Ground fault detection method and device and computer equipment

Abstract

The invention relates to a ground fault detection method and device and computer equipment, and the method comprises the steps: carrying out the feature extraction of pre-obtained zero-sequence current data of a feeder line of a power distribution network, and obtaining a fault feature; and inputting the fault features into a pre-constructed fault recognition model for ground fault detection processing to obtain a target detection result. According to the fault identification model, a training data set is updated based on zero-sequence current data, and incremental learning is carried out according to the updated training data set to dynamically adjust model parameters. The incremental learning enables the model to continuously adapt to continuously changing actual working conditions and newly appearing data conditions in the operation process of the power distribution network, and effectively overcomes the problem of performance reduction possibly appearing when a traditional model faces new data. According to the method, new knowledge can be continuously learned and integrated, so that high sensitivity and accurate recognition capability of various grounding fault types are kept, and the generalization performance and adaptability of the model are remarkably improved.

Description

Technical Field

[0001] The present application relates to the technical field of distribution network fault identification, and in particular to a ground fault detection method, device and computer equipment. Background Art

[0002] In recent years, with the increasing complexity of distribution lines, the probability of distribution network failures has also increased. Rapid identification, location, and removal of distribution network failures are important ways to enhance power supply reliability and are also a challenge for the development of modern intelligent distribution networks. According to relevant statistics, single-phase grounding faults are one of the most common types of faults in distribution networks, accounting for about 80% of the total number of distribution network failures, while HIF (High Impedance Fault) accounts for about 5%-10% of them, and the actual proportion may be higher.

[0003] Current HIF detection methods usually use static intelligent machine learning models, which are difficult to make correct decisions when faced with data that differs greatly from the distribution of training data, resulting in reduced detection performance. Summary of the invention

[0004] Based on this, it is necessary to provide a ground fault detection method, device and computer equipment that can solve the above-mentioned technical problems.

[0005] In a first aspect, the present application provides a ground fault detection method. The method comprises:

[0006] Perform feature extraction processing on the zero-sequence current data of the feeder line of the distribution network acquired in advance to obtain fault features;

[0007] The fault characteristics are input into a pre-built fault identification model for ground fault detection processing to obtain a target detection result; the model parameters of the fault identification model are updated based on the zero-sequence current data training data set, and are dynamically adjusted by incremental learning according to the updated training data set; the target detection result is used to determine the ground fault type corresponding to the fault characteristics.

[0008] In one embodiment, the fault identification model includes an input layer, a hidden layer network and an output layer; the fault characteristics are input into the pre-built fault identification model to perform ground fault detection processing to obtain a target detection result, including:

[0009] Input the fault features into the input layer for data conversion processing to obtain the fault vector;

[0010] Input the fault vector into the hidden layer network for multi-layer nonlinear transformation processing to obtain a nonlinear transformation result;

[0011] The nonlinear transformation results are input into the output layer for classification decision processing to obtain the target detection results.

[0012] In one embodiment, the method further comprises:

[0013] The updated training data set is input into the fault identification model trained in the historical period to perform ground fault detection processing, and multiple historical detection results are obtained;

[0014] According to each historical detection result and the pre-constructed target loss function, the model parameters of the fault recognition model trained in the historical period are adjusted to obtain the fault recognition model after incremental learning.

[0015] In one embodiment, the process of determining the updated training data set includes:

[0016] Performing confidence assessment processing on the zero-sequence current data to obtain a confidence assessment result of the zero-sequence current data;

[0017] When the confidence evaluation result is greater than a preset verification threshold and the total number of zero-sequence current data is less than a preset storage threshold, determining an additional training set according to the zero-sequence current data;

[0018] The training data set is updated according to the additional training set to obtain an updated training data set.

[0019] In one embodiment, the process of constructing the above objective loss function includes:

[0020] Construct a penalty term; the penalty term includes the corresponding relationship between the training data set, the updated training data set and the maximum mean difference;

[0021] Based on the penalty term and the pre-established cross entropy loss function, the target loss function is determined.

[0022] In one embodiment, the model parameters of the fault recognition model trained in the historical period are adjusted according to each historical detection result and the pre-constructed target loss function, and the fault recognition model after incremental learning is obtained, including:

[0023] Performing importance evaluation processing on multiple neurons in the hidden layer network of the fault recognition model that has been trained in the historical period, and obtaining the second norms of multiple neuron weights;

[0024] Determine multiple important neurons according to the second norm of each neuron weight and the preset norm;

[0025] According to the neuron parameters of each important neuron and the target loss function, the model parameters of the fault recognition model trained in the historical period are adjusted to obtain the fault recognition model after incremental learning.

[0026] In one embodiment, the method further comprises:

[0027] Obtain multiple historical network layer weights in the hidden layer network of the fault recognition model trained in the historical period and multiple target network layer weights of the fault recognition model after incremental learning;

[0028] The weighted two-norm calculation process is performed on each historical network layer weight and each target network layer weight respectively, so as to obtain a plurality of historical weighted two-norms and a plurality of target weighted two-norms;

[0029] Input each historical weighted two-norm and each target weighted two-norm into the average value function to calculate the average value, and obtain the historical two-norm average value and the target two-norm average value;

[0030] The historical two-norm average value and the target two-norm average value are ratio processed, and the ratio result is multiplied by the weight of each target network layer to obtain the optimized target network layer weight.

