Power transmission line hidden discharge feature extraction method and system based on deep learning

Abstract

The application discloses a power transmission line hidden danger discharge feature extraction method and system based on deep learning, and comprises the following steps: acquiring a power transmission line hidden danger discharge current waveform in historical records; performing variational mode decomposition on the acquired hidden danger discharge current to extract an intrinsic mode function; constructing a multi-scale convolutional neural network, and training a feature extraction model using different types of power transmission line hidden danger discharge currents; and extracting power transmission line hidden danger discharge features by using the trained power transmission line hidden danger discharge feature extraction neural network model. The combination of VMD and deep learning is used to extract the frequency domain and time domain multi-scale features of the hidden danger discharge signal respectively, so that the efficient extraction of the power transmission line hidden danger discharge features can be realized, and the safe and stable operation of the power transmission line is ensured.

Description

Technical Field

[0001] This invention relates to the field of discharge feature extraction technology, specifically to a method and system for extracting discharge features of hidden dangers in transmission lines based on deep learning. Background Technology

[0002] With my country's rapid economic development and surging electricity demand, the power grid is growing larger and the distances between power lines are increasing. However, during the long-term operation of overhead transmission lines, potential problems may arise such as insulator contamination, floating hardware potential, and insufficient line-to-line spacing. If these problems are not eliminated in time, they may lead to line tripping faults.

[0003] After a potential discharge occurs, it is crucial to identify the discharge type as quickly as possible to provide maintenance departments with a reference for line inspection and to promptly eliminate the discharge point on the transmission line. However, current prior knowledge about potential discharges is limited. On one hand, manually analyzing the discharge waveform type through inspection would place an excessive burden on personnel; on the other hand, it is difficult for humans to make timely and accurate judgments on a large volume of potential discharge currents on transmission lines. Using a feature extraction model for potential discharges on transmission lines to automatically extract features from the discharge waveform can quickly achieve deep feature extraction of potential discharges, thereby facilitating subsequent identification of the discharge type. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that existing methods for identifying potential hazards have problems such as high cost and slow feature extraction speed.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning-based method for extracting discharge features of hidden dangers in transmission lines, comprising:

[0007] Obtain the waveform of the discharge current of potential hazards in transmission lines from historical records;

[0008] The obtained hidden discharge current is subjected to variational mode decomposition to extract the intrinsic mode function;

[0009] Construct a multi-scale convolutional neural network and train a feature extraction model using the discharge current of potential hazards in transmission lines of different types.

[0010] The characteristics of potential discharge hazards in transmission lines are extracted using a trained neural network model for extracting these hazards.

[0011] As a preferred embodiment of the deep learning-based method for extracting features of hidden discharge in transmission lines according to the present invention, the hidden discharge current waveform includes the hidden discharge current waveform retrieved from the transmission line recording system as sample data.

[0012] Each sample data is labeled and coded according to the type of potential discharge.

[0013] As a preferred embodiment of the deep learning-based method for extracting discharge features of hidden dangers in transmission lines according to the present invention, the variational mode decomposition includes defining an optimization problem: the objective is to decompose the signal f(t) into K modes, each mode u k Around a center frequency ω k ;

[0014] This is achieved by minimizing the objective function, which means:

[0015]

[0016] in, is u k The Fourier transform of the k-th mode, where k represents the k-th mode. This represents the partial derivative with respect to time t, where j represents the imaginary part of the imaginary number, and t represents time.

[0017] Introduce constraints:

[0018]

[0019] Where f represents the original signal;

[0020] Constructing the Lagrange function:

[0021]

[0022] Where α represents the regularization parameter; λ represents the Lagrange multiplier, used to ensure that the decomposed modes can reconstruct the original signal;

[0023] Alternate updates via ADMM k ω k , until convergence.

[0024] As a preferred embodiment of the deep learning-based method for extracting discharge features of hidden dangers in transmission lines according to the present invention, the multi-scale convolutional neural network specifically comprises a multi-scale convolutional neural network with three scales, including three parallel convolutional modules of different scales. Each convolutional module includes one convolutional layer and one batch normalization layer. In the second multi-scale convolutional module, a max pooling operation is added after the batch normalization layer to downsample the mined features.

