An intelligent detection method and system for arrester faults

Abstract

The present invention discloses an intelligent detection method and system for arrester faults. The method includes: arranging a sensor network around the target arrester device to comprehensively monitor the working environment and operating parameters of the device; using the data collected by the sensor network to perform a series of preprocessing operations; adopting an improved one-dimensional signal denoising processing algorithm of OMLSA to denoise the multi-sensor data so as to remove the noise in the signal; sending the denoised data into a residual network based on channel convolution for model training, and using a genetic algorithm to search for the optimal parameters of the model; combining online learning and offline learning algorithms to realize real-time update and optimization of the model; and performing error determination after optimizing the model through a particle swarm annealing fusion algorithm. This application has significant beneficial effects in improving efficiency, accuracy, accurately grasping the health status, and enhancing adaptability compared with the traditional threshold determination method.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial fault detection, and particularly to an intelligent detection method and system for arrester faults. Background Art

[0002] It is very important to monitor the operating status of equipment in real time. However, there are some problems with traditional arrester fault detection methods, and it is impossible to accurately judge the health status of the equipment:

[0003] 1. Manually designed features: Traditional arrester fault detection methods usually require manually designed features, which is a very time-consuming and laborious task. Moreover, due to the limited experience and knowledge of people, it is difficult to design features that can effectively capture the characteristics of fault signals.

[0004] 2. Poor model generalization ability: Traditional fault diagnosis technologies usually adopt simple models, such as statistical models and machine learning models. These models have poor generalization ability and are prone to overfitting or underfitting problems in practical applications. This means that the model may not be able to adapt well to new data or situations, resulting in a decrease in the accuracy and reliability of fault detection.

[0005] 3. Difficulty in processing multi-channel sensor signals: With the development of sensor Internet of Things technology, processing multi-channel sensor signals has become a challenge for traditional fault diagnosis technologies. Traditional technologies may not be able to effectively process a large amount of sensor data or extract useful information from these data.

[0006] These problems limit the accuracy and reliability of traditional methods and cannot meet the requirements for accurately grasping the health status of equipment in modern industrial production. Summary of the Invention

[0007] In view of this, the present invention proposes an intelligent detection method and system for arrester faults to solve the problem that it is difficult to accurately grasp the health status of equipment in the prior art.

[0008] The specific technical solution of the present invention is as follows:

[0009] An intelligent detection method for arrester faults, comprising the following steps:

[0010] Step 1, arranging a sensor network around the target arrester device to comprehensively monitor the working environment and operating parameters of the device;

[0011] Step 2, using the data collected by the sensor network to perform a series of preprocessing operations, including data cleaning and data matching, to remove invalid or incorrect data and ensure the accuracy and consistency of the data;

[0012] Step 3, use the improved one-dimensional signal denoising algorithm of OMLSA to denoise the multi-sensor data, so as to remove the noise in the signal and retain the useful information;

[0013] Step 4, send the denoised data into the residual network based on channel convolution for model training. Through training, the model can learn the relationship between the operating state of the arrester and the environmental parameters, and use the genetic algorithm to search for the optimal parameters of the model;

[0014] Step 5, combine the online learning and offline learning algorithms to realize the real-time update and optimization of the model;

[0015] Step 6, optimize the model through the particle swarm annealing fusion algorithm and then conduct error determination. When the error is within the acceptable range, use the model for online prediction to monitor the operating state of the arrester in real time and discover potential fault risks;

[0016] Step 7, when the input changes greatly or meets other set conditions, end the whole process.

[0017] Specifically, the OMLSA algorithm in Step 3 includes the following execution steps: First, collect the target signal, including voice signals or one-dimensional time signals; then use the IMCRA algorithm based on maximum likelihood estimation to estimate the noise and effectively process non-stationary noise; then perform denoising processing on the signal according to the least mean square error criterion; thereafter, use the Bayesian estimation method to optimize the quality of the denoised signal; then perform weighted processing on different time and frequency components by adding a time-frequency mask to the optimization objective function to enhance the recognition of the denoised signal; at the same time, add the prior probability model of the signal to the objective function to regularize the optimization result to improve the stability of the denoised signal; further, model at different frequency resolutions, process signal components of different scales, and combine the multi-scale results to further improve the denoising performance; finally, introduce a feedback structure to realize multi-stage denoising optimization.

[0018] Specifically, enhancing the recognition of the denoised signal is achieved through a one-dimensional time signal enhancement algorithm based on adaptive gain, including the following steps: First, initialize the gain parameter, select the appropriate gain type and its parameter values; then use a fixed filter to filter the input one-dimensional time signal to reduce noise and other interferences; then calculate the gain parameter according to the filtered signal; then update the calculated gain parameter to a new value; finally, repeat the filtering, calculation, and update steps until the gain converges or reaches the preset number of iterations.

[0019] Specifically, the recognition of the signal after enhanced noise reduction is achieved through a gain calculation method: in the case of linear gain, the formula g = s_n / s_m is used to calculate the gain, where g represents the gain, s_n represents the filtered signal, and s_m represents the mean value of the signal; in the case of non-linear gain, the formula g = f(s_n) is used to calculate the gain, where f represents the non-linear function of the gain and s_n represents the filtered signal.