[0031] In one embodiment, the above-mentioned feature extraction processing is performed on the pre-acquired zero-sequence current data of the distribution network feeder line to obtain the fault features including:

[0032] The zero-sequence current data of the distribution network feeder lines are processed by wavelet decomposition to obtain multiple levels of detail coefficients;

[0033] The standard deviation of detail coefficients at each level is calculated to obtain fault characteristics.

[0034] In a second aspect, the present application also provides a ground fault detection device. The device comprises:

[0035] A feature extraction module is used to perform feature extraction processing on the pre-acquired zero-sequence current data of the distribution network feeder line to obtain fault features;

[0036] The fault detection module is used to input the fault characteristics into a pre-built fault identification model for ground fault detection processing to obtain a target detection result; the model parameters of the fault identification model are based on the zero-sequence current data to update the training data set, and are dynamically adjusted by incremental learning according to the updated training data set; the target detection result is used to determine the ground fault type corresponding to the fault characteristics.

[0037] In a third aspect, the present application further provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method in the first aspect when executing the computer program.

[0038] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method in the first aspect when executed by a processor.

[0039] In a fifth aspect, the present application further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the method in the first aspect are implemented.

[0040] The above-mentioned ground fault detection method, device and computer equipment obtain fault features by performing feature extraction processing on the pre-acquired zero-sequence current data of the distribution network feeder line; the fault features are input into the pre-built fault identification model for ground fault detection processing to obtain the target detection result. The fault identification model of the present application adopts a method of updating the training data set based on zero-sequence current data, and dynamically adjusting the model parameters through incremental learning based on the updated training data set. This incremental learning enables the model to continuously adapt to the ever-changing actual operating conditions and newly emerging data situations during the operation of the distribution network, effectively overcoming the performance degradation problem that may occur in traditional models when facing new data. It can continuously learn and integrate new knowledge, thereby maintaining high sensitivity and accurate recognition capabilities for various types of ground faults, significantly improving the generalization performance and adaptability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is an application environment diagram of a ground fault detection method in an embodiment;

[0042] Figure 2 is a schematic flow chart of a ground fault detection method in one embodiment;

[0043] Figure 3 A schematic diagram of a process for obtaining a target detection result in one embodiment;

[0044] Figure 4 is a flow chart of a ground fault detection method in another embodiment;

[0045] Figure 5 is a flowchart of a process for determining an updated training data set in one embodiment;

[0046] Figure 6 Schematic diagram of a process of constructing a target loss function in one embodiment;

[0047] Figure 7 A schematic diagram of a process for obtaining a fault identification model after incremental learning in one embodiment;

[0048] Figure 8 is a schematic diagram of a model neuron in one embodiment;

[0049] Fig. 9 A schematic diagram of a process of determining an optimized target network layer weight in one embodiment;

[0050] Fig.10 A schematic diagram of a process for obtaining fault characteristics in an embodiment;

[0051] Fig.11 A schematic diagram of obtaining multi-level detail coefficients and approximate coefficients of a waveform signal in one embodiment;

[0052] Fig.12 A radial distribution network simulation model in one embodiment;

[0053] Fig.13 is a structural block diagram of a ground fault detection device in one embodiment;

[0054] Fig.14 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0056] First of all, before specifically introducing the technical solution of the embodiments of the present application, the technical background based on the embodiments of the present application is introduced.

[0057] As the link in the power system that is most closely connected with users, the distribution network provides electricity guarantee for residents' lives and industrial production. With the increase in the construction of smart distribution networks in my country and the access of a large number of distributed power sources, the structure of the distribution network has become increasingly complex, and the probability of failure has also increased. Therefore, it is very necessary to improve the power supply quality and reliability of the distribution network. According to relevant statistics, single-phase grounding fault is one of the most important types of faults in the distribution network, accounting for about 80% of the total number of distribution network faults, while high-impedance grounding fault (HIF) accounts for about 5% to 10%, and the actual proportion may be higher. In recent years, with the increase in the complexity of distribution lines, the probability of distribution network failures has also increased. The rapid identification, location and removal of distribution network faults is an important way to enhance power supply reliability, and it is also a challenge for the development of modern smart distribution networks.

[0058] The main identification methods for high-resistance grounding faults in distribution networks include time domain method, frequency domain method, time-frequency domain method and artificial intelligence method. The time domain method usually uses the time domain characteristics such as nonlinear distortion and asymmetry of the fault signal as the basis for HIF identification, which is easy to analyze and derive, and has a clear physical meaning; the frequency domain method captures the low-frequency component and high-frequency component characteristics of the fault waveform, amplifies the fault frequency domain characteristics to achieve fault detection; the time-frequency domain method comprehensively utilizes the time domain and frequency domain characteristics of the fault signal to identify HIF, and has a strong ability to characterize features. The artificial intelligence method usually combines signal processing technology with artificial intelligence algorithms to form a HIF intelligent identification method. This method can realize adaptive learning of deep features in massive data and avoid the limitations of artificially extracted fault features on the experience of the prophet.

[0059] Based on this, the present application provides a ground fault detection method, device and computer equipment, aiming to solve the above technical problems.