[0025] As a preferred embodiment of the deep learning-based method for extracting features of hidden discharges in transmission lines according to the present invention, the multi-scale convolutional neural network further includes performing convolution operations on the input hidden discharge IMFs, assuming... This represents the i-th convolutional kernel in the l-th layer. Let j represent the j-th local region in the l-th layer. Then the convolution operation process is as follows:

[0026]

[0027] Among them, y l(i,j) represents the dot product of the kernel and the receptive region; W represents the width of the convolution kernel; X represents the i-th convolutional kernel in the l-th layer; l ( Rj ) represents the j-th local region of the l-th layer, where the convolution kernel performs convolution operations; j' represents a position index; X l(j+j′) This represents a specific location on the input feature map of the l-th layer; This represents the value of the i-th convolutional kernel in the l-th layer at position j';

[0028] Batch normalization is performed on the result of the first convolution operation:

[0029]

[0030]

[0031]

[0032] Where, d l(i,j) μ represents the output of a neuron; μ represents y. l(i,j) The average value; σ 2 Indicates y l(i,j) The variance; ε is a constant used to prevent invalid calculations when the variance is 0;

[0033] Perform the second convolution operation, the process of which is the same as the first convolution operation;

[0034] Perform max pooling operation:

[0035]

[0036] Among them, a l(i,t) It is the value of the t-th neuron within the i-th visual field of the l-th layer, w m It is the width of the max pooling window. This is the result after max pooling. This is the result after average pooling, and Maxpool represents the max pooling operation.

[0037] As a preferred embodiment of the deep learning-based method for extracting features of hidden discharges in transmission lines according to the present invention, the step of training the feature extraction model using different types of hidden discharge currents in transmission lines includes inputting the extracted hidden discharge IMFs into the corresponding convolutional neural network branches C. s In the diagram, each branch outputs a feature map F. s :

[0038] F s =C s (IMFs)

[0039] Establish a connection operation through a fusion function: multi-scale feature F s The features are integrated into the final feature representation F through the fusion function g;

[0040] F = g(F1,F2,…,F) s )

[0041] Calculate the loss and update the network parameters: given the predicted output Calculate the cross-entropy loss with the true label y;

[0042]

[0043] During training, the Adam optimization algorithm is used to update the network parameters θ based on the gradient of the loss function.

[0044]

[0045] Where η represents the learning rate; IMFs represent the intrinsic mode components decomposed from the signal; g represents the fusion function, which is the concatenation operation of F1, F2, ...; C s θ represents a branch in a convolutional neural network; θ represents the parameters in the network; θ new Indicates the updated parameter; θ old This represents the parameters before the update; L represents the cross-entropy loss mentioned above. Let L represent the gradient of the loss L with respect to θ.

[0046] As a preferred embodiment of the deep learning-based method for extracting features of hidden discharge in transmission lines according to the present invention, the step of training the feature extraction model using different types of hidden discharge currents in transmission lines further includes testing the network with test samples, evaluating the model performance, and determining whether the model has reached stability; if the performance is stable, the training ends.

[0047] The performance evaluation includes assessing the proportion of samples correctly predicted by the model to the total number of samples; after the network is tested, the loss value is calculated according to the cross-entropy loss formula. When the network's loss value tends to converge, with no fluctuation exceeding the fluctuation ratio H, and the network's prediction accuracy also reaches the convergence index, the network is determined to be stable.

[0048] A deep learning-based system for extracting discharge features of hidden dangers in transmission lines, employing any of the methods described in this invention, is characterized by:

[0049] The data acquisition module obtains the discharge current waveform of potential hazards in transmission lines from historical records.

[0050] The training module performs variational mode decomposition on the acquired hidden discharge currents to extract intrinsic mode functions; it constructs a multi-scale convolutional neural network and trains the feature extraction model using hidden discharge currents from different types of transmission lines.

[0051] The extraction module uses a trained neural network model for extracting the discharge features of hidden dangers in transmission lines to extract the discharge features of hidden dangers in transmission lines.

[0052] A computer device includes: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0053] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the present invention.

[0054] The beneficial effects of this invention are as follows: The deep learning-based method for extracting discharge features of hidden dangers in transmission lines provided by this invention has the advantage of being able to determine the number of mode decompositions (IMFs). Its adaptability is manifested in determining the number of IMFs for a given sequence based on the actual situation. In the subsequent search and solution process, it can adaptively match the optimal center frequency and finite bandwidth of each mode, and can achieve effective separation of IMFs, frequency domain partitioning of the signal, and thus obtain the effective decomposition components of the given signal, ultimately obtaining the optimal solution to the variational problem. It overcomes the problems of endpoint effects and mode component aliasing in empirical mode decomposition methods, and has a more solid mathematical theoretical foundation. It can reduce the non-stationarity of time series with high complexity and strong nonlinearity, and decompose to obtain relatively stationary subsequences containing multiple different frequency scales, making it suitable for non-stationary sequences. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 The overall flowchart of the deep learning-based method for extracting hidden discharge features of transmission lines is provided in the first embodiment of the present invention.