[0020] Specifically, the recognition of the signal after enhanced noise reduction is achieved through a one-dimensional time signal enhancement algorithm of adaptive filtering. The execution steps of this algorithm include: initializing the filter, selecting an appropriate filter type and assigning initial parameters; applying the initialized filter to the input one-dimensional time signal for filtering; calculating the parameters of the filter according to the filtered signal; updating the value of the filter according to the calculated filter parameters; repeating the filtering, calculation, and update steps until the filter converges, so as to dynamically adjust the filter parameters according to the signal characteristics and obtain a better one-dimensional time signal enhancement effect.

[0021] Specifically, the calculation of the filter parameters is achieved through a filter parameter calculation method based on the one-dimensional data of the lightning arrester: for a linear filter, the formula h_k = s_k / s_m is used to calculate the filter coefficient, where h_k represents the filter coefficient, s_k represents the filtered signal, and s_m represents the mean value of the signal; for a non-linear filter, the formula h_k = f(s_k) is used to calculate the filter parameters, where f represents the non-linear function of the filter and s_k represents the filtered signal.

[0022] Specifically, in step 4, the data after noise reduction processing is sent into a residual network based on channel convolution for model training, including model input, model output, and model calculation process. The model input includes the input joint feature x, the convolution kernel w, the gated attention weight k_g, and the self-attention weight k_s; the model output is the model output probability matrix y; in the model calculation process, through the calculation of the gated attention mechanism and the self-attention mechanism, a new context vector and output are obtained; the gated attention mechanism includes a gating mechanism and an attention mechanism, and the self-attention mechanism includes the calculation of the query matrix, key matrix, value matrix, and attention score, as well as multi-head calculation and post-processing.

[0023] Specifically, in step 4, a genetic algorithm is used for optimization, and the specific implementation steps are as follows: Initialize a group of particle swarms representing the weights and structure of the neural network, and initialize the velocity and position of the particles; Calculate the fitness of each particle to measure the quality of the particle; In each iteration, update the velocity and position of the particles until all particles in the particle swarm converge to the global optimal solution; After the particle swarm algorithm converges, use the simulated annealing algorithm to further optimize the neural network, generate new solutions and accept or reject them, and gradually reduce the annealing temperature as the number of iterations increases; Repeat the particle swarm algorithm and the simulated annealing algorithm until the preset number of iterations is reached or other termination conditions are met.

[0024] Specifically, in step 5, a model online learning strategy is used for continuous optimization and update, and the specific implementation steps are as follows: Initialize the model parameters; Obtain new input samples from the data stream; Use the current model parameters for prediction or inference; Calculate the error loss between the model output and the true label; Update the model parameters through gradient descent of the loss function; Evaluate the performance of the model on the evaluation set at regular intervals; Save the current best model according to the evaluation metrics; Continuously obtain new samples and update the model to achieve continuous optimization.

[0025] This application also proposes an intelligent arrester fault detection system, including: a sensor network arrangement module for arranging a sensor network around the target arrester device; a data preprocessing module for preprocessing and cleaning the collected data; a noise reduction processing module using an improved one-dimensional signal noise reduction processing algorithm of OMLSA to perform noise reduction on multi-sensor data; a model training module for sending the noise-reduced data into a residual network based on channel convolution for model training, and using a genetic algorithm to search for the optimal parameters of the model; a model update module for realizing real-time update and optimization of the model by combining online learning and offline learning algorithms; an error determination module for performing error determination after optimizing the model through a particle swarm annealing fusion algorithm; an online prediction module for performing online prediction within an acceptable error range to monitor the operation status of the arrester in real time and discover potential fault risks; and a process control module for ending the entire process when the input changes greatly or other set conditions are met.

[0026] The beneficial effects of the present invention are as follows:

[0027] 1. Improve efficiency: The traditional threshold determination method needs to collect a large amount of data and analyze it to determine the fault type, while this application only needs to collect a small amount of data and learn the fault characteristics through the neural network, thus greatly improving the efficiency of fault diagnosis;

[0028] 2. Improve accuracy: This application can find the optimal parameters of the neural network through the particle swarm optimization algorithm, thus improving the accuracy of fault diagnosis;

[0029] 3.Accurately grasp the health status: This application can learn fault characteristics through neural networks, so as to accurately grasp the health status of the device;

[0030] 4.Enhance adaptability: When it is found that the data distribution of the arrester changes greatly during the arrester fault reasoning process, the arrester reasoning end side will collect data in real time, send it into the model for training, and then re-implement the learning and reasoning of the model for new data, so as to ensure that the model always adapts to the data changes and provides more accurate fault diagnosis results. Brief Description of the Drawings

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0032] Figure 1 It is a schematic flow chart of the intelligent arrester fault detection method of the present invention;

[0033] Figure 2 It is a schematic flow chart of the optimization of the particle swarm joint annealing algorithm of the present invention;

[0034] Figure 3 It is a schematic application flow chart in the actual intelligent arrester fault detection process of the present invention. Detailed Description of the Embodiments

[0035] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the following further details the present invention in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0036] With the continuous progress of deep learning technology, its application in the field of arrester fault detection has attracted much attention. Deep learning technology can solve the challenges faced by traditional fault detection methods because it has the following advantages:

[0037] First of all, deep learning does not require manual feature design. In traditional fault detection methods, feature extraction is a crucial step, and appropriate features need to be designed and selected manually. However, deep learning can automatically extract useful feature representations from large-scale data, thus avoiding the cumbersome process of manual feature design.