[0060] The ground fault detection method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The server obtains the fault characteristics by performing feature extraction processing on the zero-sequence current data of the feeder line of the distribution network acquired in advance; the fault characteristics are input into the pre-built fault identification model for ground fault detection processing to obtain the target detection result. Among them, the terminal 102 can be but is not limited to a smart meter, a power sensor, and an edge computing device, and the server 104 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0061] In an exemplary embodiment, Figure 2 As shown, a ground fault detection method is provided, which is applied to Figure 1 The server in the example is used to illustrate the following steps:

[0062] S201, performing feature extraction processing on the pre-acquired zero-sequence current data of the feeder line of the distribution network to obtain fault features.

[0063] Among them, the distribution network feeder line refers to the power grid that receives electric energy from the transmission network or regional power plants and distributes it locally or step by step according to voltage to various users through distribution facilities. The feeder line is the line in the distribution network that distributes electric energy from the substation bus to various load points (such as users, distribution boxes, etc.). Zero-sequence current data is usually collected on these feeder lines to monitor and analyze whether the lines have abnormal conditions such as ground faults.

[0064] Zero-sequence current data refers to the situation in which the vector sum of the three-phase currents is zero under normal circumstances in a three-phase AC power system. When a ground fault occurs in the system (such as a single-phase ground fault), the balance of the three-phase currents is broken, and an additional current is generated. This current forms a loop through the earth, and its magnitude and direction are related to the degree of imbalance of the three-phase currents. This current is called zero-sequence current. The data obtained by monitoring and collecting the zero-sequence current on the feeder line of the distribution network is zero-sequence current data, which contains important information about the ground fault.

[0065] Fault features are some characteristic information that can characterize the characteristics of ground faults, extracted from the original zero-sequence current data through specific processing and analysis. These features can be numerical, vector or other forms of data, such as the statistical value of specific coefficients obtained through a certain transformation (such as discrete wavelet transform), etc. The fault identification model can determine whether a ground fault occurs and determine the specific type of fault based on these fault features.

[0066] The embodiment of the present application installs a high-precision zero-sequence current sensor at a key position of the distribution network feeder line, sets the sampling frequency to, for example, 8kHz, continuously collects zero-sequence current data of the distribution network feeder line, and sets the collection time to cover a time range of eight cycles to ensure that a sufficiently complete and representative zero-sequence current signal is obtained. Then, the data is subjected to signal preprocessing, discrete wavelet transform, and feature calculation to obtain the extracted fault features.

[0067] Another implementation method is to use the built-in current monitoring function of the smart meter to obtain the zero-sequence current data of the distribution network feeder line. The smart meter usually samples the zero-sequence current at a certain time interval (such as every 10 milliseconds), and the collection time span is all the sampled data in the past hour.

[0068] The collected zero-sequence current data within one hour is segmented into 10-minute segments to obtain 6 segments of data. Fast Fourier transform (FFT) is performed on each segment of zero-sequence current data to convert the time domain signal into a frequency domain signal to obtain the spectrum of each segment of data. By analyzing the peak frequency, spectrum bandwidth and other characteristic information in the spectrum, it can be preliminarily determined whether there are frequency components that may be related to ground faults.

[0069] According to the results of spectrum analysis, several representative frequency points in each spectrum diagram (such as the peak frequency and its adjacent frequency points) are selected, the amplitudes corresponding to these frequency points are calculated, and these amplitudes and their changes in different time periods (i.e., different segments of data) are used as fault characteristics.

[0070] S202, inputting the fault characteristics into a pre-built fault identification model to perform ground fault detection processing to obtain a target detection result.

[0071] Among them, the fault identification model is a model built based on machine learning or deep learning algorithms. Its purpose is to establish a mapping relationship between fault characteristics and ground fault types by learning a large amount of known sample data with ground fault labels (including corresponding fault characteristics and fault types). The model parameters of the fault identification model are based on the zero-sequence current data to update the training data set, and are dynamically adjusted through incremental learning based on the updated training data set.

[0072] Incremental learning is a learning strategy for machine learning. Different from the traditional one-time use of all training data for model training, incremental learning allows the model to gradually update and optimize its own model parameters as new data continues to arrive, based on the initial training that has been completed.

[0073] The target detection result is the result of the model outputting whether a ground fault occurs and, if a fault occurs, what type of ground fault it is after the extracted fault features are input into the fault identification model for ground fault detection processing. This result can be a classification label (such as high-resistance ground fault, low-resistance ground fault, etc.) or a probability value vector (indicating the probabilities of various possible fault types). The operation and maintenance personnel can determine the type of ground fault corresponding to the fault features based on this target detection result, and then take corresponding maintenance measures.

[0074] In the embodiment of the present application, first: First, a large amount of distribution network feeder line grounding fault sample data (including corresponding fault features and grounding fault types) collected and annotated in advance is used to initialize the fault identification model for training. The fault identification model here can be constructed based on a neural network architecture, such as a multi-layer perceptron (MLP) structure, and the number of input layer nodes is set to correspond to the number of extracted fault features (such as the 6 standard deviation features extracted by discrete wavelet transform, the number of input layer nodes is 6), the number of output layer nodes corresponds to the number of grounding fault types to be identified (assuming that there are 5 common grounding fault types, the number of output layer nodes is 5), and several hidden layers are set in the middle, and the number of hidden layer nodes is determined according to experience and experiments. Through conventional training methods such as back propagation algorithms, the model is trained using initial sample data to enable it to initially have the ability to identify different types of grounding faults. Then, the training data set is updated, incremental learning is adjusted, and fault detection is performed to determine the type of grounding fault that actually occurred.