[0057] Figure 2 This is a comparison chart of the load balancing rates of various algorithms under the same task request in the deep learning-based method for extracting discharge features of hidden dangers in transmission lines, provided in the second embodiment of the present invention. Detailed Implementation

[0058] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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 protection scope of the present invention.

[0059] Example 1

[0060] Reference Figure 1 As an embodiment of the present invention, a method for extracting discharge features of hidden dangers in transmission lines based on deep learning is provided, including:

[0061] S1: Obtain the waveform of discharge current for potential hazards in transmission lines from historical records.

[0062] Furthermore, the waveform of the discharge current of the hidden danger is retrieved from the transmission line recording system as sample data; each sample data is labeled and coded according to the type of discharge of the hidden danger.

[0063] The purpose of using a transmission line waveform recording system to retrieve the current waveform of potential discharge hazards is to obtain actual transmission line status data. This data reflects the current waveform characteristics of the transmission line during potential discharge hazards. The waveform recording system can capture the dynamic changes in the current waveform, providing accurate basic data for subsequent data analysis and feature extraction. Labeling each sample data according to the type of potential discharge hazard is to associate each waveform sample with its corresponding discharge type. This labeling not only helps supervised learning during the training process, but also facilitates effective classification and identification after model training, even if the model can learn the association between different types of discharge waveforms and specific discharge types. These two steps are the foundation for building an effective machine learning model, especially when dealing with complex and high-dimensional power system data. By accurately capturing and labeling data, the model's learning ability can be enhanced, thereby more accurately identifying and predicting potential hazards in the power grid, and ultimately improving the reliability and safety of the power system.

[0064] S2: Perform variational mode decomposition on the acquired hidden discharge current and extract the intrinsic mode functions.

[0065] Furthermore, VMD was used to extract IMFs at different center frequencies from the potential discharge current, transforming the original data from the time domain to time-frequency composite data. The VMD calculation process is as follows:

[0066] (1) Define the optimization problem. The goal is to decompose the signal f(t) into K modes, each mode u k Around a center frequency ω k This is achieved by minimizing the objective function, as shown in the following formula:

[0067]

[0068] in, is u k The Fourier transform of; k represents the k-th mode, just like i in the usual accumulation symbol, except that here it is decomposed into K modes, so k is used. This represents taking the partial derivative with respect to time t. j is the imaginary part of the imaginary number, and t is time. Since the signal is f(t), it means that the signal is a quantity related to time t.

[0069] (2) Introducing constraints. To ensure that the decomposed modal integration can reconstruct the original signal, the following constraints are introduced:

[0070]

[0071] Here, f represents the original signal, i.e., f(t) mentioned above.

[0072] (3) Construct the Lagrangian function. To simultaneously satisfy the optimization objective and constraints, the Lagrangian function is constructed as follows:

[0073]

[0074] Where α represents the regularization parameter; λ represents the Lagrange multiplier, used to ensure that the decomposed modes can reconstruct the original signal.

[0075] (4) Update u alternately using ADMM. k ω k λ continues until convergence. Specifically, the Alternating Direction Multiplier Method (ADMM) is applied. Through ADMM, u is updated alternately. k ω k λ until convergence:

[0076] Update u k : Fix other variables and update u k To minimize the Lagrange function.

[0077] Update ω k : Keep other variables fixed, update ω k To minimize the Lagrange function.

[0078] Update λ: Update the Lagrange multiplier λ according to the degree of constraint violation.

[0079] (5) Repeat the ADMM update process until the convergence condition is met: the difference between two iterations is less than the preset threshold.

[0080] S3: Construct a multi-scale convolutional neural network and train the feature extraction model using the discharge current of potential hazards in transmission lines of different types.

[0081] A multi-scale convolutional neural network with three scales is constructed, which contains three parallel convolutional modules of different scales. Each convolutional module contains one convolutional layer and one batch normalization layer. In the second multi-scale convolutional module, a max pooling operation is added after the batch normalization layer to downsample the mined features.

