A method and system for intelligent analysis of partial discharge detection signals of a reactor

Abstract

The application discloses a kind of intelligent analysis method and system of electric reactor partial discharge detection signal, by extracting time domain, frequency domain and time-frequency domain multidimensional features, then based on activation function quantization single-feature importance, pearson correlation coefficient quantization power frequency and high-frequency signal correlation, realize the weighted accurate fusion of features, both avoid the information limitation of single feature, also reduce the interference of redundant information, so that the fusion feature can fully reflect the core characteristics of electric reactor partial discharge.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment state monitoring and intelligent diagnosis, and relates to an intelligent analysis method and system for detection signals of partial discharge of a reactor. BACKGROUND

[0002] As a core device in the power system, the reactor is widely used in scenes such as harmonic suppression, surge current limitation, and line reactive power distribution, and the stability of the operation state of the reactor is directly related to the safe and reliable operation of the power system. Partial discharge is an important representation of internal insulation defects of the reactor, and through detection and analysis of the partial discharge signals, potential faults of the device can be found in time, and insulation faults can be prevented from further expanding to cause serious power accidents, so partial discharge detection has become one of the key technologies for state monitoring of the reactor.

[0003] With the development of sensor technology, a double-magnetic-core signal extraction sensor has realized synchronous collection of power frequency signals and high-frequency partial discharge signals of the reactor, providing a data basis for collaborative analysis of multiple source signals. However, the current analysis method for the output signals of the sensor still has many technical bottlenecks, and it is difficult to meet the actual needs of accurate diagnosis:

[0004] In the signal denoising link, the traditional method mostly uses fixed frequency band filtering technology, which cannot adaptively separate effective signals and noise according to the dynamic characteristics of the power frequency signals and the high-frequency partial discharge signals. Especially in a complex operating environment such as a strong electromagnetic field and vibration, noise interference easily leads to a significant decrease in the signal-to-noise ratio, affecting the accuracy of subsequent feature extraction.

[0005] In the feature processing link, the existing technology often separately analyzes the features of the power frequency signals and the high-frequency signals, or uses a simple weighting method for fusion, without fully considering the internal correlation between the two types of signals. This processing method easily causes feature information redundancy or missing of key information, and cannot form fused features that can accurately reflect the essence of partial discharge faults, thereby restricting the reliability of fault diagnosis.

[0006] In the aspect of constructing a diagnosis model, the traditional deep learning network lacks targeted structural design, and it is difficult to efficiently extract local time-frequency features and time sequence features of the signals at the same time. Moreover, the model training mostly relies on a large number of samples of a specific type of reactor, and when facing different types of reactors or small sample scenarios, the model has poor adaptability, and the diagnosis accuracy is significantly reduced. At the same time, the existing model is mostly of a fixed structure, and after training, it cannot be dynamically optimized according to new data accumulated in the long-term operation process of the device. With the extension of the operation time of the device, the diagnosis accuracy of the model will gradually decrease, and it is difficult to meet the needs of long-term monitoring.

[0007] In summary, the existing partial discharge detection signal analysis method of the electric reactor has defects in noise reduction effect, feature fusion quality, model adaptability and long-term stability, and an efficient intelligent analysis method that can adapt to the signal characteristics of the double magnetic core sensor is needed to improve the accuracy and reliability of the partial discharge fault diagnosis of the electric reactor and ensure the safe and stable operation of the power system. SUMMARY

[0008] To solve the problems in the background art, the present application provides an intelligent analysis method and system for partial discharge detection signals of electric reactors, aiming to overcome the shortcomings of existing partial discharge signal analysis methods of electric reactors and provide an intelligent analysis method based on multi-band adaptive noise reduction, attention feature fusion and migration deep learning diagnosis.

