factors that can lead to failure.

[0123] In related technologies, commonly used intelligent algorithms for fault diagnosis of overhead ground wires include decision trees, backpropagation neural networks, support vector machines, K-nearest neighbors classifiers, and deep learning. However, most of these algorithms suffer from low accuracy, poor generalization ability, or are prone to overfitting.

[0124] However, this application provides a method for diagnosing overhead ground wire faults. It collects relevant operational data of the overhead ground wire and preprocesses the collected data to eliminate noise interference in the signal. The preprocessed data is divided into a training set and a test set. The CAE model is trained using the data in the training set. After training, the model parameters are saved, and a PSO-XGBoost fault diagnosis model is constructed. The training set is then input into the XGBoost fault diagnosis model for training until the model converges, obtaining the optimal XGBoost fault diagnosis model. The data in the test set is input into a CAE feature extraction model to extract data features, and the extracted data features are then input into the XGBoost fault diagnosis model for diagnosis, obtaining the fault diagnosis result of the overhead ground wire. In other words, by preprocessing the collected vibration signal, noise interference is effectively eliminated, improving the signal-to-noise ratio of the vibration signal. Using the trained CAE feature extraction model for feature extraction, deep-level features of the vibration signal can be extracted. These extracted features are then input into the XGBoost fault diagnosis model for diagnosis, thus obtaining the fault diagnosis result of the overhead ground wire and improving the real-time performance and accuracy of fault monitoring.

[0125] Figure 2This is a flowchart of an overhead ground wire fault diagnosis method according to an embodiment of this application. Figure 2 As shown, the method includes the following steps:

[0126] Step S201: Obtain the vibration signal of the overhead ground wire.

[0127] In this embodiment, vibration signals of the overhead ground wire are acquired by sensors installed on the overhead ground wire or from monitoring equipment of the power system.

[0128] Step S202, IEWT signal decomposition.

[0129] In this embodiment, the obtained secondary partition spectrum segmentation boundary is decomposed into several modal components by IEWT.

[0130] Step S203, weighted kurtosis index T.

[0131] In this embodiment, a weighted kurtosis index for each modal component is calculated.

[0132] Step S204: Signal reconstruction and noise reduction.

[0133] In this embodiment, the T>0 mode component is selected for reconstruction and noise reduction.

[0134] Step S205: Divide the samples.

[0135] In this embodiment, the reconstructed signal is divided into samples. For example, the data samples are divided into training set and test set in a 7:3 ratio, and fault category labels are added.

[0136] Step S206: Input the test set into the CAE feature extraction model.

[0137] In this embodiment, the training set data is fed into the CAE feature extraction model for training, and the model parameters are saved after training is completed.

[0138] Step S207, CAE model training.

[0139] In this embodiment, the Convolutional Autoencoder (CAE) is a network that combines the unsupervised learning training concept of autoencoders with the convolution and pooling operations of convolutional neural networks. Similar to autoencoders, the CAE also consists of an encoder and a decoder. During network training, the feature representation of the input data is obtained through convolution and pooling, and then decoded through transposed convolution to obtain effective features of the time series within the allowable range of reconstruction error. In the encoder of the CAE, one-dimensional convolutional layers and pooling layers completely replace the fully connected layers of traditional autoencoders, effectively reducing the number of trainable parameters in the network model and enhancing the ability to extract deep latent features from high-dimensional complex data. The CAE feature training process includes unsupervised training and supervised training.

[0140] Optionally, for unsupervised training, in this stage, the convolutional autoencoder measures the difference between the input samples and the decoded output by calculating the mean squared error loss function, and then optimizes the network parameters through backpropagation. The mean squared error loss function is defined as:

[0141]

[0142] Where Batch is the number of samples fed into the model during one training iteration; n is the dimension of the samples; x b,i For each element of the input data; This is for decoding the elements of the data.