[0082] Perform convolution operation on the input hidden discharge IMFs, and set... This represents the i-th convolutional kernel in the l-th layer. Let j represent the j-th local region in the l-th layer. Then the convolution operation process is as follows:

[0083]

[0084] Among them, y l(i,j) The dot product of the kernel and the receptive region is represented by , and W represents the width of the convolution kernel. This represents the weight of the convolution kernel l in the j-th region; This represents the i-th convolutional kernel in the l-th layer. This represents the j-th local region of the l-th layer, where the convolution kernel performs convolution operations. j' represents a position index, X l(j+j′) This represents a specific location on the input feature map of layer l. This represents the value of the i-th convolutional kernel in the l-th layer at position j'. Therefore, for example, the formula means that if the input data X is a signal, then X... l(j+j′) This refers to a small segment of signal surrounding the l-th layer and j-th position in the signal. During the sliding process of the convolutional kernel Kl, each element of itself... Multiply the input feature map Xl by the elements at the corresponding positions j+j', and then sum all the products to obtain the feature map y. l(i,j) The value at position (i,j). The receptive region is the region where the convolutional kernel is located, because convolutional neural networks are inspired by the working principle of the animal retina.

[0085] Batch normalization is performed on the result of the first convolution operation:

[0086]

[0087]

[0088]

[0089] Where, d l(i,j) μ represents the output of a neuron; μ represents y. l(i,j) The average value; σ 2 Indicates y l(i,j) The variance; ε is a constant used to prevent invalid calculations when the variance is 0;

[0090] Perform the second convolution operation, the process of which is the same as the first convolution operation;

[0091] Perform max pooling operation:

[0092]

[0093] Among them, a l(i,t) It is the value of the t-th neuron within the i-th visual field of the l-th layer, w m It is the width of the max-pooling window, w a It is the width of the average pooling window. This is the result after max pooling. This is the result after average pooling. Maxpool represents the max pooling operation. l(i,t)Represents an element of the input feature map (before the pooling operation), located at position (i,t), with the remaining letters explained in the context.

[0094] Next, the feature extraction model was trained using the discharge current of potential hazards in different types of transmission lines. The calculation process is as follows:

[0095] The extracted hidden danger discharge IMFs are input into the corresponding convolutional neural network branches C. s In the diagram, each branch outputs a feature map F. s :

[0096] F s =C s (IMFs)

[0097] Establish a connection operation through a fusion function: multi-scale feature F s The features are integrated into the final feature representation F through the fusion function g.

[0098] F = g(F1,F2,…,F) s )

[0099] Calculate the loss and update the network parameters: given the predicted output Calculate the cross-entropy loss using the true label y.

[0100]

[0101] During training, the Adam optimization algorithm is used to update the network parameters θ based on the gradient of the loss function.

[0102]

[0103] Where η represents the learning rate. IMFs are the intrinsic mode components derived from signal decomposition; here, IMFs are used directly to represent them. C s This represents a branch in a convolutional neural network. Because it takes IMFs as input to the network, it is denoted by C. s (IMFs) represents the fusion function, which is the operation on the most eigenvectors F1, F2, ... (F represents eigenvectors, as explained in the text). Here, g means performing a concatenation operation on F1, F2, ..., that is, merging the vectors that are joined end to end.

[0104] θ refers to the parameters in the network. new These are the updated parameters, θ old These are the parameters before the update, where L represents the cross-entropy loss mentioned above. Let L represent the gradient of the loss L with respect to θ.

[0105] Furthermore, the network is tested using test samples to evaluate its performance and determine if it has reached stability; if the performance is stable, training ends. The evaluation metric is usually accuracy, which is the proportion of correctly predicted samples out of the total number of samples. After testing, the loss value is calculated using the cross-entropy loss formula mentioned above. When the network's loss value essentially converges (i.e., no longer fluctuates significantly), and the network's prediction accuracy also tends to converge, the network can be considered stable.

[0106] S4: Use a trained neural network model for extracting the discharge features of hidden dangers in transmission lines to extract the discharge features of hidden dangers in transmission lines.

[0107] The electrical signal for which feature extraction is required is input into a trained neural network model for extracting features of hazardous discharges in transmission lines. The model outputs the extracted features, namely the feature vector of the hazardous discharge current. This feature vector encompasses the multi-scale time-frequency features of the input hazardous discharge signal. After outputting the feature vector, it can be stored to establish a hazardous discharge feature database, or used as input to an intelligent classification algorithm to classify hazardous discharge types.