[0038] Secondly, deep learning can model large amounts of data. Lightning arrester fault detection requires processing large amounts of data, including sensor data, historical data, etc. Deep learning can process large-scale datasets and learn useful feature representations and models from them, thereby improving the accuracy and reliability of fault detection.

[0039] Finally, deep learning is very good at processing multi-dimensional matrix information. In the lightning arrester fault diagnosis scenario, it is usually necessary to process multi-dimensional data, such as sensor data, time series data, etc. Deep learning can process this multi-dimensional matrix information to better understand and analyze the fault state of the lightning arrester.

[0040] In summary, the lightning arrester fault diagnosis technology based on deep learning can solve the challenges faced by traditional fault detection methods and improve the accuracy and reliability of fault detection. With the continuous development of deep learning technology, its application prospect in the field of lightning arrester fault detection will be broader.

[0041] The present invention proposes an intelligent detection method for lightning arrester faults, as Figure 1 shown, including the following steps:

[0042] Step 1. Arrange the sensor network and build an industrial Internet system: First, we arrange a sensor network around the target device to comprehensively monitor the working environment, operating parameters, etc. of the device. Through these sensors, we can collect data in real time and build an industrial Internet system to provide a basis for subsequent data processing and analysis.

[0043] Step 2. Data collection and preprocessing: Using the data collected by the sensor network, we carried out a series of preprocessing work. First, we cleaned the data to remove invalid or incorrect data. Then, we performed data matching to ensure the accuracy and consistency of the data. These preprocessing steps provide a high-quality dataset for subsequent model training.

[0044] Step 3. Data noise reduction processing: In order to improve the quality of the data, we used an improved one-dimensional signal noise reduction processing algorithm of OMLSA to perform noise reduction on multi-sensor data. This algorithm can effectively remove the noise in the signal and retain useful information, providing a more accurate basis for subsequent analysis.

[0045] Step 4. Model training and optimization: We sent the data after noise reduction processing into a residual network based on channel convolution for model training. Through training, the model can learn the relationship between the device operating state and environmental parameters, providing support for subsequent fault prediction. At the same time, we used a genetic algorithm to search for the optimal parameters of the model to achieve the optimization of the model structure.

[0046] Step 5. Online learning and real-time update: At the algorithm level, we combine online learning and offline learning algorithms to achieve real-time update and optimization of the model. Online learning enables the model to continuously learn new data and information during operation, improving the accuracy of prediction. Offline learning, on the other hand, can be used to conduct in-depth analysis of historical data to discover potential fault patterns and rules.

[0047] Step 6. Error determination and online prediction: After optimizing the model through the particle swarm annealing fusion algorithm, we conduct error determination. If the error is within the acceptable range, the model can be used for online prediction. Online prediction can monitor the operating status of the device in real time, promptly detect potential fault risks, and provide decision-making support for device maintenance and repair.

[0048] Step 7. End and summary: When there are significant changes in the input or other end conditions are met, we end the entire process.

[0049] The above method can be summarized into three stages in the actual application process, as Figure 3 shown, to achieve comprehensive processing of arrester fault diagnosis:

[0050] First, in the first stage, large-scale offline collection of arrester data is carried out to ensure obtaining rich and diverse data samples. Subsequently, offline learning algorithms are used to train the model with these data, enabling the model to learn the potential relationship between arrester data and faults. The focus of this stage is to establish a preliminary fault diagnosis model.

[0051] Next, entering the second stage, parameter optimization of the model obtained in the first stage is carried out. By using the particle swarm and simulated annealing algorithms, we can search for the optimal parameters of the model, further improving the performance and accuracy of the model. The optimization process in this stage makes the model more adaptable to the actual needs of arrester fault diagnosis.

[0052] Finally, in the third stage, the optimized model is deployed to the actual application scenario. During this period, we closely monitor the distribution of arrester data. Once a significant change in the data distribution is found, the online learning mechanism is immediately activated. By collecting data in real time and training the model on the arrester diagnosis side, we can ensure that the model promptly adapts to the new data distribution and continuously provides accurate fault diagnosis. Once the model learns the new arrester data distribution, it will be reapplied to arrester fault diagnosis to ensure continuous diagnostic effectiveness.

[0053] All in all, through the sequential implementation of the three stages, this method realizes a complete process from data collection, model training to model optimization and online learning, thus improving the accuracy and efficiency of arrester fault diagnosis.

[0054] In an industrial environment, we often use various sensors to collect data, such as sound, temperature, pressure, etc. However, this data often contains a lot of noise, just like the background noise we hear, and this noise will interfere with our acquisition of useful information.

