Power distribution terminal automatic test method and system based on dynamic fault recording and analysis oscillograph

Abstract

The invention discloses a power distribution terminal automatic test method and system based on a dynamic fault record and analysis oscillograph, and the method comprises the steps: S1, obtaining a signal collected by the dynamic fault record and analysis oscillograph, carrying out the wavelet packet decomposition and noise reduction processing, carrying out the transformation, obtaining a signal feature containing an instantaneous frequency and an instantaneous amplitude, and carrying out the detection of the instantaneous frequency and the instantaneous amplitude; time domain features and frequency domain features in the signal features are extracted, and fault feature classification is carried out; s2, reconstructing the fault feature vector, carrying out time sequence feature association mapping, and predicting the action response of the power distribution terminal through a pre-trained intelligent prediction model; generating a test case according to the prediction result and the corresponding fault feature vector; and S3, reconstructing a test signal, outputting the test signal to the power distribution terminal to generate an actual response signal, evaluating the actual response signal by using a prediction result of the intelligent prediction model, and generating a test evaluation result responded by the power distribution terminal. The intelligent level of power distribution terminal testing can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment testing, in particular to a power distribution terminal automation testing method and system based on a dynamic fault record and analysis recording wave device. BACKGROUND

[0002] With the continuous development and intelligent upgrading of the power system, the automation testing technology of the power distribution terminal has increasingly become the focus of the power industry.

[0003] At present, the traditional power distribution terminal testing method lacks advanced signal processing technology in the signal acquisition and noise reduction processing link, and it is difficult to accurately extract the small change characteristics of the electrical signal, resulting in low accuracy and real-time of fault identification. At the same time, there are significant shortcomings in fault prediction and pattern learning, mainly relying on experience model and static threshold judgment, lacking in-depth learning and intelligent analysis ability of the dynamic response characteristics of the equipment, and it is difficult to accurately predict the potential fault and abnormal working state of the equipment.

[0004] Therefore, there is an urgent need for a solution to solve the problems in the prior art. SUMMARY

[0005] The main purpose of the present application is to provide a power distribution terminal automation testing method and system based on a dynamic fault record and analysis recording wave device, which can automatically test the correctness of the response signal output by the power distribution terminal.

[0006] The technical scheme adopted by the present application is: A power distribution terminal automation testing method based on a dynamic fault record and analysis recording wave device is provided, comprising the following steps: S1, acquire the alternating voltage signal and alternating current signal collected by the dynamic fault record and analysis recording wave device, perform wavelet packet decomposition and noise reduction processing to obtain the noise-reduced electrical signal; transform the noise-reduced electrical signal to obtain signal characteristics containing instantaneous frequency and instantaneous amplitude, and then extract the time domain characteristics and frequency domain characteristics in the signal characteristics; input the extracted time domain characteristics and frequency domain characteristics into a pre-constructed deep convolutional neural network for fault feature classification to obtain a fault feature vector; S2, reconstruct the fault feature vector and perform channel time sequence feature correlation mapping, input the correlation mapping result into a pre-trained intelligent prediction model, predict the action response of the power distribution terminal, and generate a prediction result; generate a test case according to the prediction result and the corresponding fault feature vector; S3, reconstruct the fault feature vector in the test case into a test signal and output it to the power distribution terminal, the power distribution terminal generates an actual response signal to the input test signal, evaluates the actual response signal using the prediction result of the intelligent prediction model, and generates a test evaluation result of the response of the power distribution terminal.

[0007] According to the above technical solution, in step S1, the electrical signal after noise reduction is specifically subjected to Hilbert-Huang transformation to obtain signal characteristics containing instantaneous frequency and instantaneous amplitude.

[0008] According to the above technical solution, the fault feature vector contains device operating state characteristics, electrical quantity characteristics and environmental parameter characteristics.

[0009] According to the above technical solution, in step S2, the fault feature vector is specifically encoded into an encoded vector in a latent feature space, the distribution of the encoded vector is constrained by a discriminator, a reconstructed feature vector is generated based on the encoded vector, the reconstructed feature vector and the original fault feature vector are subjected to adversarial training, network parameters are updated until the reconstruction error converges, and fault mode data is generated based on the converged reconstructed feature vector; the fault mode data is subjected to time sequence segmentation processing, the fault mode data is divided into time sequence data segments according to a time window, time sequence feature information in the time sequence data segments is extracted, a feature weight is calculated using an attention mechanism, the time sequence feature information is subjected to weighted processing, a feature correlation mapping is established based on the weighted time sequence feature information, and a correlation mapping result is obtained.

[0010] According to the above technical solution, when the intelligent prediction model is trained, if the test result output by the intelligent prediction model does not reach a preset evaluation threshold, the intelligent prediction model is online optimized through a meta-learning algorithm, and a test case is regenerated based on the optimized intelligent prediction model and the test is executed until the test evaluation result reaches the preset evaluation threshold.

[0011] According to the above technical solution, in step S1, the AC voltage signal and the AC current signal are specifically decomposed into multiple layers by wavelet packet decomposition, the energy distribution of each layer of wavelet coefficients is calculated, the wavelet coefficients are processed using a soft threshold processing method, and the wavelet coefficients after soft threshold processing are reconstructed to obtain the electrical signal after noise reduction.

[0012] According to the above technical solution, in step S1, the electrical signal after noise reduction is specifically decomposed into multiple intrinsic mode components, Hilbert transformation is performed on each intrinsic mode component, instantaneous frequency characteristics and instantaneous amplitude characteristics are extracted, and signal characteristics are obtained by combination.