[0108] On the other hand, this embodiment also provides a deep learning-based system for extracting the discharge features of hidden dangers in transmission lines, which includes:

[0109] The data acquisition module obtains the discharge current waveform of potential hazards in transmission lines from historical records.

[0110] The training module performs variational mode decomposition on the acquired hidden discharge currents to extract intrinsic mode functions; it constructs a multi-scale convolutional neural network and trains the feature extraction model using hidden discharge currents from different types of transmission lines.

[0111] The extraction module uses a trained neural network model for extracting the discharge features of hidden dangers in transmission lines to extract the discharge features of hidden dangers in transmission lines.

[0112] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product 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 described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0114] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0115] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0116] Example 2

[0117] Reference Figure 2 As an embodiment of the present invention, a method for extracting the discharge features of hidden dangers in transmission lines based on deep learning is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0118] Reference Figure 2 The process involves extracting the discharge features of the input signal. Specifically, first, VMD is performed on the original data to extract the IMFs (Integrated Frequency Components). IMFs represent the signal components at different center frequencies, thus implicitly containing frequency domain features at different scales. Next, the IMFs are input into a multi-scale convolutional neural network (VMN) to perform convolution operations and extract their time-domain features. Different convolution kernel sizes can extract time-domain features at different scales. Therefore, the multi-scale convolutional neural network can extract the multi-scale time-domain features of the IMFs. Through VMD processing and the multi-scale convolutional neural network, the multi-scale time-frequency feature information of the original data is extracted.

[0119] The performance of this invention was compared through comparative experiments. First, a simulated power transmission line environment was established, and a simulator was used to generate different types of potential discharge current waveforms. These waveforms mimic various discharge conditions that may occur in actual power transmission lines, including discharges caused by internal insulation degradation, poor contact, moisture corrosion, and other problems.

[0120] Each type of discharge waveform data was preprocessed using variational mode decomposition (VMD) to decompose each waveform into multiple intrinsic mode functions (IMFs). Next, a multi-scale convolutional neural network model was constructed and trained. This model was specifically designed with three parallel convolutional modules of different scales to process these IMFs, aiming to capture key features across different frequency ranges.

[0121] During model training, a label encoding strategy was employed to classify waveform samples of different hazard types, and the cross-entropy loss function and Adam optimizer were used for network training. In the experiments, performance data of the model was collected during the training and validation phases to evaluate its ability to identify various types of hazard discharges.

[0122] A control group was set up to perform the same task using traditional signal processing techniques (such as Fourier transform and simple statistical learning methods; the results of the traditional method were obtained through a mixed test of the two methods) for effective comparison. After data collection, the performance of the two methods will be evaluated and compared using a series of metrics, including accuracy, recall, F1 score, and processing time. The deep learning model was trained with a small number of samples, while the augmented deep learning model was trained with a much larger number of samples.

[0123] Table 1 Experimental Data

[0124]

[0125]

[0126] It is evident that the multi-scale convolutional neural network model of this invention demonstrates superiority in identifying the discharge characteristics of potential hazards in transmission lines. As can be seen from the accuracy, recall, and F1 score, the deep learning model shows significant improvements across all performance metrics compared to traditional methods. Particularly noteworthy is the accuracy, where the deep learning model achieves 95%, while the traditional method only reaches 80%, and the enhanced deep learning model further improves to 98%, indicating the potential for further improvement through optimization of the network structure and training process.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting hidden discharge features of transmission lines based on deep learning, characterized in that, include: Obtain the waveform of the discharge current of potential hazards in transmission lines from historical records; The obtained hidden discharge current is subjected to variational mode decomposition to extract the intrinsic mode function; Construct a multi-scale convolutional neural network and train a feature extraction model using the discharge current of potential hazards in transmission lines of different types. The characteristics of hidden discharge in transmission lines are extracted using a trained neural network model for extracting hidden discharge features in transmission lines. The multi-scale convolutional neural network is specifically constructed by building a multi-scale convolutional neural network with three scales, including three parallel convolutional modules of different scales. Each convolutional module includes one convolutional layer and one batch normalization layer. In the second multi-scale convolutional module, a max pooling operation is added after the batch normalization layer to downsample the mined features. The method of training the feature extraction model using different types of transmission line hazard discharge currents includes inputting the extracted hazard discharge IMFs into the corresponding convolutional neural network branches. In the middle, each branch outputs a feature map. : Establish connection operations through fusion functions: multi-scale features The features are integrated into the final feature representation F through the fusion function g; Where IMFs represent the intrinsic mode components decomposed from the signal; g represents the fusion function, which is the concatenation operation of F1, F2, ...; C s This represents a branch in a convolutional neural network.