[0075] Another implementation method: A model architecture based on a support vector machine (SVM) can be selected as a fault identification model, and a kernel function (such as a radial basis function kernel) can be used to handle nonlinear separable situations. The model is initialized and trained using pre-collected distribution network feeder line grounding fault sample data.

[0076] When new zero-sequence current data is obtained, it is processed for feature extraction (using another feature extraction method suitable for SVM, such as principal component analysis (PCA) combined with local linear embedding (LLE), first reducing the data dimension through PCA, and then further extracting nonlinear features through LLE) to obtain new fault features. These new fault features and the corresponding actual grounding fault types are then added to the original training data set to form an updated training data set.

[0077] For the fault identification model based on SVM, the incremental support vector machine (ISVM) algorithm is used for incremental learning and dynamic adjustment. When processing new data, the ISVM algorithm will adapt to the new data situation by adjusting the support vector, kernel function parameters, etc. according to the relationship between the new data and the existing data, thereby improving the generalization ability of the model. In this process, indicators that measure the difference in data distribution (such as the maximum mean difference (MMD)) can also be embedded in the loss function as a penalty term (such as embedding MMD into the objective function in the dual problem of SVM).

[0078] Finally, the newly extracted fault features are input into the fault identification model after incremental learning and dynamic adjustment for grounding risk detection. According to the principle of SVM, the model will output a classification result as the target detection result based on the input fault features by calculating the distance between it and the support vector.

[0079] The above-mentioned ground fault detection method, device and computer equipment obtain fault features by performing feature extraction processing on the pre-acquired zero-sequence current data of the distribution network feeder line; the fault features are input into the pre-built fault identification model for ground fault detection processing to obtain the target detection result. The fault identification model of the present application adopts a method of updating the training data set based on zero-sequence current data, and dynamically adjusting the model parameters through incremental learning based on the updated training data set. This incremental learning enables the model to continuously adapt to the ever-changing actual operating conditions and newly emerging data situations during the operation of the distribution network, effectively overcoming the performance degradation problem that may occur in traditional models when facing new data. It can continuously learn and integrate new knowledge, thereby maintaining high sensitivity and accurate recognition capabilities for various types of ground faults, significantly improving the generalization performance and adaptability of the model.

[0080] In an exemplary embodiment, based on the above embodiment, see Figure 3 The fault identification model of the embodiment of the present application includes an input layer, a hidden layer network and an output layer; the embodiment of the present application involves a process of inputting fault features into a pre-built fault identification model to perform ground fault detection processing and obtain a target detection result, including the following steps:

[0081] S301, input the fault feature into the input layer for data conversion processing to obtain a fault vector.

[0082] In the embodiment of the present application, according to the number of fault features extracted, for example, the input layer is set to have 6 nodes. Each node receives a fault feature value. Data normalization: First, the input fault features are normalized to ensure that feature values ​​of different scales have the same weight and influence in subsequent processing, and the normalized fault feature values ​​are arranged in the order of the input layer nodes to form a vector, namely the fault vector.

[0083] S302, inputting the fault vector into the hidden layer network for multi-layer nonlinear transformation processing to obtain a nonlinear transformation result.

[0084] The hidden layer network of the embodiment of the present application can adopt a multi-layer perceptron (MLP) structure with 3 hidden layers, and each hidden layer has 10, 8, and 6 neurons respectively.

[0085] Before starting to train the model, first initialize the connection weights of each neuron in the hidden layer network. Using the random initialization method, assign a random value in the interval [-0.1, 0.1] to the connection weights between each neuron and the previous layer node. Select a suitable activation function for each neuron in the hidden layer network, and input the fault vector into the first hidden layer. For each neuron in the first hidden layer, calculate according to its connection weight and activation function. Take the output vector of the first hidden layer as input, and continue to input it into the second hidden layer and the third hidden layer in turn for processing according to the above method. For each neuron in the second hidden layer, also calculate according to its connection weight and activation function to obtain the output vector of the second hidden layer; then input it into the third hidden layer to obtain the final nonlinear transformation result.

[0086] S303, inputting the nonlinear transformation result into the output layer for classification decision processing to obtain the target detection result.

[0087] The embodiment of the present application inputs the nonlinear transformation result (i.e., the output vector of the last layer of the hidden layer network) into the output layer. For each node of the output layer, the output value of each node is calculated according to its connection weight and the input vector. According to the characteristics of the Softmax function, this output value can be regarded as the probability value of the occurrence of the ground fault type. The target detection result can be determined according to the probability value, that is, the ground fault type corresponding to the node with the largest probability value is the ground fault type that the model determines is most likely to occur.

[0088] In the process of performing data conversion processing on the fault features at the input layer to obtain the fault vector, the embodiment of the present application can make the input data more standardized and unified, whether through normalization processing or other preprocessing methods, so as to facilitate subsequent processing by the hidden layer network and reduce errors caused by inconsistent data formats or differences in data scales.