[0013] According to the technical scheme, in step S2, the specific generation process of the test case is as follows: a test case search space is constructed based on the prediction result, the test case search space contains test scene information, test condition information and expected response information, sampling and simulation are performed in the test case search space through a Monte Carlo tree search algorithm, a search path is selected based on a confidence upper limit criterion, a candidate test case set is obtained, the candidate test case set is input into a genetic algorithm for optimization iteration, the candidate test case set is evaluated based on a pre-constructed test case fitness evaluation function, test cases with high fitness are selected for cross and mutation operations, the population is repeatedly updated until a maximum evolution generation is reached, an optimized test case is generated, and the optimized test case contains device model information, test trigger conditions, test input parameters, test execution timing and expected output response; wherein the test input parameters include voltage variation, current variation and phase variation, and the expected output response includes action time requirement and action power requirement.

[0014] According to the technical scheme, in step S3, the response feature vector is specifically compared with the standard response in the test case, the response similarity is calculated by combining a deep metric learning algorithm, and the test evaluation result is generated based on the response similarity and a fuzzy comprehensive evaluation algorithm.

[0015] The application also provides a power distribution terminal automation test system based on a dynamic fault record and analysis recorder, comprising: The first unit is configured to acquire AC voltage signals and AC current signals collected by the dynamic fault record and analysis recorder, perform wavelet packet decomposition and noise reduction processing to obtain noise-reduced electrical signals, transform the noise-reduced electrical signals to obtain signal features containing instantaneous frequency and instantaneous amplitude, extract time domain features and frequency domain features from the signal features, input the extracted time domain features and frequency domain features into a pre-constructed deep convolutional neural network for fault feature classification, and obtain a fault feature vector; The second unit is configured to reconstruct the fault feature vector and perform channel time sequence feature association mapping, input the association mapping result into a pre-trained intelligent prediction model, predict the action response of the power distribution terminal, generate a prediction result, and generate a test case according to the prediction result and the corresponding fault feature vector; The third unit is configured to reconstruct the fault feature vector in the test case into a test signal and output the test signal to the power distribution terminal, the power distribution terminal generates an actual response signal to the input test signal, evaluates the actual response signal based on the prediction result of the intelligent prediction model, and generates a test evaluation result of the response of the power distribution terminal.

[0016] The application further provides a computer storage medium, which stores a computer program executable by a processor, and the computer program executes the power distribution terminal automation test method based on the dynamic fault record and analysis recorder.

[0017] The application has the advantages that: the original electrical signal is denoised by wavelet packet decomposition, the instantaneous frequency and amplitude characteristics are extracted, the time-frequency domain characteristics are extracted and the fault classification is performed, the multi-domain and multi-scale characteristics of the fault signal are fully mined, the accuracy and efficiency of the fault feature extraction are improved, and a good foundation is laid for subsequent fault pattern learning and test case generation; the test signal is reconstructed according to the fault characteristic vector, the actual response signal of the power distribution terminal to the test signal is evaluated together with the prediction result of the intelligent prediction model, the test evaluation result of the power distribution terminal response is generated, and the intelligent level of the power distribution terminal test is improved.

[0018] Further, Monte Carlo tree search and genetic algorithm are introduced to search and evolve the test cases, a balance is obtained between sample efficiency and optimization quality, high-quality test cases with comprehensive coverage and strong pertinence are generated, and the effectiveness and reliability of the test are improved.

[0019] Further, the response characteristics are extracted by the convolutional neural network, and the deep metric learning is introduced to calculate the response similarity, the problems of high-dimension and multi-modal of the response signal are overcome, the accurate characterization and similarity measurement of the response characteristics are realized, and a quantitative index is provided for the test result evaluation.

[0020] Further, the online learning and dynamic optimization mechanism are introduced, so that the test system has the ability of self-improvement and continuous evolution, and important technical support is provided for the intelligent operation and maintenance of the substation and the equipment health management.

[0021] Of course, implementing any product of the present application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0023] Figure 1 is a flow chart of the power distribution terminal automation test method based on the dynamic fault record and analysis recorder according to the embodiment of the present application; Figure 2 is a structural schematic diagram of the power distribution terminal automation test system based on the dynamic fault record and analysis recorder according to the embodiment of the present application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] It should be noted that the illustrations provided in the embodiments of the present invention are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0026] In this invention, it should also be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first" and "second" are used only for descriptive and distinguishing purposes and should not be construed as indicating or implying relative importance.

[0027] Furthermore, it should be noted that the features of the various embodiments of the present invention can be combined or integrated in whole or in part, and as those skilled in the art will understand, they can interact and operate in different ways. Each embodiment can be implemented independently of each other or in association with one another.

[0028] like Figure 1 As shown, the automated testing method for power distribution terminals based on dynamic fault recording and analysis recorders in this embodiment of the invention includes the following steps: S1. Acquire AC voltage and AC current signals collected by the dynamic fault recorder and analyzer, perform wavelet packet decomposition and noise reduction processing to obtain the noise-reduced electrical signal; transform the noise-reduced electrical signal to obtain signal features containing instantaneous frequency and instantaneous amplitude, extract time-domain and frequency-domain features from the signal features through a pre-constructed long short-term memory neural network; input the extracted time-domain and frequency-domain features into a pre-constructed deep convolutional neural network for fault feature classification to obtain the fault feature vector; S2, reconstruct the fault feature vector and perform channel time sequence feature correlation mapping, input the correlation mapping result into the pre-trained intelligent prediction model, predict the action response of the power distribution terminal, generate a prediction result; generate a test case according to the prediction result and the corresponding fault feature vector; S3, reconstruct the fault feature vector in the test case into a test signal, and output it to the power distribution terminal. The power distribution terminal generates an actual response signal to the input test signal, and evaluates the actual response signal using the prediction result of the intelligent prediction model to generate a test evaluation result of the power distribution terminal response.