[0055] To remove this noise, we need to perform noise reduction processing. Noise reduction processing is to separate the noise from the data through some algorithms and technologies, so that we can hear or see useful information more clearly.

[0056] When processing sound data, a commonly used method is the algorithm based on the time domain. This method directly processes the time series of the sound. For example, we can find a time point, and then intercept the sound data before and after this time point as new training samples.

[0057] Another method is the algorithm based on the frequency domain. This method converts the sound into a spectrum and then processes it on the spectrum. For example, we can add some noise to the sound or change the frequency of the sound, etc., to generate new training samples.

[0058] In step 3, the specific description of the one-dimensional signal noise reduction processing algorithm of the improved OMLSA is as follows:

[0059] The OMLSA algorithm, whose full name is optimally-modified log-spectral amplitude, has an excellent technical background when dealing with speech noise reduction and other one-dimensional signals. This algorithm is based on the MMSE criterion, Bayesian estimation, and IMCRA noise estimation algorithm, aiming to improve the signal processing performance.

[0060] The following are the execution steps of the OMLSA algorithm:

[0061] Step 31. Collect the signal: First, collect the target signal, including speech signals or one-dimensional time signals, such as leakage current or temperature.

[0062] Step 32. Noise estimation: Use the IMCRA algorithm to estimate the noise. This algorithm is based on maximum likelihood estimation and can effectively process non-stationary noise.

[0063] Step 33. Signal noise reduction: According to the MMSE criterion, minimize the mean square error between the estimated value and the true value to perform noise reduction processing on the signal.

[0064] Step 34. Bayesian estimation: Use the Bayesian estimation method to calculate the posterior probability according to the prior probability and the likelihood function, and further optimize the quality of the noise-reduced signal.

[0065] Step 35. Time-frequency mask processing: Add the time-frequency mask W(t,f) to the optimization objective function to perform weighted processing on different time and frequency components, so as to enhance the recognition of the denoised signal.

[0066] Step 36. Prior model constraint: Add the prior probability model P(s) of the signal to the objective function to regularize the optimization result and improve the stability of the denoised signal.

[0067] Step 37. Multi-scale processing: Model at different frequency resolutions, process signal components of different scales, and further improve the denoising performance by combining multi-scale results.

[0068] Step 38. Recursive optimization: By introducing a feedback structure, use the result of the previous round to initialize the next round of optimization to achieve multi-stage denoising optimization.

[0069] Through the above steps, the OMLSA algorithm can be effectively applied to speech denoising and other processing scenarios of one-dimensional signals, improving the signal processing performance.

[0070] After performing denoising processing, we also need to enhance the data. Data enhancement is to generate new and more representative training data through some technical means. For example, we can translate the sound data along the time axis by a certain distance, or add some noise to the sound. Both denoising processing and data enhancement are to make our data clearer and more accurate, so as to better extract useful information.

[0071] In step 35, to enhance the recognition of the denoised signal, one of the following methods can be used to achieve it:

[0072] (1) A one-dimensional time signal enhancement algorithm based on adaptive gain, which is used to enhance one-dimensional time signals. This algorithm can dynamically adjust the gain parameters according to the characteristics of the signal, so as to obtain a better enhancement effect. The specific implementation steps are as follows:

[0073] Step 3511. Initialize the gain parameter: First, it is necessary to select appropriate gain parameters, which can be linear gain or non-linear gain. For linear gain, the gain can be initialized to 1; for non-linear gain, a suitable initial value can be selected.

[0074] Step 3512. Filter the signal: Use a fixed filter to filter the input one-dimensional time signal to reduce noise and other interferences.

[0075] Step 3513. Calculate the gain parameter: Calculate the gain parameter according to the filtered signal. This can be achieved by analyzing the characteristics of the signal or using other algorithms.

[0076] Step 3514. Update the gain parameter: Update the calculated gain parameter to a new value. This update process can be iterative step by step until the gain converges or reaches a preset threshold.

[0077] Step 3515. Repeat steps 3512 - 3514: Repeat steps 3512 - 3514 until the gain converges or reaches a preset number of iterations. This process can be implemented through a loop structure.

[0078] Through the above steps, the one - dimensional time signal enhancement algorithm based on adaptive gain can automatically adjust the gain parameter to better adapt to the characteristics of different signals, thereby obtaining a better enhancement effect. This algorithm has the advantages of being simple and easy to implement, with a small computational load, and can be widely applied to the processing and enhancement of various one - dimensional time signals.

[0079] (2) A gain calculation method for the one - dimensional data of lightning arresters can be determined according to the characteristics of the signal and the purpose of enhancement. For linear gain, the following formula can be used for calculation: g = \frac{s_n}{s_m}, where: g is the gain; s_n is the filtered signal; s_m is the mean value of the signal. For non - linear gain, the following formula can be used for calculation: g = f(s_n), where: f is the non - linear function of the gain; s_n is the filtered signal.

[0080] This gain calculation method can select appropriate formulas and parameters according to specific situations to obtain a better enhancement effect. At the same time, this method also has the advantages of being simple and easy to implement, with a small computational load, and can be widely applied to the processing and enhancement of the one - dimensional data of lightning arresters.