[0009] The first aspect of the present application provides an intelligent analysis method for partial discharge detection signals of electric reactors, comprising:

[0010] Receiving the power frequency signal and the high frequency signal output by the double magnetic core sensor, and obtaining a plurality of modal components through variational modal decomposition; after screening the effective modal components, calculating the adaptive noise reduction threshold based on the noise standard deviation and the number of sampling points;

[0011] Extracting the time domain, frequency domain and time-frequency domain features of the power frequency signal and the high frequency signal after noise reduction to form a feature vector, calculating the local and global weights based on the activation function and the Pearson correlation coefficient and summing them up to obtain a fusion feature vector;

[0012] Constructing a deep learning network containing a residual connection convolutional subnetwork and an improved forget gate long short-term memory subnetwork, training a basic model through a large number of samples, fine-tuning a target model based on feature distribution differences, inputting the fusion feature vector to output the fault type and determining the fault severity combined with the high frequency signal parameters;

[0013] Setting the diagnosis error and data accumulation threshold, when the error continuously meets the standard or the sample size meets the standard, fine-tuning the model with new samples and historical samples, updating the target model or storing the samples according to the change of the validation set accuracy.

[0014] Optionally, the adaptive noise reduction threshold calculation formula is , wherein is the adaptive noise reduction threshold of the Kth effective modal component, is the noise standard deviation, N is the number of sampling points, and lg represents the logarithm operation with base 10, represents the square root operation.

[0015] Optionally, the local weight calculation formula includes: , , wherein is the local weight of the i-th feature of the power frequency feature vector, is the local weight of the jth feature of the high-frequency feature vector, S represents a sigmoid activation function, is the ith feature value of the power frequency feature vector, is the jth feature value of the high-frequency feature vector, e represents a natural constant, ∑ represents a summation operation, and u and v are traversal variables in the summation operation.

[0016] Optionally, the global weight calculation formula is as follows: wherein γ is the global weight of the power frequency feature vector, is the power frequency feature vector is the high-frequency feature vector, e represents a natural constant, is the Pearson correlation coefficient between the power frequency feature vector and the high-frequency feature vector; 1-γ is the global weight of the high-frequency feature vector.

[0017] Optionally, the improved forgetting gate calculation formula is as follows: wherein, is the output value of the improved forgetting gate at time t, S represents a sigmoid activation function, is a weight matrix, is the hidden layer output value of the long short-term memory subnetwork at time t-1, is the input feature value of the hidden layer at time t, is a bias term, is a decay coefficient, is the difference between the hidden layer output values at times t-1 and t-2, and , is the hidden layer output value at time t-2, | | represents an absolute value operation, and • represents a matrix multiplication operation. Optionally,

[0018] Optionally, the feature distribution difference is calculated by KL divergence, and the formula is as follows: wherein is the KL divergence value, P(x) is the feature distribution output by the basic model, Q(x) is the feature distribution of the target type reactor sample set, x is a feature variable, lg represents a logarithm operation with a base of 10, and ∑ represents a summation operation, represents the ratio of the corresponding feature variable x between the feature distribution output by the basic model and the feature distribution of the target type reactor sample set.

[0019] Optionally, the fault types include: no fault, insulation aging, air gap discharge, and surface discharge.

[0020] A second aspect of this application provides an intelligent analysis system for partial discharge detection signals of a reactor, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned intelligent analysis method for partial discharge detection signals of a reactor.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] This invention provides an intelligent analysis method and system for partial discharge detection signals of reactors. By extracting multi-dimensional features in the time domain, frequency domain, and time-frequency domain, and then quantifying the importance of single features based on activation functions and the correlation between chemical frequency and high frequency signals based on Pearson correlation, the weighted and accurate fusion of features is achieved. This avoids the information limitations of single features and reduces the interference of redundant information, so that the fused features can fully reflect the core characteristics of partial discharge of reactors.

[0023] This invention constructs a deep learning network containing a residual-connected convolutional subnetwork and an improved forgetting gate long short-term memory subnetwork, which not only enhances the ability to extract local time-frequency features but also improves the sensitivity to time-series abrupt signals. By training the basic model with large-scale samples and fine-tuning the target model based on the differences in feature distribution, it effectively solves the sample adaptation problem of different types of reactors. Especially in small sample scenarios, it can still accurately output the fault type and severity, and significantly reduce the fault misjudgment rate. Attached Figure Description

[0024] Figure 1 This is a flowchart of an intelligent analysis method for partial discharge detection signals of a reactor according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of an intelligent analysis system for partial discharge detection signals of a reactor according to an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] In one embodiment, such as Figure 1 As shown, an intelligent analysis method for partial discharge detection signals of reactors is provided, which is then applied to... Figure 1 Taking China as an example, the following specific steps will be used:

[0028] S10: Receives the power frequency signal and high frequency signal output from the dual magnetic core sensor, performs variational mode decomposition to obtain multiple sets of modal components, filters the effective modal components, and calculates the adaptive noise reduction threshold based on the noise standard deviation and the number of sampling points.