[0029] Specifically, in step S1, the Hilbert-Huang transform is performed on the noise-reduced electrical signal to obtain signal features including instantaneous frequency and instantaneous amplitude. A long short-term memory neural network is used to construct a time sequence feature extraction model. The signal features are input into the time sequence feature extraction model to extract time domain features and frequency domain features. The time domain features and the frequency domain features are input into a deep convolutional neural network for fault feature classification to obtain a fault feature vector.

[0030] In step S2, when performing time sequence feature correlation mapping of the channel, the fault feature vector channel in the recording waveform file obtained by the recording device is correlated with the output channel of the fault amount output device output to the power distribution terminal.

[0031] In an embodiment of the present application, the dynamic fault recording and analysis recording wave device collects the alternating voltage signal and the alternating current signal of the power distribution terminal. The sampling frequency of the dynamic fault recording and analysis recording wave device is 10 kHz, the sampling time is 100 ms, the sampling accuracy is 16 bits, the amplitude range of the alternating voltage signal is 0-220V, and the amplitude range of the alternating current signal is 0-5A.

[0032] The alternating voltage signal and the alternating current signal are decomposed by wavelet packet, the alternating voltage signal and the alternating current signal are decomposed into 5 layers, db4 wavelet basis function is used for decomposition, the energy distribution of each layer wavelet coefficient is calculated, the soft threshold processing method with threshold value of 0.5 is used to process the wavelet coefficient, the wavelet coefficient after soft threshold processing is reconstructed to obtain the noise-reduced electrical signal. The soft threshold processing method for processing the wavelet coefficient is shown in the following formula: ; Wherein, represents the processed wavelet coefficient of the i-th layer, j d j represents the original wavelet coefficient of the i-th layer, sign( j d j represents a sign function, which is used to indicate​​d j median(|x|) denotes the median of the symbols, d j |) denotes the median of the absolute values of the wavelet coefficients of the j N j denotes the length of the wavelet coefficients of the j

[0033] The denoised electrical signal can also be decomposed into 10 intrinsic modal components, a Hilbert transform is performed on each intrinsic modal component to extract instantaneous frequency features and instantaneous amplitude features, the value range of the instantaneous frequency features is 45-55 Hz, the instantaneous amplitude features of the alternating voltage signal have a unitary value of 0.9-1.1, and the instantaneous amplitude features of the alternating current signal have a unitary value of 0.8-1.2, and the instantaneous frequency features and the instantaneous amplitude features of all the intrinsic modal components are combined to obtain signal features.

[0034] A long short-term memory neural network is constructed, the long short-term memory neural network includes an input layer, a first hidden layer, a second hidden layer and an output layer, the first hidden layer contains 128 neurons, and the second hidden layer contains 64 neurons, the signal features are segmented into segmented signal features according to a time window of 10 ms, the segmented signal features are input into the long short-term memory neural network, and time domain features and frequency domain features are extracted through the long short-term memory neural network.

[0035] A deep convolutional neural network is constructed, the deep convolutional neural network includes three convolutional layers and two fully connected layers, the three convolutional layers contain 32, 64 and 128 3x3 convolutional kernels respectively, and the two fully connected layers contain 256 and 128 neurons respectively, the time domain features and the frequency domain features are combined into a feature vector and input into the deep convolutional neural network, and the feature vector is classified through the deep convolutional neural network to obtain a fault feature vector.

[0036] In an embodiment of the present application, the test case generation process can include: The fault feature vector is obtained, the fault feature vector contains device operating state features, electrical quantity features and environmental parameter features, a variational autoencoder generative adversarial network is used for encoding and reconstruction processing of the fault feature vector, the fault feature vector is encoded into an encoded vector in a latent feature space, a discriminator is used for constraining the distribution of the encoded vector, a reconstructed feature vector is generated based on the encoded vector, the reconstructed feature vector and the original fault feature vector are subjected to adversarial training, network parameters are updated until the reconstruction error converges, and fault mode data is generated based on the converged reconstructed feature vector.

[0037] ​​The fault mode data is subjected to time sequence segmentation processing, and the fault mode data is divided into time sequence data segments according to a time window, the length of the time window is 20 milliseconds, the sliding step length of the time window is 5 milliseconds, the time sequence data segments are subjected to encoding and decoding processing by a dynamic time sequence correlation model, time sequence feature information in the time sequence data segments is extracted by a bidirectional long short-term memory network, feature weights are calculated by using an attention mechanism, the time sequence feature information is subjected to weighting processing, a feature correlation mapping is established based on the weighted time sequence feature information, and a correlation mapping result is obtained.

[0038] The correlation mapping result is added to a preset double deep Q network, a state space and an action space are constructed based on the correlation mapping result, the state space includes device real-time running state, fault feature parameters and environmental monitoring data, the action space includes device control instruction sequence and protection action strategy, the state space and the action space are subjected to interactive learning by the double deep Q network, a state-action value function is constructed, the state-action value function is optimized based on an experience replay mechanism, a device action response strategy is generated, the device action response strategy is input into a target network, an action response of the power distribution terminal is predicted based on the target network, and a prediction result is obtained. The experience replay mechanism is a reinforcement learning technology that stores historical interaction experience and randomly extracts samples for training to break time correlation and improve learning efficiency. The use case search space refers to all possible test paths or combinations that can be explored in the test case generation or selection process, and is usually used for optimization of automated testing. The upper confidence bound criterion is a heuristic algorithm used in Monte Carlo tree search, which balances the strategy of exploring unknown nodes and utilizing known nodes by combining the average reward and visit number of the current node to determine the search direction of the next step.