[0143] Optionally, for supervised fine-tuning, the network parameters are optimized at this stage using the cross-entropy loss function between the softmax layer output and the true label. Cross-entropy reflects the difference between the true and predicted distributions, and the cross-entropy loss function is defined as:

[0144]

[0145] Where M is the number of categories; N is the number of samples; y ic p is a symbolic function; ic Let c be the predicted probability that the observed sample belongs to c.

[0146] Step S208: Does the CAE model meet the iteration termination condition?

[0147] In this embodiment, after training the CAE model, it can be further determined whether the trained CAE model meets the iteration termination condition. If the iteration termination condition is met, step S209 is executed; otherwise, step S207 is executed.

[0148] Alternatively, the convergence of the loss function can be used to determine whether the CAE model meets the iteration termination condition. For example, during training, the CAE model continuously adjusts its parameters to minimize the loss function, such as mean squared error (MSE) or cross-entropy loss, using the backpropagation algorithm. When the change in the loss function is less than a preset threshold for several consecutive iterations, i.e., when the loss function reaches a stable state, the model can be considered to have converged, and training can be stopped at this point.

[0149] Alternatively, the CAE model can be judged to determine whether it meets the iteration termination condition by checking whether the CAE model has reached the preset number of training iterations.

[0150] Step S209: Training complete and save the CAE model.

[0151] In this embodiment, if the trained CAE model meets the iteration termination condition, the CAE model is saved.

[0152] Step S210: Input the test set into the trained CAE model.

[0153] In this embodiment, a test set can be input into the trained CAE model for feature extraction. Feature extraction can be achieved using convolutional layers and pooling layers. For example, local features are extracted using convolutional layers, and dimensionality selection is performed on the output features of the convolutional layers using pooling layers.

[0154] Step S211: Output the feature dataset.

[0155] In this embodiment, the data feature set extracted in step S210 can be output.

[0156] Step S212, training set features.

[0157] In this embodiment, 70% of the output feature dataset can be used as the training set.

[0158] Step S213, test set features.

[0159] In this embodiment, 30% of the output feature dataset can be used as the test set.

[0160] Step S214: PSO optimizes the parameters of XGBoost.

[0161] In summary, the PSO algorithm can be used to optimize the model parameters of the XGBoost model.

[0162] Step S215, XGBoost training.

[0163] In this embodiment, the XGBoost model can be trained using data from the training set.

[0164] Step S216: Does the XGBoost model meet the iteration termination condition?

[0165] In this embodiment, after training the XGBoost model, it can be further determined whether the XGBoost model meets the iteration termination condition. If the XGBoost model meets the iteration termination condition, step S217 is executed; if the XGBoost model does not meet the iteration termination condition, step S214 is executed.

[0166] Alternatively, it can be determined whether the XGBoost model meets the iteration stopping condition by judging whether the XGBoost model has reached the maximum number of iterations.

[0167] Optionally, the change in the loss function of the XGBoost model in each iteration can be monitored. The iteration can be stopped when the change in the loss function is less than a preset threshold, indicating convergence.

[0168] Step S217: Training complete and save the XGBoost model.

[0169] In this embodiment, an XGBoost model that satisfies the iteration termination condition can be saved.

[0170] Step S218: Use the trained XGBoost model to perform fault classification.

[0171] In this embodiment, features from the test set can be input into a trained XGBoost model for fault classification.

[0172] In steps S201 to S218 above, the optimal CAE feature extraction model and the optimal XGBoost fault diagnosis model are obtained by training the CAE feature extraction model and the XGBoost model respectively. The trained CAE feature extraction model is then used for feature extraction. This not only automatically learns the deep features of the signal, reducing reliance on expert knowledge, but also ensures the discriminative power and effectiveness of the extracted features through a combination of unsupervised and supervised training, laying a solid foundation for subsequent fault diagnosis. Using the extracted features for fault classification through the trained XGBoost fault diagnosis model can improve the accuracy of the diagnostic results.

[0173] Figure 3 This is a flowchart illustrating a method for optimizing an XGBoost model using a particle swarm optimization algorithm according to an embodiment of this application, as shown below. Figure 3 As shown, the method includes the following steps:

[0174] Step S301: Randomly initialize the particle swarm.