2. The method for extracting hidden discharge features of transmission lines based on deep learning as described in claim 1, characterized in that: The hidden discharge current waveform includes sample data retrieved from the transmission line recording system. Each sample data is labeled and coded according to the type of potential discharge.

3. The method for extracting discharge features of hidden dangers in transmission lines based on deep learning as described in claim 2, characterized in that: The variational mode decomposition includes defining an optimization problem: the objective is to transform the signal... Decomposed into K modes, each mode Around a center frequency ; This is achieved by minimizing the objective function, which means: in, yes The Fourier transform of the k-th mode, where k represents the k-th mode. This represents the partial derivative with respect to time t, where j represents the imaginary part of the imaginary number and t represents time. Introduce constraints: Where f represents the original signal; Constructing the Lagrange function: in, Represents the regularization parameter; This represents the Lagrange multiplier, used to ensure that the decomposed modes can reconstruct the original signal; Alternate updates via ADMM , , Until convergence.

4. The method for extracting hidden discharge features of transmission lines based on deep learning as described in claim 3, characterized in that: The multi-scale convolutional neural network further includes performing convolution operations on the input hidden danger discharge IMFs, setting... This represents the i-th convolutional kernel in the l-th layer. Let j represent the j-th local region in the l-th layer. Then the convolution operation process is as follows: in, represents the dot product of the kernel and the receptive region; W represents the width of the convolution kernel; This represents the i-th convolutional kernel in the l-th layer; represents the j-th local region of the l-th layer, where the convolution kernel performs convolution operations; j' represents a position index. This represents a specific location on the input feature map of the l-th layer; This represents the value of the i-th convolutional kernel in the l-th layer at position j'; Batch normalization is performed on the result of the first convolution operation: in, This represents the output of a neuron; express The average value; express The variance; This is a constant used to prevent invalid calculations when the variance is 0; Perform the second convolution operation, the process of which is the same as the first convolution operation; Perform max pooling operation: in, It is the value of the t-th neuron within the i-th visual field of the l-th layer. It is the width of the max pooling window. This is the result after max pooling. This is the result after average pooling, and Maxpool represents the max pooling operation.

5. The method for extracting hidden discharge features of transmission lines based on deep learning as described in claim 4, characterized in that: Calculate the loss and update the network parameters: given the predicted output and real labels Calculate the cross-entropy loss; The Adam optimization algorithm is used during training to update network parameters based on the gradient of the loss function. : in, Indicates the learning rate; Represents parameters in the network; This indicates the updated parameters; This represents the parameters before the update; L represents the cross-entropy loss mentioned above. Indicates loss L to The gradient.

6. The method for extracting hidden discharge features of transmission lines based on deep learning as described in claim 5, characterized in that: The method of training the feature extraction model using the discharge current of hidden dangers in different types of transmission lines also includes testing the network with test samples, evaluating the model performance, and determining whether the model has reached stability. Training ends if performance is stable. The performance includes evaluating the proportion of samples correctly predicted by the model out of the total number of samples; After the network is tested, the loss value is calculated according to the cross-entropy loss formula. When the network's loss value tends to converge, with no fluctuation exceeding the fluctuation ratio H, and the network's prediction accuracy also reaches the convergence index, the network is determined to be stable.

7. A deep learning-based system for extracting discharge features of hidden dangers in transmission lines, employing the method described in any one of claims 1-6, characterized in that: The data acquisition module obtains the discharge current waveform of potential hazards in transmission lines from historical records. The training module performs variational mode decomposition on the acquired hidden discharge currents to extract intrinsic mode functions; it constructs a multi-scale convolutional neural network and trains the feature extraction model using hidden discharge currents from different types of transmission lines. The extraction module uses a trained neural network model for extracting the discharge features of hidden dangers in transmission lines to extract the discharge features of hidden dangers in transmission lines.

8. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of a method for extracting discharge features of hidden dangers in transmission lines based on deep learning.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of a deep learning-based method for extracting discharge features of hidden dangers in power transmission lines.

Images