[0039] A test case search space is constructed based on the prediction result, the test case search space includes test scene information, test condition information and expected response information, sampling and simulation are performed in the test case search space by a Monte Carlo tree search algorithm, a search path is selected based on an upper confidence bound criterion, a candidate test case set is obtained, the candidate test case set is input into a genetic algorithm for optimization iteration, a test case fitness evaluation function is constructed, the candidate test case set is evaluated based on the test case fitness evaluation function, test cases with high fitness are selected for crossover and mutation operations, the population is repeatedly updated until the maximum evolution generation is reached, and an optimized test case is generated. The optimized test case includes device model information, test trigger condition, test input parameter, test execution time sequence and expected output response, the test input parameter includes voltage change amount, current change amount and phase change amount, and the expected output response includes action time requirement and action power requirement, which are used to judge the correctness and timeliness of the device action.

[0040] Specifically, sampling and simulation are performed in the test case search space by the Monte Carlo tree search algorithm, and the search path is selected based on the confidence upper bound criterion as shown in the following formula: ; Wherein, represents the path obtained by the Monte Carlo tree search, v represents the identifier of the current node, p represents the parent node of the current node v , Q ( v ) represents the average return of the current node v , N ( v ) represents the visit times of the current node v , C p represents the exploration constant, N ( p ) represents the visit times of the parent node p , gamma represents the weighted coefficient of the neural network output, f NN ( v ) represents the predicted value of the current node v given by the neural network.

[0041] The confidence upper bound criterion comprehensively considers the average return and visit times of the node when selecting the search path, and realizes the balance of the optimal solution within a limited search times, further improving the search efficiency. In the calculation of the confidence upper bound criterion, the predicted value of the neural network is fused, which can more accurately evaluate the potential of the node, provide strong domain knowledge guidance for the search algorithm, make the search proceed in a more promising direction, reduce blind exploration, and through Monte Carlo simulation to evaluate the search path, which can timely find and prune the hopeless branches, avoid unnecessary resource waste, and continuously improve the quality and efficiency of test case generation. In summary, the embodiment has universality and scalability, and can be widely applied to dynamic fault recording and analysis of substations, even to other industrial fields of automated testing, to improve test efficiency and system reliability, and to promote the development of smart grid.

[0042] In step S3, the fault feature vector in the test case can be reconstructed into a test signal through wavelet inverse transform, the actual response signal of the power distribution terminal to the test signal is collected and time alignment is performed, the actual response signal after time alignment is input into a graph convolutional neural network to extract a response feature vector, the response feature vector is compared with the standard response in the test case, a response similarity is calculated by combining a deep metric learning algorithm, and a test evaluation result is generated based on the response similarity and based on a fuzzy comprehensive evaluation algorithm.

[0043] The wavelet inverse transform is a signal processing method that reconstructs a signal decomposed by wavelet transform back to the original signal through the inverse process of wavelet transform. Wavelet transform decomposes a signal into components of multiple frequency bands, and wavelet inverse transform recombines these components to restore the original signal or an approximate signal. The deep metric learning algorithm is a kind of deep learning method that aims to learn an embedding space in which similar data points are closer and dissimilar data points are farther apart, and is commonly used in tasks such as face recognition and image retrieval. The response similarity is a measure of the similarity between two signals or system responses, commonly used in fault diagnosis and pattern recognition, which evaluates the similarity by calculating the correlation or similarity between signals.

[0044] The training process of the intelligent prediction model mainly includes: The fault feature vector obtained in the above embodiment is identified for power frequency component, harmonic component and transient component, the fault feature vector is decomposed by five layers of wavelet based on db4 wavelet basis function, the high frequency coefficients obtained by decomposition are denoised by soft thresholding, and the low frequency coefficients are denoised by hard thresholding, to generate a test signal containing three-phase voltage, three-phase current and zero sequence component; The protection starting signal, trip signal and reclosing signal output by the secondary equipment are collected, and direct current bias removal, power frequency interference filtering and sampling zero drift elimination are performed, time alignment is performed between the test signal and the secondary equipment through a dynamic time warping algorithm, the time alignment uses an iterative calculation method with a time window width of one hundred milliseconds and a sliding step of ten milliseconds; alignment is achieved through iterative calculation.

[0045] The signal after time alignment is converted into graph structure data, and an eight-layer graph convolutional neural network is used to extract response features, the first four layers of the graph convolutional neural network extract local features, the last four layers extract global features, each layer contains sixty-four three-by-three convolution kernels, which are processed through batch normalization layer and ReLU activation function, and a dropout rate of zero point three is introduced; The response similarity is calculated by a triplet loss network, the response feature is taken as an anchor sample, the standard response feature is taken as a positive sample, and the response feature of the unselected test case is taken as a negative sample, the Euclidean distance is calculated by a deep neural network with a margin parameter of 0.2 mapping to a 128-dimensional metric space, and a fuzzy comprehensive evaluation system is constructed based on the response similarity, including three first-level indexes of time characteristics, amplitude characteristics and waveform characteristics, with weights of 0.4, 0.3 and 0.3 respectively, and three second-level indexes under each first-level index, and a five-level evaluation grade is used to calculate the test evaluation score. When the test evaluation score is less than 0.85, a model-independent meta-learning algorithm is used for optimization, a task distribution is constructed based on the last five test data, the model parameters are updated by an adaptive moment estimation algorithm in an inner loop, the learning rate is determined by a second-order derivative optimization in an outer loop, sixteen tasks are randomly sampled for training in each round, and the test case is regenerated and tested using the optimized model until the test evaluation score is greater than or equal to 0.85. Wherein, the meta-learning algorithm is an algorithm that learns how to learn, which quickly adapts to new tasks or data distribution by training the model, and is usually used for few-shot learning and fast model adaptation problems.