[0029] Specifically, receiving the power frequency signal and high-frequency signal output from the dual-core sensor is the initial step in this process. The power frequency signal, output from the coil corresponding to the signal-taking core of the dual-core sensor, has a frequency of 50Hz and contains information related to the power frequency load current during reactor operation. The high-frequency signal, also output from the coil corresponding to the signal-taking core, has a frequency range of 1MHz to 100MHz and contains high-frequency pulse signals generated by partial discharge in the reactor. After being output by the sensor, both types of signals are transmitted to the analysis system through a preset signal transmission path, serving as the raw input for subsequent processing.

[0030] Variational mode decomposition (VMD) involves decomposing the received signals into multiple modal components. For power frequency signals, VMD decomposes them into three modal components, corresponding to the fundamental frequency and the 2nd and 3rd harmonics. For high-frequency signals, VMD decomposes them into five modal components, covering different frequency components related to partial discharge within the 1MHz to 100MHz frequency band. By constructing and solving a variational constraint problem, VMD ensures that each modal component has an independent center frequency, achieving signal separation across different frequency dimensions and laying the foundation for subsequent screening of effective information.

[0031] Screening effective modal components is the process of extracting relevant information for partial discharge detection from the multiple sets of modal components obtained by decomposition. Specifically, the signal-to-noise ratio (SNR) of each set of modal components is calculated, which is the ratio of the power of the effective signal to the power of the noise in that component. Modal components with an SNR of not less than 15 dB are retained as effective modal components, while those with an SNR of less than 15 dB are discarded, thereby avoiding interference from noise in low SNR components to subsequent analysis.

[0032] Calculating the adaptive noise reduction threshold based on the noise standard deviation and the number of sampling points is a process of noise suppression for the selected effective modal components. First, the noise standard deviation of each effective modal component is calculated. This value is obtained by analyzing the signal of the component during the period without partial discharge, and is used to quantify the noise level in the component. Next, the number of sampling points N for each effective mode component is determined, i.e., the number of discrete sampled data points contained in the component; subsequently, according to the formula... Calculate the adaptive noise reduction threshold , where lg represents the logarithmic operation with base 10. This represents the square root operation. Finally, select the effective modal components whose absolute values ​​are less than [a certain value]. The sampling points are set to 0, and the absolute values ​​are not less than 0. The sampling points are then used to reconstruct all the effective modal components after processing, resulting in the noise-reduced power frequency signal and high frequency signal, thus completing the noise reduction process.

[0033] Taking the partial discharge detection of a 220kV parallel reactor as an example: The power frequency signal output by the dual-core sensor is 50Hz. After variational mode decomposition, three modal components are obtained, corresponding to the 50Hz fundamental wave, the 100Hz second harmonic, and the 150Hz third harmonic, respectively. The high-frequency signal ranges from 1MHz to 100MHz. After variational mode decomposition, five modal components are obtained, corresponding to frequency components around 5MHz, 20MHz, 40MHz, 60MHz, and 80MHz, respectively. After calculating the signal-to-noise ratio (SNR) of each modal component, the three power frequency modal components with SNRs of 18dB, 20dB, and 16dB, and the five high-frequency modal components with SNRs of 17dB, 19dB, 21dB, 16dB, and 18dB are retained as effective modal components. For one set of effective high-frequency modal components, its noise standard deviation is... With a voltage of 2.5V and a sampling point count N of 10000, the adaptive threshold is calculated using the formula. = = ≈7.07V; Set the sampling points with an absolute value less than 7.07V in this component to 0, retain the remaining sampling points, and then reconstruct them with other effective modal components to obtain the noise-reduced high-frequency signal. Similarly, the noise reduction of the power frequency signal is completed.

[0034] S20: Extract the time-domain, frequency-domain, and time-frequency-domain features of the noise-reduced power frequency signal and high frequency signal to form a feature vector. Calculate the local and global weights based on the activation function and Pearson correlation coefficient, and sum them up to obtain the fused feature vector.