[0081] (3) A one - dimensional time signal enhancement algorithm based on adaptive filtering can dynamically adjust the parameters of the filter according to the characteristics of the signal, thereby obtaining a better enhancement effect. The following are the specific implementation steps of this algorithm:

[0082] Step 3531. Initialize the filter: Select an appropriate filter type and assign initial parameters to it. For a linear filter, initialize the filter coefficients; for a non - linear filter, initialize the filter parameters.

[0083] Step 3532. Filter the signal: Apply the initialized filter to the input one - dimensional time signal to filter the signal.

[0084] Step 3533. Calculate the filter parameters: Calculate the parameters of the filter according to the filtered signal. This can be achieved by analyzing the characteristics of the signal or using other algorithms.

[0085] Step 3534. Update the filter: Update the filter values according to the calculated filter parameters. This update process can be iterative step by step until the filter converges or reaches a preset threshold.

[0086] Step 3535. Repeat steps 3532 - 3534 until the filter converges.

[0087] Through the above steps, the one - dimensional time - signal enhancement algorithm of adaptive filtering can automatically adjust the filter parameters to better adapt to the characteristics of different signals, thus obtaining better enhancement effects. This algorithm has the advantages of being simple to implement and having a small computational load, and can be widely applied to the processing and enhancement of various one - dimensional time signals.

[0088] At the same time, this algorithm can be further optimized and improved as needed. For example, more complex filter types and parameter update strategies can be introduced to improve the adaptability and enhancement effect of the algorithm. In addition, other signal - processing techniques, such as noise suppression and feature extraction, can be combined to further improve the processing ability of one - dimensional time signals.

[0089] In step 3533, the calculation of filter parameters can adopt the following method:

[0090] A method for calculating filter parameters based on one - dimensional data of lightning arresters is determined according to the characteristics of the signal and the purpose of enhancement. For a linear filter, the filter coefficients can be calculated using the following formula: \(h_k=\frac{s_k}{s_m}\), where: \(h_k\) is the filter coefficient; \(s_k\) is the filtered signal; \(s_m\) is the mean value of the signal. For a non - linear filter, the filter parameters can be calculated using the following formula: \(h_k = f(s_k)\), where: \(f\) is the non - linear function of the filter; \(s_k\) is the filtered signal.

[0091] This method for calculating filter parameters based on one - dimensional data of lightning arresters can select appropriate formulas and parameters according to specific situations to obtain better enhancement effects. At the same time, this method also has the advantages of being simple to implement and having a small computational load, and can be widely applied to the processing and enhancement of one - dimensional data of lightning arresters.

[0092] After the data pre - processing is completed in the fault intelligent detection method of the present invention, various sensor signals will be stacked into a joint feature matrix, and then step 4 is executed and sent into a deep - learning fault diagnosis algorithm based on an improved channel convolutional residual network. The specific implementation steps of this algorithm are as follows:

[0093] Step 411. Model input:

[0094] *x: Input joint feature, with the shape of (N, H, W, C).

[0095] *w: Convolution kernel, with shape (K, K, C, C).

[0096] *k_g: Weight of gated attention, with shape (C, C).

[0097] *k_s: Weight of self attention, with shape (C, C).

[0098] Step 412. Model output:

[0099] *y: Model output probability matrix.

[0100] Step 413. Model calculation process:

[0101] *y = x * w * k_g * k_s.

[0102] *Among them, k_g is the weight of gated attention and can be calculated in the following way:

[0103] *k_g = sigmoid(f(x)).

[0104] *Among them, f(x) is the activation function of gated attention, and the GELU function can be selected.

[0105] *k_s is the weight of self attention and can be calculated in the following way:

[0106] *k_s = softmax(f(x)).

[0107] *Among them, f(x) is the activation function of self attention, and the GELU function can be used.

[0108] Step 414. The Gated Attention Mechanism usually includes two main parts: the gating mechanism and the attention mechanism.

[0109] *The gating mechanism usually uses a gate to determine which information should be updated or ignored. The formula of the gating mechanism can be expressed as: g = sigmoid(W_g * h_{t - 1}+b_g).

[0110] *Among them, g is a gating signal, W_g and b_g are the weights and biases of the gate, and h_{t - 1} is the hidden state of the previous step.

[0111] *The attention mechanism usually uses a weight distribution to determine which information should be focused on or emphasized. The formula of the attention mechanism can be expressed as: a = softmax(W_a * h_{t - 1}+

[0112] b_a).

[0113] *Among them, a is a weight distribution, W_a and b_a are the weights and biases of the attention weights, and h_{t - 1} is the hidden state of the previous step.

[0114] *The gated attention mechanism combines the gating mechanism and the attention mechanism to determine which information should be updated and focused on. The formula of the gated attention mechanism can be expressed as: c = g * a * h_{t - 1}

[0115] *Among them, c is a new context vector used to update the hidden state, g is the gating signal, a is the weight distribution, and h_{t - 1} is the hidden state of the previous step. This formula means that only when the gating signal g is large, the new context vector c is used to update the hidden state.