[0046] Wherein, the harmonic component is the periodically changing part of the signal, usually based on the frequency component of the integer multiple of the fundamental frequency, existing in many periodic signals, the transient component is the part of the signal that changes greatly in a short time, usually characterized by sudden or instantaneous events, the sampling zero drift elimination is a signal processing technique used to eliminate the zero drift caused by equipment or environmental factors during sampling, usually achieved by calibration and removal of offset, the triplet loss network is a deep learning loss function used to optimize the relative distance relationship between triplets, commonly used in metric learning tasks such as face verification, the adaptive moment estimation algorithm in the inner loop is an algorithm used to optimize model parameters in meta-learning, which adaptively adjusts the gradient and optimization step in multiple training steps so that the model can quickly adapt to new tasks.

[0047] The time-aligned signal is converted into graph structure data. For example, a graph convolutional neural network containing eight layers is used to extract response features. The first four layers of the network are used to extract local features, and the last four layers are used to extract global features. Each layer contains 64 convolution kernels with a size of 3x3, and is processed through a batch normalization layer and a ReLU activation function. To prevent overfitting, a dropout layer with a dropout rate of 0.3 is introduced at each layer; The response similarity can be calculated using a triple loss network. The response features extracted in step S3 can be taken as anchor samples, the standard response features as positive samples, and the response features of the test cases not selected as negative samples. These samples are mapped to a 128-dimensional metric space by a deep neural network with a margin parameter of 0.2, and the Euclidean distance is calculated. Based on the calculated response similarity, a fuzzy comprehensive evaluation system is constructed. The evaluation system includes three first-level indicators of time characteristics, amplitude characteristics, and waveform characteristics, and their weights are set to 0.4, 0.3, and 0.3, respectively. Three second-level indicators are set under each first-level indicator. A five-level evaluation grade is used to calculate the test evaluation score.

[0048] If the test evaluation score is less than 0.85, a model-independent meta-learning algorithm is used to optimize the model. Based on the task distribution constructed from the last five test data, the model parameters are updated by the self-adaptive matrix estimation algorithm in the inner loop, and the learning rate is determined by the second-order derivative optimization in the outer loop. 16 tasks are randomly sampled for training in each round. The test cases are regenerated and tested using the optimized model, and the process is repeated until the test evaluation score is greater than or equal to 0.85.

[0049] Exemplarily, taking the fault test of a dynamic fault recording and analysis recorder in a substation as an example, a group of fault feature vectors are collected from the dynamic fault recording and analysis recorder, and the power frequency component, harmonic component and transient component are identified. The db4 wavelet basis function is used for five-layer wavelet decomposition of the fault feature vector, and the soft and hard threshold denoising is used for the high and low frequency coefficients respectively, and the test signal containing three-phase voltage, three-phase current and zero sequence component is generated after reconstruction.

[0050] The protection starting signal, tripping signal and reclosing signal output by the secondary equipment of the dynamic fault recording and analysis recorder are collected, and the direct current bias removal, power frequency interference filtering and sampling zero drift elimination preprocessing are performed. The dynamic time warping algorithm is used to time-align the preprocessed signal and the test signal, and the time window width is set to 100 milliseconds and the sliding step is 10 milliseconds.

[0051] The aligned signal is converted into graph structure data, and a response feature is extracted using an eight-layer graph convolutional neural network. The first four layers of the network extract local features, and the last four layers extract global features, each layer has 64 3x3 convolution kernels, and is subjected to batch normalization, ReLU activation and dropout processing with a dropout rate of 0.3.

[0052] The response similarity is calculated using the ternary tuple loss network. The extracted response features, standard response features, and other test case response features are used as anchor samples, positive samples, and negative samples, respectively. The Euclidean distance is calculated by mapping the deep neural network with a margin parameter of 0.2 to a 128-dimensional metric space. A fuzzy comprehensive evaluation system is constructed based on the similarity, including time characteristics, amplitude characteristics, and waveform characteristics, with weights of 0.4, 0.3, and 0.3, respectively. Each first-level index has three second-level indexes, and a five-level evaluation grade is used to give the test score.

[0053] If the test score is less than 0.85, a model-independent meta-learning algorithm is used for optimization. Based on the last five test data, a task distribution is constructed, and the parameters are updated by an adaptive matrix estimation algorithm in the inner loop and a second-order derivative optimization learning rate in the outer loop. Every round randomly samples 16 tasks for training, and the test case is tested using the optimized model until the test score reaches 0.85 or above.

[0054] In this embodiment, test signals containing three-phase voltage, three-phase current, and zero sequence components are generated, covering a variety of fault types and characteristics, ensuring the effectiveness and comprehensiveness of fault feature extraction. The sliding window and iterative optimization mechanism ensure the alignment accuracy and efficiency. Through the end-to-end deep learning method, a global and hierarchical response feature representation is automatically constructed, avoiding the limitations of manually designed features and enhancing the richness and discriminability of feature representation. This lays a foundation for subsequent similarity calculation and evaluation analysis. The measured response and the standard response are accurately similar, and the multi-dimensional response features are measured by the Euclidean distance, forming a comprehensive similarity quantitative index, providing a precise reference for evaluating the test results. In summary, this embodiment realizes the full-process automation and intelligence of the distribution terminal fault test, providing strong technical support for improving the intelligent level of substation operation and maintenance and ensuring the safe and stable operation of the power grid. It has important significance for promoting the development and application of intelligent testing technology for secondary equipment.