[0035] Specifically, extracting the time-domain, frequency-domain, and time-frequency-domain features of the two types of denoised signals to form feature vectors is the process of feature quantization of the denoised power frequency signal and high-frequency signal. The two types of denoised signals refer to the power frequency signal and the high-frequency signal after variational mode decomposition and adaptive denoising threshold processing, respectively.

[0036] For the denoised power frequency signal, the extracted time-domain features include the effective current value and the total harmonic distortion rate (THD). The effective current value is the square root of the average of the squares of the current physical quantity of the power frequency signal over one cycle. Specifically, it is calculated by squaring the instantaneous current value of the power frequency current signal within a complete cycle, obtaining the arithmetic mean of all squares, and then taking the square root of this average. The result is the effective current value of the power frequency signal. The THD is the ratio of the root mean square (RMS) of all harmonic components to the RMS of the fundamental component. Frequency-domain features include the fundamental frequency and the amplitude of the third harmonic. The fundamental frequency is the frequency of the 50Hz component in the power frequency signal, and the amplitude of the third harmonic is the amplitude of the 150Hz component. Time-frequency domain features include the marginal spectral entropy and instantaneous frequency standard deviation extracted based on the Hilbert-Huang transform. The marginal spectral entropy is the information entropy of the Hilbert marginal spectrum, and the instantaneous frequency standard deviation is the quantized value of the discreteness of the instantaneous frequency sequence of the signal. These six features are arranged in order to form the power frequency feature vector.

[0037] For the denoised high-frequency signal, the extracted time-domain features include pulse peak value and pulse interval. The pulse peak value is the maximum amplitude of the partial discharge pulse in the high-frequency signal, and the pulse interval is the time difference between two adjacent partial discharge pulses. The frequency-domain features include center frequency and bandwidth. The center frequency is the center value of the frequency band where the high-frequency signal energy is concentrated, and the bandwidth is the frequency range of the main energy distribution of the high-frequency signal. The time-frequency domain features include marginal spectral entropy and instantaneous frequency standard deviation extracted based on the Hilbert-Huang transform, defined consistent with the time-frequency domain features of the power frequency signal. These six features are arranged in order to form a high-frequency feature vector.

[0038] The process of fusing two types of feature vectors involves calculating local and global weights based on activation functions and Pearson correlation coefficients, and then summing them to obtain a fused feature vector. The activation function is the sigmoid function, which maps feature values ​​to the [0,1] interval to quantify feature importance; the Pearson correlation coefficient is used to quantify the correlation between the chemical frequency feature vector and the high-frequency feature vector.

[0039] When calculating local weights, for the power frequency feature vector, the formula is used. Calculate the local weights for each feature Where S represents the sigmoid activation function, Let be the i-th eigenvalue of the power frequency eigenvector, e be the natural constant, and ∑ be the summation operation; for high-frequency eigenvectors, the formula is used... Calculate the local weights for each feature ,in is the j-th eigenvalue of the high-frequency eigenvector. The meanings of the other parameters are consistent with the local weight formula of the power frequency eigenvector. u and v are the traversal variables in the summation operation, respectively.

[0040] When calculating the global weight, the formula is used. Calculate the global weight γ of the power frequency eigenvector, where γ is the Pearson correlation coefficient between the power frequency feature vector and the high frequency feature vector. The global weight of the high frequency feature vector is 1-γ, and the meanings of other parameters are consistent with the local weight formula.

[0041] In the weighted summation process, the weighted sum of the power frequency eigenvectors is first calculated, that is... Then calculate the weighted sum of the high-frequency eigenvectors, i.e. Where u and v are the traversal variables in the summation operation; finally, the feature vectors are fused... The fused feature vector is obtained.

[0042] Taking the partial discharge detection of a 220kV parallel reactor as an example: the features extracted from the noise-reduced power frequency signal are: effective value 5A, total harmonic distortion rate 3%, fundamental frequency 50Hz, third harmonic amplitude 0.8A, marginal spectral entropy 0.6, and instantaneous frequency standard deviation 2Hz, forming a power frequency feature vector. =[5,3,50,0.8,0.6,2]; The features extracted from the denoised high-frequency signal are: pulse peak value 80V, pulse interval 50ms, center frequency 30MHz, bandwidth 15MHz, marginal spectral entropy 0.7, and instantaneous frequency standard deviation 5MHz, forming a high-frequency feature vector. =[80,50,30,15,0.7,5].