[0116] Step 415. The working process and main formula of Self - Attention are as follows:

[0117] *The input sequence X ∈ R^{n×d} is represented as a query matrix Q, a key matrix K, and a value matrix V: Q = XW_Q, K = XW_K, V = XW_V. Among them, W_Q, W_K, and W_V are all learnable parameter matrices.

[0118] *Calculate the attention scores, that is, the dot product of the query and the key: Attention(Q, K, V) =

[0119] softmax(QK^T / √d). Here d is the dimension of Q, and it is scaled to avoid being too large.

[0120] *Weight the value matrix: Output = Attention(Q, K, V), that is, use the attention scores to perform weighted summation on the value matrix to obtain the output.

[0121] *Perform multi - head calculation on the above formula, calculate different Q, K, and V respectively, and then concatenate the outputs of the heads: MultiHead(Q, K, V) = Concat(head_1,..., head_h). Among them, head_i = Attention(Q_i, K_i, V_i).

[0122] *Finally, perform post - processing such as linear transformation and residual connection.

[0123] The specific description of using the genetic algorithm for optimization in step 4 is as follows:

[0124] The method of optimizing a neural network based on particle swarm and simulated annealing, as Figure 2 shown, the specific implementation steps are as follows:

[0125] Step 421. Initialize the particle swarm:

[0126] * Create a particle swarm consisting of a group of particles, where each particle represents the weights and structure of a neural network.

[0127] * Initialize the velocity and position of the particles. The velocity can be randomly initialized or initialized based on the genetic algorithm.

[0128] Step 422. Calculate the fitness of the particles:

[0129] * Calculate the fitness of each particle. The fitness measures the quality of the particle. For a neural network, the sum of squared errors can be used as the fitness of the particle.

[0130] Step 423. Update the velocity and position of the particles:

[0131] * In each iteration, each particle updates its position according to its velocity and position.

[0132] * The velocity update formula for the particle is: v(t + 1) = wv(t) + c1r1(gbest - x(t)) +

[0133] c2r2(pbest - x(t)), where v(t) is the velocity of the particle in the t-th iteration, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, gbest is the global optimal solution, and pbest is the best solution of the particle.

[0134] * The position update formula for the particle is: x(t + 1) = x(t) + v(t + 1).

[0135] Step 424. Determine whether convergence has occurred:

[0136] * If all the particles in the particle swarm converge to the global optimal solution, the algorithm ends.

[0137] * Otherwise, enter the next iteration.

[0138] Step 425. Use the simulated annealing algorithm for optimization:

[0139] * After the particle swarm algorithm converges, use the simulated annealing algorithm to further optimize the neural network.

[0140] *In each iteration, a new solution is generated. If the fitness of the new solution is better, the new solution is accepted; if the fitness of the new solution is worse than the original solution, the new solution is accepted with a certain probability.

[0141] *As the number of iterations increases, the annealing temperature of the simulated annealing algorithm gradually decreases. When the annealing temperature drops to a certain value, the algorithm ends.

[0142] Step 426. Repeat steps 423 to 425 until a preset number of iterations is reached or other termination conditions are met.

[0143] Through the above steps, this method combines the advantages of the particle swarm algorithm and the simulated annealing algorithm, and can effectively optimize the weights and structure of the neural network, improving the performance and generalization ability of the neural network.

[0144] In step 5 mentioned above, model online learning is a continuous learning strategy that can achieve continuous optimization and update of the trained model. The following are the specific steps of model online learning:

[0145] Step 51. Initialize the model: At the beginning of online learning, the parameters of the model need to be initialized. Pre-trained model parameters can be used, or the parameters can be randomly initialized.

[0146] Step 52. Obtain new samples: Obtain new input samples from the data stream. These samples can be obtained sequentially or in batches.

[0147] Step 53. Model prediction: Use the current model parameters to predict or infer the new samples to obtain the output results.

[0148] Step 54. Calculate the loss: Compare the model output with the true labels and calculate the error loss. This loss function can help guide the direction of model update.

[0149] Step 55. Update the parameters: Update the parameters of the model through gradient descent of the loss function. This step is carried out according to the backpropagation algorithm.

[0150] Step 56. Model evaluation: Evaluate the performance of the model on the evaluation set at regular intervals. This can help understand the performance of the model on unknown data, as well as whether the model is overfitting or underfitting.

[0151] Step 57. Model saving: Save the current best model according to the evaluation metrics. This can ensure that we have a backup of the best-performing model.

[0152] Step 58. Return to step 52: Continuously obtain new samples in a loop, continuously update the model, and achieve online learning. This step is a loop process that continuously obtains new samples and updates the model to achieve continuous optimization of the model.

[0153] Through the above steps, we can achieve the online learning of the model. This method is particularly suitable for situations where the model needs to be continuously updated to adapt to data changes.