[0089] In an exemplary embodiment, based on the above embodiment, see Figure 4 , the method of the embodiment of the present application further includes the following steps:

[0090] S401, inputting the updated training data set into the fault recognition model trained in the historical period to perform ground fault detection processing, and obtaining multiple historical detection results.

[0091] In the embodiment of the present application, it is first ensured that the updated training data set has been completely collected and organized, wherein each sample data contains the fault features after feature extraction and the corresponding actual ground fault type label. Assume that the updated training data set contains N samples, and the fault feature vector dimension of each sample is M.

[0092] Get the fault recognition model that has been trained in the historical period. The model is assumed to be built based on the multi-layer perceptron (MLP) structure, has completed the initial training phase, and has an input layer, several hidden layers, and an output layer. The number of input layer nodes is set according to the fault feature dimension determined in the previous training. Here, it is assumed to be consistent with the fault feature vector M dimension of the currently updated training data set; the number of output layer nodes is set according to the number of ground fault types to be identified. Assuming that K is the type of ground fault, the number of output layer nodes is K.

[0093] Each sample of the updated training data set is input into the fault identification model that has been trained in the historical period in turn. For the fault feature vector of each sample, the data format is adapted according to the input layer requirements of the model (for example, normalization and other preprocessing operations may be required to keep consistent with the input processing method when the model was trained before), and then processed by multi-layer nonlinear transformation of the hidden layer (such as using neurons with ReLU activation function for calculation processing, the specific calculation method is as described above: first calculate the weighted sum of the input vector and the connection weight, and then process it through the activation function), and finally obtain the historical detection result of each sample in the output layer according to the pre-set classification algorithm (such as Softmax function is used to convert the output into the probability value of various fault types). This historical detection result is a probability value vector of length K, which indicates the probability that the sample belongs to each type of grounding fault, or it can also be converted into a clear fault type classification label according to certain threshold rules.

[0094] S402, adjusting the model parameters of the fault recognition model trained in the historical period according to each historical detection result and the pre-constructed target loss function, and obtaining the fault recognition model after incremental learning.

[0095] Among them, the target loss function is a function specially constructed for optimizing the fault identification model. It is used to measure the difference between the model prediction results (i.e., historical detection results) and the true label (i.e., the actual ground fault type corresponding to each sample in the updated training data set). By quantifying this difference into a numerical value, it can guide the model in how to adjust the model parameters during the training process to reduce this difference, thereby improving the accuracy and performance of the model.

[0096] In the embodiment of the present application, first, the loss value of each sample is calculated according to the pre-constructed target loss function. Assume that the target loss function adopts the form of a cross entropy loss function combined with the maximum mean difference (MMD) as a penalty term. Finally, the cross entropy loss value and the value of the MMD penalty term are added to obtain the total loss value of each sample.

[0097] According to the calculated total loss value, the back propagation algorithm is used to calculate the gradient of the model parameters of each layer of the fault recognition model that has been trained in the historical period. For the model with a multi-layer perceptron (MLP) structure, starting from the output layer, the gradient of the connection weight between each node and the node in the next layer, as well as the weight gradient of each hidden layer neuron are calculated in sequence. According to the calculated gradient value, the model parameters of the fault recognition model that has been trained in the historical period are adjusted. By continuously repeating this update process, all model parameters of the model are adjusted to obtain the fault recognition model after incremental learning.

[0098] Another implementation method: Assume that the objective loss function adopts a form based on the mean square error (MSE) loss function and adds a custom penalty term for changes in data distribution. Calculate the gradient: According to the calculated total loss value, use the gradient-based optimization algorithm to calculate the gradient of the model parameters of each layer of the fault recognition model that has been trained in the historical period. For models built based on some principles of convolutional neural networks (CNN), start from the output layer and calculate the gradient of the connection weights between the output layer and the last layer of the hidden layer, as well as the weight gradients of the convolution kernels of each layer of the hidden layer. By repeating this update process, all model parameters of the model are adjusted to obtain a fault recognition model after incremental learning.

[0099] The embodiment of the present application obtains historical detection results by inputting the updated training data set into the fault recognition model that has been trained in the historical period, and adjusts the model parameters according to the target loss function, so that the model can more accurately learn the relationship between the fault characteristics in the new data and the ground fault type. The new data contains new situations that continue to appear during the operation of the distribution network. After such adjustments, the model can better adapt to these new situations, thereby improving the recognition accuracy of various ground fault types and reducing misjudgments and missed judgments.

[0100] In an exemplary embodiment, based on the above embodiment, see Figure 5 The embodiment of the present application relates to a process of determining an updated training data set, which includes the following steps:

[0101] S501, performing confidence assessment processing on zero-sequence current data to obtain a confidence assessment result of the zero-sequence current data.

[0102] In an embodiment of the present application, the data collected in real time by the data acquisition module is evaluated, and a human-computer interaction strategy is adopted to perform manual verification, storage and other operations on new data whose label confidence is lower than the verification threshold. For new data whose confidence is higher than the verification threshold, only new data whose confidence is lower than the storage threshold is stored.