[0055] For example, a primary fault in a certain substation is taken as an example, and the AC voltage signal and AC current signal at the fault time are collected. The voltage signal amplitude is 110V, and the current signal amplitude is 2A. The dynamic fault recording and analysis wave recorder is used to collect the voltage and current signals, with a sampling frequency of 10kHz, a time length of 100ms, and a precision of 16 bits. The sampled data is decomposed by 5-layer wavelet packet, using db4 wavelet basis function, and the wavelet coefficients are processed by energy calculation and soft threshold (λ=0.5). Then the noise reduction signal is reconstructed; The noise reduction signal is subjected to empirical mode decomposition to obtain 10 intrinsic mode components. The Hilbert transform is performed on each intrinsic mode component to obtain the instantaneous frequency and instantaneous amplitude. The instantaneous frequency is in the range of 45-55Hz, the instantaneous amplitude of the voltage signal is in the range of 0.9-1.1, and the instantaneous amplitude of the current signal is in the range of 0.8-1.2. The signal features are obtained by combining the instantaneous frequency and amplitude.

[0056] The long short-term memory neural network is constructed, the input layer receives the signal features, the first hidden layer has 128 neurons, and the second hidden layer has 64 neurons. The signal features are segmented at 10ms and input into the long short-term memory neural network to extract the time-frequency features, and the deep convolutional neural network is constructed, which has three convolutional layers and two fully connected layers (256, 128). The time-frequency features extracted by the long short-term memory neural network are input into the deep convolutional neural network for classification, and finally the fault feature vector [0.85, 0.92, 0.35, 0.12, …] is output.

[0057] In an embodiment of the present application, the classification of the deep learning network can realize accurate diagnosis of different types of faults, provide reliable basis for subsequent fault location and cause analysis, the standardized fault feature vector is convenient for data sharing and comparison analysis between different systems, based on the fault knowledge base, the algorithm model can be further optimized to realize fault warning and trend prediction, and the automatic test can obviously improve the equipment reliability and prolong the equipment life. At the same time, the accurate and efficient fault diagnosis can avoid unnecessary shutdown and maintenance, reduce the maintenance cost, improve the system economy, and the embodiment can effectively guide the operation and maintenance of the power distribution terminal, improve the system reliability and economy, and has important significance for guaranteeing the safe and stable operation of the power grid and improving the power supply service quality.

[0058] In an embodiment of the present application, the threshold value can be flexibly set according to the statistical characteristics of each layer of wavelet coefficients, effectively adapt to different decomposition levels and different signal types, the median value is not sensitive to abnormal values, can avoid the interference of individual larger noise coefficients on the threshold value, improve the denoising robustness, the adaptive threshold strategy simplifies the denoising process, improves the practicability and engineering value of the algorithm, retains the sign information of the coefficients, and the processed coefficients continuously change around the original values, which is helpful to improve the smoothness and continuity of the reconstructed signal, reduce the artificial distortion introduced by the threshold processing, and the embodiment can significantly improve the denoising performance and fault feature extraction accuracy, has strong engineering practical value, and provides strong support for the popularization and application of the dynamic fault recording and analysis recorder.

[0059] In step S2, based on the fault feature vector, a variational autoencoder generative adversarial network is used to build a fault mode learning model to obtain fault mode data, a sequence-to-sequence model is used to process the fault mode data, a dynamic time sequence correlation model is established to obtain a correlation mapping result, an intelligent prediction model is built based on the correlation mapping result and a deep reinforcement learning algorithm, the action response of the power distribution terminal is predicted through the intelligent prediction model to obtain a prediction result, and an optimized test case is generated through a Monte Carlo tree search algorithm and a genetic algorithm based on the prediction result.

[0060] The variational autoencoder is a generative model that learns the probability distribution of latent variables by maximizing the lower bound of evidence, and is used to generate new data samples while having the ability to reconstruct input data and regularize latent space. The sequence-to-sequence model is a deep learning architecture commonly used for tasks where input and output are sequences, such as machine translation and speech recognition, which maps one sequence to another through an encoder-decoder structure. The Monte Carlo tree search algorithm is a search algorithm used in decision-making processes that gradually searches for the optimal strategy through four steps of selection, expansion, simulation, and backtracking, and is widely used in game problems such as Go and chess.

[0061] For example, taking the 110kV line protection device of a certain substation as an example, the voltage and current data at the time of a fault are collected by a dynamic fault recorder and analyzer, and a fault feature vector is extracted. The fault feature vector includes voltage, current, frequency, and other electrical quantity characteristics before and after the fault, as well as temperature, humidity, and other environmental parameter characteristics.

[0062] The variational autoencoder generative adversarial network is used to encode the fault feature vector and map it to a latent feature space, while the discriminator is used to constrain the encoding vector to follow a standard normal distribution. The fault feature is reconstructed based on the encoding vector, and the reconstruction error is minimized through adversarial training. After 1000 iterations of training, the reconstructed feature vector is obtained, and the fault mode data is generated.