[0175] In this embodiment, the particle swarm optimization algorithm mimics the process of birds flying and foraging to find the optimal solution. This algorithm is easy to implement, has high accuracy, and converges quickly, and is widely used. Figure 4 This is a schematic diagram of a particle swarm optimization algorithm according to an embodiment of this application. Assuming there is a target location in a region for a flock of birds to find, firstly, define m numbers of birds distributed at arbitrary locations in the feasible space. Each bird can be considered as a particle Z = {z1, z2, ... zn}. m Each particle's position represents a possible target position Z. i ={z i1 z i2 ...z in Particles determine their fitness for their current and target positions using a fitness function. Each time the flock iterates its flight path, it records its best position (called "local optimum Pbest") and the best position found by the entire flock (called "global optimum Gbest"). Each bird combines these two extreme positions to determine its next flight direction and speed V. i ={v i1 v i2 ...v in They then forage again. This iterative process gradually shifts the foraging center of the entire flock towards the globally optimal direction, until they find the best overall position.

[0176] Optionally, random initialization of the particle swarm optimization (PSO) can be understood as randomly setting the relevant parameters of XGBoost. This includes setting parameters such as the number of iterations, inertia weights, and learning rate for the PSO algorithm. It also involves setting the feasible range for the learning rate and penalty term coefficients during the optimization process.

[0177] Step S302: Calculate the fitness of each particle.

[0178] In this embodiment, the fitness of each particle is calculated iteratively using the particle swarm optimization algorithm.

[0179] Step S303: Update Pbest and Gbest based on fitness, and update the particle's movement speed and position.

[0180] In this embodiment, the particle swarm optimization algorithm updates Pbest and Gbest based on fitness, as well as the particle's velocity and position. The velocity vector iteration formula is as follows:

[0181]

[0182] The position vector iteration formula is:

[0183]

[0184] In the above formula, ω is the inertia weight, which is between [0,1]; c1 and c2 are learning factors; r1 and r2 are random probability values ​​between [0,1] and [0,1]; and k is the number of iterations of PSO.

[0185] Step S304: Determine whether the XGBoost model meets the iteration stopping condition.

[0186] In this embodiment, after updating the particle's moving speed and position, it can be determined whether the XGBoost model meets the iteration conditions. If the iteration stopping condition is met, the entire process ends; otherwise, it returns to step S302. The method for determining whether the XGBoost model meets the iteration stopping condition can refer to the aforementioned step S216.

[0187] In steps S301 to S304 above, optimizing the parameters of the XGBoost model using the particle swarm optimization algorithm can significantly improve the model's performance, stability, and generalization ability, while reducing the risk of overfitting, making the model more effective and reliable in complex problems such as overhead ground wire fault diagnosis.

[0188] Figure 5 This is a schematic diagram of a fault detection device for an overhead ground wire according to an embodiment of the present invention. Figure 5 As shown, the fault detection device 500 for the overhead ground wire may include: an acquisition unit 501, a processing unit 502, an extraction unit 503, and a detection unit 504.

[0189] The acquisition unit 501 is used to acquire the vibration signal of the overhead ground wire, wherein the vibration signal is used to at least characterize the operating status of the overhead ground wire.

[0190] The processing unit 502 is used to preprocess the vibration signal based on the target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal. The target empirical wavelet transform algorithm is an improvement on the empirical wavelet transform algorithm, and different modal components are used to represent different frequency components in the vibration signal.

[0191] Extraction unit 503 is used to extract features from multiple modal components using a feature extraction model to obtain feature vectors. The feature extraction model is obtained by training a convolutional autoencoder model, and the feature vectors are used to characterize the deep-level features related to the fault of the overhead ground wire in multiple modal components.

[0192] The detection unit 504 is used to detect the feature vector using a fault detection model to obtain the fault detection result. The fault detection model is obtained by pre-training the ensemble learning algorithm model using the particle swarm optimization algorithm. The fault detection result includes at least the fault type of the overhead ground wire and the fault location on the overhead ground wire.