[0043] When calculating the local weights, after mapping with the sigmoid function, the local weights of the power frequency feature are determined as [0.1, 0.1, 0.2, 0.1, 0.3, 0.2], and the local weights of the high-frequency feature are determined as [0.3, 0.2, 0.1, 0.1, 0.2, 0.1]. When calculating the global weights, given that the Pearson correlation coefficient C = 0.7 between the power frequency feature vector and the high-frequency feature vector, the global weight formula is used. First, calculate the exponent term, e. 0.7 ≈2.0138, e 1-0.7 =e 0.3≈1.3499, then calculate the denominator as 2.0138+1.3499≈3.3637, substitute into the formula to get γ≈2.0138 / 3.3637≈0.60, the corresponding high-frequency feature vector global weight 1-γ≈0.40. Before calculating the fused feature vector, the weighted sum of the power frequency features and the weighted sum of the high frequency features are calculated separately. The weighted sum of the power frequency features is 0.1 × 5 (the product of the first eigenvalue 5 and its corresponding local weight 0.1) + 0.1 × 3 (the product of the second eigenvalue 3 and its corresponding local weight 0.1) + 0.2 × 50 (the product of the third eigenvalue 50 and its corresponding local weight 0.2) + 0.1 × 0.8 (the product of the fourth eigenvalue 0.8 and its corresponding local weight 0.1) + 0.3 × 0.6 (the product of the fifth eigenvalue 0.6 and its corresponding local weight 0.3) + 0.2 × 2 (the product of the sixth eigenvalue 2 and its corresponding local weight 0.2). After calculating each item, the sum is 0.5 + 0.3 + 10 + 0.08 + 0.18 + 0.4 = 11.46. The weighted sum of the high frequency features is 0.3 × 80 (the product of the first eigenvalue 50 and its corresponding local weight 0.2). The product of the first eigenvalue (80) and its corresponding local weight (0.3) + 0.2 × 50 (the product of the second eigenvalue (50) and its corresponding local weight (0.2)) + 0.1 × 30 (the product of the third eigenvalue (30) and its corresponding local weight (0.1)) + 0.1 × 15 (the product of the fourth eigenvalue (15) and its corresponding local weight (0.1)) + 0.2 × 0.7 (the product of the fifth eigenvalue (0.7) and its corresponding local weight (0.2)) + 0.1 × 5 (the product of the sixth eigenvalue (5) and its corresponding local weight (0.1)) is calculated item by item and then summed to get 24 + 10 + 3 + 1.5 + 0.14 + 0.5 = 39.14. Finally, the above results are substituted into the formula for calculating the fused eigenvector, i.e., fused eigenvector = 0.60 × 11.46 + 0.40 × 39.14, and the result is 6.876 + 15.656 ≈ 22.53.

[0044] S30: Construct a deep learning network with a convolutional subnetwork containing residual connections and an improved long short-term memory subnetwork with a forget gate. The basic model is trained with a large number of samples, and the target model is fine-tuned based on the differences in feature distribution. The input is a fused feature vector, the output is the fault type, and the fault severity is determined by combining high-frequency signal parameters.

[0045] Specifically, constructing a deep learning network containing a residual-connected convolutional subnetwork and an improved forgetting gate long short-term memory subnetwork is the core model structure design process for realizing intelligent diagnosis of partial discharge signals. The residual-connected convolutional subnetwork consists of three convolutional layers, each with a 3×3 kernel size and a stride of 1. Residual connections are established between adjacent convolutional layers, meaning that the output of the previous layer is directly superimposed onto the output of the next layer through skip connections. This alleviates the gradient vanishing problem in deep network training and enhances the extraction capability of local time-frequency features.

[0046] The improved forgetting gate long short-term memory subnetwork contains two hidden layers, each with 64 neurons. The formula for calculating the improved forgetting gate is as follows: ,in To improve the output value of the forget gate at time t, This represents the sigmoid activation function. Its weight matrix, This represents the hidden layer output value of the Long Short-Term Memory (LSTM) subnetwork at time t-1. Let be the input feature value at time t. For its bias term, The attenuation coefficient is... The difference between the hidden layer output values ​​at time t-1 and time t-2 is... , t-2 represents the hidden layer output value, | represents absolute value operation, and • represents matrix multiplication operation. This improvement can enhance the network's sensitivity to time-varying signals such as partial discharge pulses in high-frequency signals.