[0063] The fault mode data is time-sequentially divided into time sequence segments with a 20ms time window and a 5ms sliding step. Bidirectional long short-term memory networks are used to extract time sequence features, and attention mechanisms are used to calculate the weights of each feature. The weighted time sequence features are correlated and mapped to establish relationships between features.

[0064] The association mapping result is input into a double deep Q network to construct a state space and an action space. The state space includes real-time running states such as line voltage, current and power, and fault characteristic parameters such as fault type and fault distance. The action space includes protection action strategies such as tripping and reclosing, and instruction sequences for controlling switch actions. The mapping from state to action is learned through the Q network, and the experience replay optimization strategy is used. Finally, the action response strategy for different fault conditions is generated, and the target network is used for prediction verification.

[0065] According to the prediction result, a test case search space is constructed, including scene information such as fault type, fault location and fault impedance, and expected response such as action time and action current. 1000 times are sampled in the search space using Monte Carlo tree search, and the optimal search path is selected based on the confidence upper limit to obtain 100 candidate test cases. The candidate test cases are input into a genetic algorithm, and the test coverage and fault detection rate are used as the fitness function for 50 generations of evolution optimization. Finally, 10 optimized test cases are obtained, each including fault scene setting, test input parameter and expected action response.

[0066] In this embodiment, the association rules between fault features are automatically learned, without the need for manual design of complex feature engineering, greatly improving the efficiency of feature extraction. By constructing feature association mapping, the spatio-temporal relationship and dynamic rules hidden in the fault mode can be found, providing more comprehensive and accurate basis for subsequent decision optimization. The reinforcement learning algorithm can autonomously find the best decision path in complex and variable fault scenarios through continuous trial and error and experience accumulation, avoiding the blindness and limitations of manually designed strategies. Through automated and intelligent test schemes, the test efficiency and accuracy can be significantly improved, the manual input and error risk can be reduced, and the intelligent operation and maintenance of the substation can be effectively supported. In summary, this embodiment can significantly improve the test level and operation and maintenance efficiency of the power distribution terminal, reduce hidden dangers, and improve system safety and reliability. It is of great significance to promote the construction of intelligent operation and maintenance system of substations, ensure the safe and stable operation of power grids, and provides a good demonstration and reference for intelligent testing and diagnosis in other industrial fields.

[0067] In summary, in the best embodiment of the present application, the original electrical signal is denoised by wavelet packet decomposition, the instantaneous frequency and amplitude characteristics are extracted by combining Hilbert-Huang transform, and the time-frequency domain features are extracted and fault classification is performed using long short-term memory neural network and deep convolutional neural network, which fully excavates the multi-domain and multi-scale features of the fault signal, improves the accuracy and efficiency of fault feature extraction, lays a good foundation for subsequent fault mode learning and test case generation, establishes dynamic time sequence association by combining sequence-to-sequence model, reveals the fault evolution law, obtains comprehensive and accurate fault mode knowledge, provides sufficient prior information and data support for intelligent testing, introduces Monte Carlo tree search and genetic algorithm for test case search and evolution, balances between sample efficiency and optimization quality, generates high-quality test cases with comprehensive coverage and strong pertinence, improves the effectiveness and reliability of testing, extracts response features by graph convolutional neural network, and introduces deep metric learning to calculate response similarity, overcoming the difficulty of high-dimensional and multi-modal response signal, realizing accurate characterization and similarity measurement of response features, and providing quantitative indicators for test result evaluation. In summary, the present application greatly saves manpower and resources, reduces human errors, improves test quality and reliability, and introduces online learning and dynamic optimization mechanism, enabling the testing system to have the ability of self-improvement and continuous evolution, providing important technical support for intelligent operation and maintenance of substations and equipment health management.

[0068] Figure 2 The structure schematic diagram of the power distribution terminal automatic test system based on dynamic fault record and analysis recorder according to the embodiment of the present application, the system comprises: The first unit is used for acquiring the AC voltage signal and the AC current signal collected by the dynamic fault record and analysis recorder, performing wavelet packet decomposition denoising processing to obtain the denoised electrical signal, transforming the denoised electrical signal to obtain signal characteristics containing instantaneous frequency and instantaneous amplitude, and extracting time domain features and frequency domain features in the signal characteristics; the extracted time domain features and frequency domain features are input into a pre-constructed deep convolutional neural network for fault feature classification to obtain a fault feature vector; The second unit is used for reconstructing the fault feature vector and performing channel time sequence feature association mapping, inputting the association mapping result into a pre-trained intelligent prediction model, predicting the action response of the power distribution terminal, generating a prediction result, and generating a test case according to the prediction result and the corresponding fault feature vector; The third unit is used for reconstructing the fault feature vector in the test case into a test signal and outputting the test signal to the power distribution terminal, the power distribution terminal generating an actual response signal to the input test signal, evaluating the actual response signal by using the prediction result of the intelligent prediction model, and generating a test evaluation result of the response of the power distribution terminal.

[0069] The units of the system are mainly used for implementing the steps of the above method embodiments, and will not be described here.

[0070] The application further provides a computer readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card memory (for example, an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read only memory (ROM), an electrically erasable programmable read only memory (EEPROM), a programmable read only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, a server, an App application store, etc., which stores a computer program, and the program is executed by a processor to realize corresponding functions. The computer readable storage medium of the embodiment is executed by the processor to realize the power distribution terminal automation testing method based on dynamic fault recording and analysis of the recorder according to the method embodiment.

[0071] It should be noted that, according to the needs of implementation, each step / component described in the application can be split into more steps / components, or two or more steps / components or part of the operation of the step / component can be combined into a new step / component, to achieve the purpose of the application.

[0072] The size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the application.