[0193] Optionally, the processing unit 502 further includes: a conversion module for converting the vibration signal from the time domain to the frequency domain based on the fast Fourier transform algorithm to obtain the signal spectrum of the vibration signal; a determination module for determining the spectral segmentation points of the signal spectrum based on the scale space method; a first partitioning module for dividing the signal spectrum into multiple spectral intervals based on the spectral segmentation points, wherein the multiple spectral intervals are used to represent different frequency components in the signal spectrum; a fitting module for fitting the frequency maxima points in each of the multiple spectral intervals based on the cubic Hermite interpolation algorithm to generate a spectral envelope, wherein the spectral envelope is used to represent the trend of the spectrum; and a second partitioning module for using the local minima points on the spectral envelope as boundary points to perform secondary partitioning of the signal spectrum to obtain multiple modal components of the vibration signal.

[0194] Optionally, the first partitioning module is also used to: smooth the signal spectrum using a Gaussian function to determine the spectrum partitioning point.

[0195] Optionally, the extraction unit 503 is further configured to: extract local features of multiple modal components using the convolutional layer of the feature extraction model, wherein the local features are features related to the fault of the overhead ground wire, and the local features are used to characterize the intensity and change trend of multiple modal components at a specified location; and perform dimensionality reduction processing on the local features of multiple modal components using the pooling layer of the feature extraction model to obtain feature vectors.

[0196] Optionally, the detection unit 504 is further configured to: input the feature vector into the fault detection model for fault detection; receive the output result of the fault detection model; and determine the fault detection result based on the output result.

[0197] Optionally, the fault detection device 500 for overhead ground wires is further configured to: acquire vibration signal samples of the overhead ground wires; preprocess the vibration signal samples to obtain multiple modal component samples of the vibration signal samples; divide the multiple modal component samples into a first training set and a first test set according to a preset ratio; train the convolutional autoencoder feature extraction model using the first training set to obtain the trained convolutional autoencoder feature extraction model; verify the trained convolutional autoencoder feature extraction model using the first test set to obtain a first verification result, wherein the first verification result is used to evaluate the feature extraction capability of the trained convolutional autoencoder feature extraction model; and, in response to the first verification result indicating that the feature reconstruction error of the trained convolutional autoencoder feature extraction model is less than a preset threshold, determine the trained convolutional autoencoder feature extraction model as a feature extraction model, wherein the feature reconstruction error is used to characterize the degree of difference between the reconstructed signal output by the convolutional autoencoder feature extraction model and the input signal of the convolutional autoencoder feature model in the first test set.

[0198] Optionally, the fault detection device for overhead ground wires is further configured to: acquire multiple feature vector samples obtained by the feature extraction model from multiple modal component samples in the first test set; divide the multiple feature vector samples into a second training set and a second test set; input the second training set into the ensemble learning algorithm model and use the particle swarm optimization algorithm to obtain the target parameter combination of the ensemble learning algorithm model, wherein the target parameter combination includes at least a learning rate parameter and a regularization parameter; adjust the parameters in the ensemble learning algorithm model using the parameters in the target parameter combination to obtain the adjusted ensemble learning algorithm model; perform fault identification verification on the adjusted ensemble learning algorithm model using the second test set to obtain a second verification result, wherein the second verification result is used to indicate the performance of the adjusted ensemble learning algorithm model on the second test set; and, in response to the second verification result indicating that the accuracy of the adjusted ensemble learning algorithm model in identifying fault types is greater than a second preset threshold, determine the adjusted ensemble learning algorithm model as a fault detection model.