[0103] The confidence of new data is calculated through the output result of the output layer of the neural network. Under the action of the Sigmoid activation function of the output layer, the output value is between 0 and 1, which can be interpreted as probability. The closer it is to 0 or 1, the higher the confidence of the model in the prediction result. Therefore, the calculation formula for the confidence of new data is applied, as shown in formula (1):

[0104] ………………………………(1)

[0105] S502: When the confidence evaluation result is greater than a preset verification threshold and the total number of zero-sequence current data is less than a preset storage threshold, determine an additional training set according to the zero-sequence current data.

[0106] In the embodiment of the present application, C represents the data confidence, and out represents the rated HIF prediction probability of the model output. When the detection confidence of the model for a new data is higher than both the verification threshold and the storage threshold, it can be considered that the model has full confidence in the detection result of the new data, and no additional learning is performed on the data. When the detection confidence of the model for a new data is higher than the verification threshold but lower than the storage threshold, it can be considered that the model still has good discrimination for the data. At the same time, in order to reduce the cost of manual verification, the model prediction label is used to perform additional learning and update on the data. When the detection confidence of the model for a new data is lower than the verification threshold, the model cannot reliably detect the data. At this time, a human-computer interaction strategy is introduced to update the data label, and the manually verified label is used to perform additional learning and update on the data.

[0107] S503: update the training data set according to the additional training set to obtain an updated training data set.

[0108] In the embodiment of the present application, the data that needs to be learned and updated is stored in the new training set. When the amount of HIF samples and non-HIF samples to be learned in the new training set reaches the set threshold, samples with the respective threshold data amounts are selected to be included in the model, and the remaining data enter the next incremental cycle. The setting of the two quantity thresholds is coordinated to avoid the potential impact of unbalanced data sets on the recognition model during the training process.

[0109] In the confidence evaluation processing step of the embodiment of the present application, various methods (such as linear regression model, energy-based evaluation, etc.) are used to evaluate the zero-sequence current data, and data with higher credibility can be screened out. This helps to exclude low-quality data that may be affected by noise interference, measurement errors, etc., thereby improving the data quality used to train the fault identification model and providing a more reliable basis for the model to accurately learn the characteristics of ground faults.

[0110] In an exemplary embodiment, based on the above embodiment, see Figure 6 The present application embodiment relates to a process of constructing a target loss function, which includes the following steps:

[0111] S601, construct a penalty term.

[0112] S602, determining a target loss function according to the penalty term and a pre-established cross entropy loss function.

[0113] The penalty term includes the corresponding relationship between the training data set, the updated training data set and the maximum mean difference.

[0114] In an embodiment of the present application, an indicator for measuring the difference in data distribution is embedded in the cross-entropy loss function as a penalty term to guide the incremental learning of the model so that the model can learn the same data features in the old data distribution and the new data distribution.

[0115] Based on the aforementioned HIF intelligent recognition model based on network weight regularization, the indicator that measures the difference in data distribution is embedded as a penalty term in the cross entropy loss function to guide the incremental learning of the model so that the model can learn the same data features in the old data distribution and the new data distribution. Maximum Mean Discrepancy (MMD) is an indicator that measures the distribution difference distance between two data sets. It can be used as a distance measurement method to measure the distribution difference between the new and old data sets. Therefore, the penalty term L and loss function based on MMD are They are expressed as formulas (2) and (3) respectively:

[0116] ………………(2)

[0117] ………………………………(3)

[0118] in, is the cross entropy loss function, is the penalty coefficient.

[0119] In the embodiment of the present application, the cross entropy loss function part in the target loss function can accurately measure the difference between the model prediction result and the true label, prompting the model to continuously adjust parameters during the training process to reduce this difference, thereby improving the model's recognition accuracy for different types of grounding faults.

[0120] In an exemplary embodiment, based on the above embodiment, see Figure 7 The embodiment of the present application relates to a process of adjusting the model parameters of the fault recognition model trained in the historical period according to each historical detection result and a pre-constructed target loss function to obtain the fault recognition model after incremental learning, including the following steps:

[0121] S701, performing importance evaluation processing on multiple neurons in the hidden layer network of the fault recognition model that has been trained in the historical period, and obtaining the second norms of multiple neuron weights.

[0122] In an embodiment of the present application, the importance of neurons in the original BP neural network is evaluated, and the parameters of important neurons are retained to maintain the parameters of these neurons unchanged during the subsequent incremental learning process. Existing models often encounter catastrophic forgetting of old knowledge when incrementally learning new knowledge. The main reason for catastrophic forgetting is that traditional models assume that the data distribution is fixed or stable, and the training samples can be obtained in full at the first time, and the model can fully learn the data features in the complete data set. However, if the data distribution of the new training set is significantly different from the data distribution of the original training set, the old model will be trained on the new training set. The new training set will change the model parameters and performance to a large extent, making it impossible for the model to retain the knowledge of the original training set, resulting in a decrease in the performance of the model in the old tasks. For this reason, consider regularizing the model parameters to achieve the purpose of consolidating the knowledge that the model has learned while learning new knowledge. For the BP neural network, it is composed of many interconnected neurons. Therefore, neurons can be regarded as the above-mentioned model parameters, and the neuron structure is as follows: Figure 8 shown.

[0123] Depend on Figure 8 It can be seen that the input-output relationship of neuron j is formula (4):

[0124] ………………(4)

[0125] Among them, in the formula, (i=1,2,3,…,n) is the input of neuron j, (i=1,2,3,…,n) is the weight of neuron j, is the threshold of neuron j, is the activation function, is the output of neuron j.