[0047] The two sub-networks are connected in sequence to form a complete deep learning network. The convolutional sub-network outputs 32-dimensional local time-frequency features, which serve as the input to the long short-term memory sub-network. Finally, the long short-term memory sub-network outputs 64-dimensional temporal features for processing by the subsequent diagnostic layer.

[0048] The basic model, trained on a large-scale sample set, is the process of enabling the deep learning network to possess general diagnostic capabilities. The large-scale sample set contains 10,000 data points, covering partial discharge signals from different types of reactors, including 110kV, 220kV, and 500kV. Each sample set includes corresponding power frequency feature vectors, high-frequency feature vectors, fused feature vectors, and fault type labels. During training, the fused feature vectors are expanded into a 32×32 time-frequency matrix as input, with the fault type as the label. The backpropagation algorithm is used to iteratively optimize the network parameters until the accuracy on the validation set stabilizes. After training, the first 80% of the weights of the convolutional and long short-term memory sub-networks are frozen, retaining only the weights of the fully connected layer (containing 32 neurons) and the last 20% of the weights of the long short-term memory sub-network as fine-tunable parameters, forming the basic model. This basic model can extract common features from the partial discharge signals of different types of reactors.

[0049] The process of fine-tuning the basic model based on the differences in feature distribution to obtain the target model is the process of adapting the basic model to a specific type of reactor.

[0050] The difference in characteristic distributions is calculated using the KL divergence formula. , Let P(x) be the KL divergence value, P(x) be the feature distribution output by the basic model, Q(x) be the feature distribution of the target model reactor sample set, x be the feature variable, lg represent the logarithm operation to base 10, and ∑ represent the summation operation. This represents the ratio of the feature distribution output by the basic model to the feature distribution corresponding to the feature variable x in the target model reactor sample set.

[0051] In both the feature distribution P(x) and the feature distribution Q(x), x refers to the output vector of the last feature layer before the fully connected (classification) layer in the deep learning network. Specifically, the fused feature vector of a large-scale sample set is input into a pre-trained base model, and the output of that layer is extracted. Gaussian distribution modeling is then applied to these output vectors, ensuring that each dimension conforms to a Gaussian distribution. Estimating the parameters yields the feature distribution P(x) of the source domain. Similarly, the fused feature vector of a target type reactor sample set is input into the same base model, the output of the same feature layer is extracted, and statistical modeling is performed to obtain the feature distribution Q(x) of the target domain. The Gaussian distribution modeling method, as a fundamental parameter estimation method in probability theory and statistics, will not be elaborated upon here.

[0052] For a small sample set of 100-500 samples of the target reactor type, such as 220kV, calculate its KL divergence value with the basic model: If <0.2 indicates that the feature distribution output by the basic model is similar to the feature distribution of the target reactor sample set, requiring only fine-tuning of the fully connected layer parameters; if A value ≥0.2 indicates a large difference, requiring the unfreezing of 20% of the weights in the Long Short-Term Memory subnetwork, and adjusting the weights to 10. -5 The learning rate is fine-tuned for 5-10 rounds to finally obtain the target model that is adapted to the target model.

[0053] The diagnostic output process of the target model involves inputting a fused feature vector, outputting the fault type, and combining it with high-frequency signal parameters to determine the fault severity. After inputting the fused feature vector into the target model, local time-frequency features are extracted by a convolutional sub-network, and temporal features are captured by a long short-term memory sub-network. Finally, a fully connected layer outputs the probability distribution of four types of faults through a softmax activation function. The category with the highest probability is the fault type, including no fault, insulation aging, air gap discharge, and surface discharge.

[0054] The severity of the fault is classified into five levels, from 1 to 5, which need to be determined by combining the average pulse amplitude and pulse frequency of the high-frequency signal. Level 1: Average pulse amplitude D ≤ 10V and pulse frequency E ≤ 10 times / minute; Level 2: 10V < average pulse amplitude D ≤ 50V and 10 times / minute < pulse frequency E ≤ 50 times / minute; Level 3: 50V < average pulse amplitude D ≤ 100V and 50 times / minute < pulse frequency E ≤ 100 times / minute; Level 4: 100V < average pulse amplitude D ≤ 200V and 100 times / minute < pulse frequency E ≤ 200 times / minute; Level 5: Average pulse amplitude D > 200V or pulse frequency E > 200 times / minute.