[0199] In this embodiment, the vibration signal of the overhead ground wire is acquired, and the vibration signal is preprocessed using an improved empirical wavelet transform algorithm to remove redundant information and noise interference, thereby obtaining multiple modal components in the vibration signal. This improves the accuracy of subsequent feature extraction for fault diagnosis. After obtaining multiple modal components, a feature extraction model is used to further extract features from the multiple modal components, obtaining feature vectors that characterize the deep-level features related to faults in the overhead ground wire. Then, a fault detection model is used to detect the feature vectors to obtain fault detection results. Since the fault detection model is pre-trained using a particle swarm optimization algorithm on an ensemble learning algorithm, it can more accurately identify faults in the overhead ground wire, improving the accuracy of fault detection results. This achieves the technical effect of improving the fault detection accuracy of the overhead ground wire, providing strong technical support for the safe operation and maintenance of the power system, and thus solving the technical problem of low fault detection accuracy of the overhead ground wire in related technologies.

[0200] According to an embodiment of the present invention, a computer-readable storage medium is also provided, the storage medium including a stored program, wherein the program executes the overhead ground wire fault detection method in the embodiment.

[0201] According to an embodiment of the present invention, a processor is also provided for running a program, wherein the program executes the overhead ground wire fault detection method in the embodiment.

[0202] According to another aspect of the present invention, a computer program product is also provided. The program product includes computer instructions that, when executed by a processor, implement the overhead ground wire fault detection method described in the embodiments.

[0203] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0204] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0205] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0206] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0207] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0208] If the integrated unit is implemented as a software functional unit and sold or used as an independent functional component, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software functional component. This computer software functional component is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0209] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fault detection method for overhead ground wires, characterized in that, include: The vibration signal of the overhead ground wire is acquired, wherein the vibration signal is used to at least characterize the operating state of the overhead ground wire; The vibration signal is preprocessed based on the target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal. The target empirical wavelet transform algorithm is an improvement on the empirical wavelet transform algorithm. Different modal components are used to represent different frequency components in the vibration signal. Feature extraction models are used to extract features from the multiple modal components to obtain feature vectors. The feature extraction model is obtained by training a convolutional autoencoder model. The feature vectors are used to characterize the deep-level features related to the fault of the overhead ground wire in the multiple modal components. The feature vector is detected using a fault detection model to obtain a fault detection result. The fault detection model is pre-trained using a particle swarm optimization algorithm on an ensemble learning algorithm model. The fault detection result includes at least the fault type of the overhead ground wire and the fault location on the overhead ground wire. The preprocessing of the vibration signal based on the target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal includes: transforming the vibration signal from the time domain to the frequency domain based on the fast Fourier transform algorithm to obtain the signal spectrum of the vibration signal; determining the spectral segmentation points of the signal spectrum based on the scale space method; dividing the signal spectrum into multiple spectral intervals based on the spectral segmentation points, wherein the multiple spectral intervals are used to represent different frequency components in the signal spectrum; fitting the frequency maxima points in each of the multiple spectral intervals based on the cubic Hermitian interpolation algorithm to generate a spectral envelope, wherein the spectral envelope is used to represent the trend of the spectrum; and using the local minima points on the spectral envelope as boundary points to perform a secondary division of the signal spectrum to obtain the multiple modal components of the vibration signal.

2. The method according to claim 1, characterized in that, Determining the spectral segmentation points of the signal spectrum based on the scale-space method includes: The signal spectrum is smoothed using a Gaussian function to determine the spectrum segmentation point.

3. The method according to claim 1, characterized in that, Feature extraction models are used to extract features from the multiple modal components to obtain feature vectors, including: The convolutional layers of the feature extraction model are used to extract local features of the multiple modal components, wherein the local features are features related to the fault of the overhead ground wire, and the local features are used to characterize the intensity and variation trend of the multiple modal components at a specified location; The feature vector is obtained by using the pooling layer of the feature extraction model to perform dimensionality reduction on the local features of the multiple modal components.

4. The method according to claim 1, characterized in that, The feature vector is detected using a fault detection model to obtain fault detection results, including: The feature vector is input into the fault detection model for fault detection. Receive the output of the fault detection model; Based on the output, the fault detection result is determined.