[0126] From formula (4), we can see that As the weight of neuron j, the larger its absolute value, the larger the corresponding component in the input vector. The greater the proportion of the neuron representation process, the higher the importance. In order to effectively evaluate the importance of neurons, the neuron weight binorm can be calculated by formula (5) and used to evaluate the importance of neurons.

[0127] S702, determining a plurality of important neurons according to the second norm of each neuron weight and a preset norm.

[0128] …………………………(5)

[0129] in, is the weighted norm of neuron j.

[0130] From formula (5), we can see that the larger the bi-norm of the neuron weight is, the more important the neuron is in the entire model decision process. Therefore, according to the bi-norm value of each neuron in each hidden layer, several important neurons in each hidden layer are selected and their parameters are saved, that is, the sensitivity of the model to historical data is retained.

[0131] S703, adjusting the model parameters of the fault recognition model trained in the historical period according to the neuron parameters of each important neuron and the target loss function, to obtain the fault recognition model after incremental learning.

[0132] The embodiment of the present application uses the back propagation algorithm to calculate the gradient of the model parameters of each layer of the fault recognition model that has been trained in the historical period according to the pre-constructed target loss function. For the hidden layer network, the focus is on the gradient corresponding to the neuron parameters of important neurons. Assuming that the target loss function adopts the form of a cross entropy loss function combined with the maximum mean difference (MMD) as a penalty term, for the historical detection results of each sample and the corresponding actual grounding fault type label, the cross entropy loss value is first calculated and then the maximum mean difference is calculated. The feature calculation of the sample under different data distributions and the application of related concepts such as the reproducing kernel Hilbert space are performed. According to this total loss value, the back propagation algorithm is used to start from the output layer, and the gradient of the connection weight between each node and the next layer of nodes, as well as the weight gradient of each hidden layer neuron are calculated in turn, focusing on the weight gradient calculation of important neurons. For the determined important neurons, their current neuron parameters are kept unchanged, that is, the weight parameters and bias parameters connected to other neurons are not changed. This is because these important neurons have learned relatively important knowledge in the previous training process, and retaining their parameters can help the model better retain the ability to recognize previous fault types. For the secondary neurons in the hidden layer network except for the important neurons, their neuron parameters are adjusted according to the calculated gradient values. By repeating this updating process, the model parameters of the secondary neurons are adjusted, so that the model can learn new data knowledge while ensuring the retention of important neuron knowledge, and obtain the fault recognition model after incremental learning.

[0133] After evaluating the importance of neurons in the hidden layer network and determining important neurons, the embodiment of the present application can enable the model to better learn the relationship between the ground fault characteristics and the ground fault type in the new data by reasonably processing the parameter adjustment of the important neurons and the secondary neurons. The important neurons retain the important knowledge learned in the previous training process, and the secondary neurons can be adjusted according to the new data, thereby improving the model's recognition accuracy for various types of ground faults and reducing misjudgments and missed judgments.

[0134] In an exemplary embodiment, based on the above embodiment, see Fig. 9 , the method of the embodiment of the present application also includes:

[0135] S801, obtaining multiple historical network layer weights in the hidden layer network of the fault recognition model trained in the historical period and multiple target network layer weights of the fault recognition model after incremental learning.

[0136] The embodiment of the present application uses an optimization constraint method to slow down the update speed of model weights during incremental learning.

[0137] In addition to retaining important neuron parameters, specific measures need to be taken to update other minor neurons in the network to fit new data. Assume that the network layer weights before and after model training are as shown in formula (6).

[0138] ………………(6)

[0139] In the formula, (i=1,2,3,…,m) and (i=1,2,3,…,m) represent the weights of the i-th neuron in the network layer before and after training, and m is the total number of neurons in the network layer.

[0140] S802, performing weighted second norm calculation processing on each historical network layer weight and each target network layer weight respectively, to obtain a plurality of historical weight second norms and a plurality of target weight second norms.

[0141] In the embodiment of the present application, each acquired historical network layer weight is calculated according to the weight second norm as calculated in formula (7), and such calculation is performed on all historical network layer weights to obtain multiple historical weight second norms.

[0142] Similarly, each target network layer weight is calculated according to the above weight two-norm calculation formula, and all target network layer weights are calculated in this way to obtain multiple target weight two-norms.

[0143] S803, input each historical weighted two-norm and each target weighted two-norm into the average value function to calculate the average value, and obtain the historical two-norm average value and the target two-norm average value.

[0144] In the embodiment of the present application, all historical weighted two norms are input into an average value function, as shown in formula (7), and the historical two norm average value is calculated by this formula. Similarly, all target weighted two norms are input into the average value function, and the target two norm average value is calculated according to the calculation formula of the above average value function.

[0145] S804, performing ratio processing on the historical two-norm average value and the target two-norm average value, and performing product processing on the ratio result and each target network layer weight to obtain the optimized target network layer weight.

[0146] In order to prevent the model from over-learning new data knowledge and ignoring old data knowledge, the embodiment of the present application should optimize the update of the network layer weight of the model to effectively control the degree of learning of the model for new data. The bi-norm of the new and old weights of the network layer is calculated and used as a proportion to adjust the new weight of the network layer to ease the update speed of the network layer weight. The specific optimization constraint method is shown in formula (7).