[0055] Taking the partial discharge detection of a 220kV parallel reactor as an example: In the constructed deep learning network, the three convolutional layers of the residual connection convolutional subnetwork output 64, 128, and 32-dimensional features respectively, and the two hidden layers of the long short-term memory subnetwork each contain 64 neurons. The decay coefficient λ of the improved forgetting gate is set to 0.2. A large-scale sample set includes 3000 sets of 110kV samples, 4000 sets of 220kV samples, and 3000 sets of 500kV samples. After training, the validation set accuracy of the basic model reaches 94%. For 300 sets of samples from the target 220kV reactor, the KL divergence value is calculated. =0.15 < 0.2, the target model is obtained after fine-tuning the fully connected layer for 5 rounds; after inputting the fused feature vector, such as 23.8, the target model outputs the fault type as air gap discharge with a probability of 0.92; combined with the high-frequency signal parameters, the average pulse amplitude is 70V and the pulse frequency is 80 times / minute, the fault severity is determined to be level 3.

[0056] S40: Set diagnostic error and data accumulation thresholds. When the error continuously meets the target or the sample size meets the target, fine-tune the model with new samples and historical samples, and update the target model or store samples based on the change in the accuracy of the validation set.

[0057] Specifically, setting diagnostic error and data accumulation thresholds is the process of establishing triggering conditions for the iteration of the target model. Diagnostic error refers to the proportion of difference between the fault type output by the target model and the manually labeled results, such as results verified by infrared detection or partial discharge detectors. The diagnostic error threshold is set to 10%, meaning that when a single diagnostic error exceeds 10%, it is considered a single error meeting the standard; continuous meeting the standard means that the diagnostic error exceeds 10% for 10 consecutive times. The data accumulation threshold is set to 200 sets. Here, the samples are valid samples, i.e., a dataset containing complete power frequency signals, high frequency signals, fault type labels, and fault severity labels. When the number of newly collected valid samples reaches 200 sets, the sample size is considered to have met the standard. The two thresholds are independent of each other; when either threshold is met, the model fine-tuning process is triggered.

[0058] When the error rate consistently meets the target or the sample size meets the target, fine-tuning the model using new and historical samples is the core operation of the target model iteration. New samples refer to the newly collected valid samples when the iteration is triggered, while historical samples refer to valid samples previously stored in the device database that have been used for target model training or have not participated in training. After merging the new and historical samples, they are randomly divided into training and validation sets in a 7:3 ratio. The training set is used to fine-tune the target model parameters, and the validation set is used to evaluate the fine-tuning effect. The fine-tuning process uses the Adam optimizer, with a learning rate set to... Using the target model as the initial model, iterate for 3-5 rounds, and calculate the diagnostic accuracy of the validation set after each round of training.

[0059] Updating the target model or storing samples based on changes in validation set accuracy is the result processing procedure for target model iteration. Changes in validation set accuracy refer to the difference between the accuracy of the fine-tuned model on the validation set and the accuracy of the target model before fine-tuning. If this difference is not less than 3%, meaning the improvement in validation set accuracy meets the preset requirement, the fine-tuning is effective, and the target model is updated to the fine-tuned new model, replacing the original target model for subsequent diagnosis. If the difference is less than 3%, meaning the improvement in validation set accuracy does not meet the preset requirement, the fine-tuning has not significantly improved the target model's performance, the original target model is retained, and the new samples are stored in the device database as a supplement to historical samples, to be used in model training during the next iteration.

[0060] Taking the partial discharge detection of a 220kV parallel reactor as an example: the diagnostic error threshold is set to 10%, the consecutive compliance count is 10, and the data accumulation threshold is 200 sets. During equipment operation, the error between the fault type output by the target model and the manually labeled result is 12% for 10 consecutive times, meeting the error compliance condition and triggering fine-tuning of the target model. At this time, there are 200 newly collected effective samples and 800 historical samples, totaling 1000 samples after merging. These are divided into a 7:3 ratio: 700 training sets and 300 validation sets. Using the original target model as the initial model, the Adam optimizer is used for 5 rounds of fine-tuning. After fine-tuning, the validation set accuracy is 96%, an improvement of 4% compared to the 92% before fine-tuning. Therefore, the target model is updated to the fine-tuned new model. In another scenario, when the number of newly collected valid samples reaches 200, which meets the data accumulation threshold, the historical samples are merged and divided into training and validation sets. After three rounds of fine-tuning, the accuracy of the validation set is 93%, which is 1% higher than the 92% before fine-tuning. In this case, the original target model is retained, and only the 200 new samples are stored in the database.