5. The method according to claim 1, characterized in that, The method further includes: Obtain vibration signal samples of the overhead ground wire; The vibration signal sample is preprocessed to obtain multiple modal component samples of the vibration signal sample; According to a preset ratio, the multiple modal component samples are divided into a first training set and a first test set; The convolutional autoencoder model is trained using the first training set to obtain the trained convolutional autoencoder model. The trained convolutional autoencoder model is validated using the first test set to obtain a first validation result, wherein the first validation result is used to evaluate the feature extraction capability of the trained convolutional autoencoder model. In response to the first verification result indicating that the feature reconstruction error of the trained convolutional autoencoder model is less than a preset threshold, the trained convolutional autoencoder model is determined as the feature extraction model, wherein the feature reconstruction error is used to characterize the degree of difference between the reconstructed signal output by the convolutional autoencoder model and the input signal of the convolutional autoencoder model in the first test set.

6. The method according to claim 5, characterized in that, The method further includes: The feature extraction model is used to extract features from multiple modal component samples in the first test set, resulting in multiple feature vector samples. The multiple feature vector samples are divided into a second training set and a second test set; The second training set is input into the ensemble learning algorithm model, and the particle swarm optimization algorithm is used to obtain the target parameter combination of the ensemble learning algorithm model, wherein the target parameter combination includes at least a learning rate parameter and a regularization parameter; The parameters in the ensemble learning algorithm model are adjusted using the parameters in the target parameter combination to obtain the adjusted ensemble learning algorithm model; The adjusted ensemble learning algorithm model is used to perform fault identification verification using the second test set to obtain a second verification result, wherein the second verification result is used to indicate the performance of the adjusted ensemble learning algorithm model on the second test set; In response to the second verification result indicating that the accuracy of the adjusted ensemble learning algorithm model in identifying fault types is greater than a second preset threshold, the adjusted ensemble learning algorithm model is determined as the fault detection model.

7. A fault detection device for an overhead ground wire, characterized in that, include: An acquisition unit is used to acquire vibration signals of an overhead ground wire, wherein the vibration signals are used to at least characterize the operating state of the overhead ground wire; The processing unit is used to preprocess the vibration signal based on the target empirical wavelet transform algorithm to obtain multiple modal components of the vibration signal, wherein the target empirical wavelet transform algorithm is an improvement on the empirical wavelet transform algorithm, and different modal components are used to represent different frequency components in the vibration signal; An extraction unit is used to extract features from the multiple modal components using a feature extraction model to obtain feature vectors. The feature extraction model is obtained by training a convolutional autoencoder model, and the feature vectors are used to characterize the deep-level features related to the fault of the overhead ground wire in the multiple modal components. The detection unit is used to detect the feature vector using a fault detection model to obtain a fault detection result. The fault detection model is obtained by pre-training an ensemble learning algorithm model using a particle swarm optimization algorithm. The fault detection result includes at least the fault type of the overhead ground wire and the fault location of the fault on the overhead ground wire. The processing unit is further configured to preprocess the vibration signal based on the target empirical wavelet transform algorithm through the following steps to obtain multiple modal components of the vibration signal: transforming the vibration signal from the time domain to the frequency domain based on the fast Fourier transform algorithm to obtain the signal spectrum of the vibration signal; determining the spectral segmentation points of the signal spectrum based on the scale space method; dividing the signal spectrum into multiple spectral intervals based on the spectral segmentation points, wherein the multiple spectral intervals are used to represent different frequency components in the signal spectrum; fitting the frequency maxima points in each of the multiple spectral intervals based on the cubic Hermitian interpolation algorithm to generate a spectral envelope, wherein the spectral envelope is used to represent the trend of the spectrum; and using the local minima points on the spectral envelope as boundary points to perform a secondary division of the signal spectrum to obtain the multiple modal components of the vibration signal.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein when the program is run by a processor, it controls the device in which the storage medium is located to perform the method of any one of claims 1 to 6.

9. A processor, characterized in that, The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 6 when it runs.

Images