[0073] It should be understood that, for those skilled in the art, improvements or changes can be made according to the above description, and all these improvements and changes should belong to the protection scope of the appended claims of the application.

Claims

1. An automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder, characterized in that, Includes the following steps: S1. Acquire AC voltage and AC current signals collected by the dynamic fault recorder and analyzer, perform wavelet packet decomposition and noise reduction processing to obtain the noise-reduced electrical signal; transform the noise-reduced electrical signal to obtain signal features containing instantaneous frequency and instantaneous amplitude, and then extract the time-domain and frequency-domain features from the signal features; input the extracted time-domain and frequency-domain features into a pre-constructed deep convolutional neural network for fault feature classification to obtain the fault feature vector; S2. Reconstruct the fault feature vector and perform temporal feature association mapping of the channel. Input the association mapping result into the pre-trained intelligent prediction model to predict the action response of the power distribution terminal and generate prediction results. Generate test cases based on the prediction results and the corresponding fault feature vector. S3. Reconstruct the fault feature vector in the test case into a test signal and output it to the power distribution terminal. The power distribution terminal generates an actual response signal to the input test signal, evaluates the actual response signal using the prediction results of the intelligent prediction model, and generates a test evaluation result of the power distribution terminal response.

2. The automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder according to claim 1, characterized in that, In step S1, the noise-reduced electrical signal is subjected to Hilbert-Huang transform to obtain signal characteristics containing instantaneous frequency and instantaneous amplitude.

3. The automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder according to claim 1, characterized in that, The fault feature vector includes equipment operating status characteristics, electrical quantity characteristics, and environmental parameter characteristics.

4. The automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder according to claim 1, characterized in that, In step S2, the fault feature vector is specifically encoded into an encoding vector in the latent feature space. The distribution of the encoding vector is constrained by a discriminator. A reconstructed feature vector is generated based on the encoding vector. The reconstructed feature vector is then used in adversarial training with the original fault feature vector. The network parameters are updated until the reconstruction error converges. Fault mode data is generated based on the converged reconstructed feature vector. The fault mode data is subjected to time-series segmentation processing, which divides the fault mode data into time-series data segments according to time windows. The time-series feature information in the time-series data segments is extracted, and the feature weights are calculated using an attention mechanism. The time-series feature information is then weighted, and a feature association mapping is established based on the weighted time-series feature information to obtain the association mapping result.

5. The automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder according to claim 1, characterized in that, If the test results output by the intelligent prediction model do not reach the preset evaluation threshold during training, the intelligent prediction model is optimized online through a meta-learning algorithm. Test cases are then regenerated and executed based on the optimized intelligent prediction model until the test evaluation results reach the preset evaluation threshold.

6. The automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder according to claim 1, characterized in that, In step S1, the AC voltage signal and the AC current signal are specifically decomposed into multiple layers using wavelet packets, the energy distribution of wavelet coefficients in each layer is calculated, the wavelet coefficients are processed using a soft thresholding method, and the wavelet coefficients after soft thresholding are reconstructed to obtain the noise-reduced electrical signal.

7. The automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder according to claim 1, characterized in that, In step S1, the noise-reduced electrical signal is specifically decomposed into multiple intrinsic mode components. A Hilbert transform is performed on each intrinsic mode component to extract instantaneous frequency features and instantaneous amplitude features, which are then combined to obtain signal features.

8. The automated testing method for power distribution terminals based on a dynamic fault recording and analysis waveform recorder according to claim 1, characterized in that, In step S2, the specific process of generating test cases is as follows: A test case search space is constructed based on the prediction results. This search space includes test scenario information, test condition information, and expected response information. A Monte Carlo tree search algorithm is used to sample and simulate within the test case search space. A search path is selected based on the upper confidence limit criterion to obtain a candidate test case set. This candidate test case set is then input into a genetic algorithm for optimization iteration. The candidate test case set is evaluated based on a pre-constructed test case fitness evaluation function. Test cases with high fitness are selected for crossover and mutation operations. The population is repeatedly updated until the maximum number of generations is reached, generating optimized test cases. The optimized test cases include device model information, test trigger conditions, test input parameters, test execution timing, and expected output response. The test input parameters include voltage changes, current changes, and phase changes. The expected output response includes action time requirements and action power requirements.

9. An automated testing system for power distribution terminals based on a dynamic fault recording and analysis waveform recorder, characterized in that, include: The first unit is used to acquire AC voltage and AC current signals collected by the dynamic fault recorder and analyze them, perform wavelet packet decomposition and noise reduction processing to obtain the noise-reduced electrical signal; transform the noise-reduced electrical signal to obtain signal features containing instantaneous frequency and instantaneous amplitude, and then extract time-domain and frequency-domain features from the signal features; input the extracted time-domain and frequency-domain features into a pre-constructed deep convolutional neural network for fault feature classification to obtain the fault feature vector; The second unit is used to reconstruct the fault feature vector and perform temporal feature association mapping of the channel. The association mapping result is input into the pre-trained intelligent prediction model to predict the action response of the power distribution terminal and generate prediction results. Test cases are generated based on the prediction results and the corresponding fault feature vector. The third unit is used to reconstruct the fault feature vector in the test case into a test signal and output it to the power distribution terminal. The power distribution terminal generates an actual response signal to the input test signal, evaluates the actual response signal using the prediction results of the intelligent prediction model, and generates a test evaluation result of the power distribution terminal response.

10. A computer storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the automated testing method for power distribution terminals based on a dynamic fault recording and analysis recorder as described in any one of claims 1-8.

Images