[0061] In one embodiment, such as Figure 2 As shown, an intelligent analysis system for partial discharge detection signals of a reactor is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements an intelligent analysis method for partial discharge detection signals of a reactor.

[0062] Specific limitations regarding the intelligent analysis system for reactor partial discharge detection signals can be found in the limitations of the intelligent analysis method for reactor partial discharge detection signals described above, and will not be repeated here. Each module in the aforementioned intelligent analysis system for reactor partial discharge detection signals can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0063] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0064] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent analysis of partial discharge detection signals of a reactor, characterized in that, The method comprises the following steps: Receiving power frequency signals and high frequency signals output by a double magnetic core sensor, and obtaining a plurality of groups of modal components through variational modal decomposition; after screening effective modal components, calculating an adaptive noise reduction threshold based on noise standard deviation and the number of sampling points; Extracting time domain, frequency domain and time-frequency domain features of the power frequency signals and the high frequency signals after noise reduction to form a feature vector, calculating local weights and global weights based on an activation function and a Pearson correlation coefficient, and performing weighted summation to obtain a fusion feature vector; Constructing a deep learning network comprising a residual connection convolutional subnetwork and an improved forget gate long short-term memory subnetwork, training a basic model through a large number of samples, fine-tuning a target model based on feature distribution differences, inputting the fusion feature vector to output a fault type, and determining a fault severity in combination with high frequency signal parameters; Setting a diagnosis error and a data accumulation threshold, fine-tuning the model with new samples and historical samples when the error continuously meets the standard or the sample size meets the standard, and updating the target model or storing the samples according to the accuracy rate change of the verification set; wherein the calculation formula of the improved forget gate is wherein, is the output value of the improved forget gate at time t, represents a sigmoid activation function, is a weight matrix, is the hidden layer output value of the long short-term memory subnetwork at time t-1, is the input feature value of the hidden layer at time t, is a bias term, is a decay coefficient, is the difference between the hidden layer output values at times t-1 and t-2, and , is the hidden layer output value at time t-2, and || represents an absolute value operation, represents a matrix multiplication operation; The feature distribution difference is calculated by KL divergence, and the formula is wherein is the KL divergence value, P(x) is the feature distribution output by the basic model, Q(x) is the feature distribution of the target type reactor sample set, x is the feature variable, lg represents the logarithmic operation with 10 as the base, and ∑ represents the summation operation, represents the ratio of the corresponding feature variable x of the feature distribution output by the basic model and the feature distribution of the target type reactor sample set.

2. The method according to claim 1, wherein, The adaptive noise reduction threshold calculation formula is wherein, is the adaptive noise reduction threshold of the Kth effective modal component, is the noise standard deviation, N is the number of sampling points, and lg represents the logarithm operation with base 10, represents the square root operation.

3. The method according to claim 1, wherein, The local weight calculation mode comprises: , , wherein, is a local weight of an i-th feature of a power frequency feature vector, is a local weight of a j-th feature of a high frequency feature vector, S represents a sigmoid activation function, is an i-th feature value of the power frequency feature vector, is a j-th feature value of the high frequency feature vector, e represents a natural constant, and ∑ represents a summation operation.

4. The method of claim 1, wherein the method further comprises: The global weight calculation mode comprises: Wherein, γ is the global weight of the power frequency feature vector, is the power frequency feature vector is the high frequency feature vector, e represents a natural constant, is the Pearson correlation coefficient between the power frequency feature vector and the high frequency feature vector; 1-γ is the global weight of the high frequency feature vector.

5. The method of claim 1, wherein the method further comprises: The fault types include: no fault, insulation aging, air gap discharge, and surface discharge.

6. An intelligent analysis system for a reactor partial discharge detection signal, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The processor executes the computer program to implement the intelligent analysis method of the partial discharge detection signal of the electric reactor according to any one of claims 1 to 5.

Images