[0154] The present invention also proposes an intelligent arrester fault detection system, including: a sensor network arrangement module for arranging a sensor network around the target arrester device; a data preprocessing module for preprocessing and cleaning the collected data; a noise reduction processing module using an improved one-dimensional signal noise reduction processing algorithm of OMLSA to perform noise reduction on multi-sensor data; a model training module for sending the noise-reduced data into a residual network based on channel convolution for model training, and using a genetic algorithm to search for the optimal parameters of the model; a model update module for realizing the real-time update and optimization of the model by combining online learning and offline learning algorithms; an error determination module for performing error determination after optimizing the model through a particle swarm annealing fusion algorithm; an online prediction module for performing online prediction within an acceptable error range to monitor the operating state of the arrester in real time and discover potential fault risks; and a process control module for ending the entire process when the input changes greatly or other set conditions are met.

[0155] The intelligent arrester fault detection method based on neural network combined with particle swarm optimization algorithm of the present invention has the following remarkable beneficial effects compared with the traditional threshold determination method:

[0156] 1. Improve efficiency: The traditional threshold determination method needs to collect a large amount of data and analyze it to determine the fault type. However, the method based on neural network combined with particle swarm optimization algorithm only needs to collect a small amount of data and learn the fault characteristics through the neural network, thus greatly improving the efficiency of fault diagnosis.

[0157] 2. Improve accuracy: The accuracy of the threshold determination method depends on the setting of the threshold. If the setting is unreasonable, it may lead to errors in fault diagnosis. However, the method based on neural network combined with particle swarm optimization algorithm can find the optimal parameters of the neural network through the particle swarm optimization algorithm, thus improving the accuracy of fault diagnosis.

[0158] 3. Accurately grasp the health status: The traditional threshold determination method can only judge whether the device has a fault and cannot accurately grasp the health status of the device. However, the method based on neural network combined with particle swarm optimization algorithm can learn the fault characteristics through the neural network, thus accurately grasping the health status of the device.

[0159] 4. Enhanced adaptability: This method adds a third-stage strategy for online learning. When significant changes are found in the data distribution of the lightning arrester during the fault inference process, the inference end of the lightning arrester will collect data in real time, send it into the model for training, and then re-implement the model's learning and inference of new data. This can ensure that the model always adapts to data changes and provides more accurate fault diagnosis results.

[0160] In summary, the intelligent fault detection method for lightning arresters based on neural networks combined with the particle swarm optimization algorithm has significant beneficial effects in terms of improving efficiency, accuracy, accurately grasping the health status, and enhancing adaptability compared to the traditional threshold determination method.

[0161] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent detection method for arrester faults, characterized in that, It includes the following steps: Step 1: Arrange a sensor network around the target lightning arrester device to comprehensively monitor the working environment and operating parameters of the device; Step 2: Use the data collected by the sensor network to perform a series of preprocessing operations, including data cleaning and data matching, to remove invalid or incorrect data and ensure the accuracy and consistency of the data; Step 3: Use the one-dimensional signal denoising processing algorithm of the improved OMLSA to denoise the multi-sensor data to remove the noise in the signal and retain useful information; The one-dimensional signal denoising processing algorithm of the improved OMLSA includes the following execution steps: First, collect the target signal, including voice signals or one-dimensional time signals; then use the IMCRA algorithm based on maximum likelihood estimation to estimate the noise and effectively process non-stationary noise; then perform denoising processing on the signal according to the minimum mean square error criterion; thereafter, use the Bayesian estimation method to optimize the quality of the denoised signal; Furthermore, by adding a time-frequency mask to the optimization objective function to perform weighted processing on different time and frequency components, the recognition degree of the denoised signal is enhanced; at the same time, a prior probability model of the signal is added to the objective function to regularize the optimization result to improve the stability of the denoised signal; furthermore, model is built at different frequency resolutions to process signal components of different scales, and the multi-scale results are combined to further improve the denoising performance; finally, a feedback structure is introduced to achieve multi-stage denoising optimization; Step 4: Send the denoised data into a residual network based on channel convolution for model training, so that the model can learn the relationship between the operating state of the lightning arrester and the environmental parameters through training, and use the genetic algorithm to search for the optimal parameters of the model; Step 5: Combine online learning and offline learning algorithms to achieve real-time update and optimization of the model; Step 6: After optimizing the model through the particle swarm annealing fusion algorithm, perform error determination. When the error is within an acceptable range, use the model for online prediction to monitor the operating state of the lightning arrester in real time and discover potential fault risks; Step 7: When there are large changes in the input or other set conditions are met, end the entire process.

2. The intelligent fault detection method for lightning arresters according to claim 1, characterized in that, The enhancement of the recognition degree of the denoised signal is achieved through a one-dimensional time signal enhancement algorithm based on adaptive gain, including the following steps: First, initialize the gain parameters, select a suitable gain type and its parameter values; then use a fixed filter to filter the input one-dimensional time signal to reduce noise and other interferences; then calculate the gain parameters according to the filtered signal; then update the calculated gain parameters to new values; finally, repeat the filtering, calculation and update steps until the gain converges or reaches the preset number of iterations.

3. The intelligent fault detection method for arrester according to claim 1, wherein The recognition degree of the enhanced noise-reduced signal is achieved through a gain calculation method: in the case of linear gain, the formula g = s_n / s_m is used to calculate the gain, where g represents the gain, s_n represents the filtered signal, and s_m represents the mean value of the signal; in the case of non-linear gain, the formula g = f(s_n) is used to calculate the gain, where f represents the non-linear function of the gain and s_n represents the filtered signal.

4. The intelligent fault detection method for lightning arresters according to claim 1, wherein, The recognition degree of the enhanced noise-reduced signal is achieved through a one-dimensional time signal enhancement algorithm with adaptive filtering. The execution steps of this algorithm include: initializing the filter, selecting a suitable filter type and assigning initial parameters; applying the initialized filter to the input one-dimensional time signal for filtering; calculating the parameters of the filter according to the filtered signal; updating the value of the filter according to the calculated filter parameters; repeating the filtering, calculation, and update steps until the filter converges, so as to dynamically adjust the filter parameters according to the signal characteristics and obtain a better one-dimensional time signal enhancement effect.

5. The intelligent fault detection method for lightning arresters according to claim 4, characterized in that, The calculation of the filter parameters is achieved through a filter parameter calculation method based on the one-dimensional data of the lightning arrester: for a linear filter, the formula h_k = s_k / s_m is used to calculate the filter coefficient, where h_k represents the filter coefficient, s_k represents the filtered signal, and s_m represents the mean value of the signal; for a non-linear filter, the formula h_k = f(s_k) is used to calculate the filter parameters, where f represents the non-linear function of the filter and s_k represents the filtered signal.

6. The intelligent detection method for arrester faults according to claim 1, wherein, In step 4, the data after noise reduction processing is sent into a residual network based on channel convolution for model training, including model input, model output, and model calculation process. The model input includes the input joint feature x, convolution kernel w, gated attention weight k_g, and self-attention weight k_s. The model output is the model output probability matrix y; in the model calculation process, through the calculation of the gated attention mechanism and the self-attention mechanism, a new context vector and output are obtained; the gated attention mechanism includes a gating mechanism and an attention mechanism, and the self-attention mechanism includes the calculation of the query matrix, key matrix, value matrix, and attention score, as well as multi-head calculation and post-processing.

7. The intelligent fault detection method for arrester according to claim 1, characterized in that, In step 4, a genetic algorithm is used for optimization. The specific implementation steps are as follows: initialize a group of particle swarms representing the weights and structures of the neural network, and initialize the velocities and positions of the particles; calculate the fitness of each particle to measure the quality of the particle; in each iteration, update the velocities and positions of the particles until all particles in the particle swarm converge to the global optimal solution; after the particle swarm algorithm converges, use the simulated annealing algorithm to further optimize the neural network, generate new solutions and accept or reject them, and gradually reduce the annealing temperature as the number of iterations increases; repeat the particle swarm algorithm and the simulated annealing algorithm until the preset number of iterations is reached or other termination conditions are met.

8. The intelligent fault detection method for lightning arresters according to claim 1, characterized in that, In step 5, an online learning strategy of the model is adopted for continuous optimization and update. The specific implementation steps are as follows: initialize the model parameters; obtain new input samples from the data stream; use the current model parameters for prediction or inference; calculate the error loss between the model output and the true label; update the model parameters through gradient descent of the loss function; evaluate the performance of the model on the evaluation set at regular intervals; save the current best model according to the evaluation metrics; and continuously obtain new samples and update the model to achieve continuous optimization.

9. An intelligent detection system for arrester faults, characterized in that, It includes: a sensor network arrangement module for arranging a sensor network around the target lightning arrester device; a data preprocessing module for preprocessing and cleaning the collected data; a noise reduction processing module for performing noise reduction on multi-sensor data using an improved one-dimensional signal noise reduction processing algorithm of OMLSA; a model training module for sending the noise-reduced data into a residual network based on channel convolution for model training and using a genetic algorithm to search for the optimal parameters of the model; a model update module for realizing real-time update and optimization of the model by combining online learning and offline learning algorithms; an error determination module for performing error determination after optimizing the model through a particle swarm annealing fusion algorithm; an online prediction module for performing online prediction within an acceptable error range to monitor the operating state of the lightning arrester in real time and discover potential fault risks; and a process control module for ending the whole process when the input changes greatly or other set conditions are met; The improved one-dimensional signal noise reduction processing algorithm of OMLSA includes the following execution steps: first, collect the target signal, including a voice signal or a one-dimensional time signal; then use the IMCRA algorithm based on maximum likelihood estimation to estimate the noise and effectively process non-stationary noise; then perform noise reduction processing on the signal according to the minimum mean square error criterion; thereafter, use the Bayesian estimation method to optimize the quality of the noise-reduced signal; further, perform weighted processing on different time and frequency components by adding a time-frequency mask to the optimization objective function to enhance the recognition of the noise-reduced signal; at the same time, add a prior probability model of the signal to the objective function to regularize the optimization result to improve the stability of the noise-reduced signal; further, model at different frequency resolutions to process signal components of different scales, and combine multi-scale results to further improve the noise reduction performance; finally, introduce a feedback structure to achieve multi-stage noise reduction optimization.

Images