[0147] ………………(7)

[0148] in, is the average value, is the weight of the network layer after optimization constraints.

[0149] The embodiment of the present application compares the average values ​​of the two-norm values ​​of the hidden layer network weights of the fault recognition model trained in the historical period and the fault recognition model after incremental learning, and optimizes the target network layer weights according to their ratio, so that after the model learns new knowledge (incremental learning process), its hidden layer network weights maintain a certain degree of coordination with the historical model in terms of overall scale. This coordination helps prevent the model from experiencing excessively drastic weight changes under the impact of new data, thereby improving the stability of the model and enabling it to maintain relatively stable performance under different data inputs and fault scenarios.

[0150] In an exemplary embodiment, based on the above embodiment, see Fig.10 The embodiment of the present application relates to a process of extracting features from pre-acquired zero-sequence current data of a feeder line of a distribution network, and obtaining fault features includes the following steps:

[0151] S901, performing wavelet decomposition processing on zero-sequence current data of the feeder line of the distribution network to obtain detail coefficients of multiple levels.

[0152] The embodiment of the present application collects the zero-sequence current of the feeder line of the distribution network at a sampling frequency of 8kHz, with a length of eight cycles. The zero-sequence current signal is decomposed into 6 layers of wavelets by discrete wavelet transform (DWT), and the standard deviation of the detail coefficients from the first level to the sixth level is calculated, and the standard deviation of the 6 detail coefficients is used as the fault feature.

[0153] DWT obtains multi-level detail coefficients and approximate coefficients of waveform signals through multi-layer low-pass and high-pass filters, such as Fig.11As shown. Where LF represents a low-pass filter, HF represents a high-pass filter, cAi represents an approximation coefficient, which represents the large-scale low-frequency information of the waveform signal, and cDi represents a detail coefficient, which represents the small-scale high-frequency information of the waveform signal. Therefore, DWT can be used to decompose the waveform signal into multiple scales, and its expression is shown in (8):

[0154] …………(8)

[0155] Among them, X is the waveform signal, cDr is the detail coefficient of the rth layer, cAr is the approximation coefficient of the rth layer, and R is the total number of decomposition layers.

[0156] S902, performing standard deviation calculation processing on detail coefficients of each level to obtain fault characteristics.

[0157] In order to reduce the amount of redundant information from the original waveform or its transformed format, the embodiment of the present application needs to perform feature extraction to find unique parameters that can characterize the important features of the waveform. The standard deviation of the signal, as an indicator reflecting the degree of discreteness of the signal distribution, can provide information about the level of change in the signal frequency distribution, as shown in formula (9):

[0158] …………………………(9)

[0159] Where x is the detail coefficient and n is the number of elements in that coefficient.

[0160] Theoretically, the more decomposition levels of DWT, the more detailed the decomposed signal, and the more significant the extracted feature effect. However, considering that the sampling rate of the data collected in this paper is 8kHz, and the sampling theorem shows that the maximum frequency of the signal is 4kHz, then the 6-layer wavelet decomposition can obtain the standard deviation characteristics of the detail components representing the relatively high frequency bands of 2000~4000Hz and 1000~2000Hz, the characteristics representing the relatively medium frequency bands of 500~1000Hz and 250~500Hz, and the characteristics representing the relatively low frequency bands of 125~250Hz and 62.5~125Hz. HIF and its interference basically do not show any difference in the fundamental frequency, so there is no need to further decompose the zero-sequence current signal at the existing sampling rate, so the standard deviation of the detail coefficients from the first to the sixth level is used as the fault feature.

[0161] like Fig.12 As shown, the system power frequency is 50Hz, the sampling rate is 8kHz, and the parameters of the cable line and overhead line are shown in Table 1.

[0162] Table 1 Line parameters

[0163]

[0164] Table 2 describes the HIF or HIF interference events that occur at different fault locations and on different lines. Among them, the capacitor switching adopts the parallel three-phase capacitor model, the excitation inrush current is simulated by a single-phase transformer without load, the low-resistance grounding fault is simulated by a low-resistance model (5Ω-100Ω), the load switching adopts the three-phase asymmetric load model, and the HIF model is the Emanuel arc model. In addition, in order to conform to the actual engineering, the asynchronous closing of capacitor switching is added to the experiment to simulate the non-fault transient situation. In this study, the three-phase asynchronous closing means that phase A is connected to the system first, and phases B and C are connected to the system at the same time with the same delay. The initial fault angle is set to 0°, 30°, 60°, 90° and 120°.

[0165] Table 2 Line parameters

[0166]

[0167] In order to simulate the data flow scenario faced in the incremental learning application, 150 groups of samples in Table 3 were randomly selected as training sets, and the remaining data were randomly divided into simulated data streams and test sets in a ratio of 7:3. The model was incrementally learned through the data in the simulated data stream. This application measured and calculated the model incremental learning accuracy, and the model detection accuracy calculation formula used in the present invention is as follows.

[0168] …………………………(10)

[0169] In the formula, TP is the number of positive samples predicted as positive; TN is the number of positive samples predicted as negative; FP is the number of negative samples predicted as positive; FN is the number of negative samples predicted as negative; the total number of samples is TP+TN+FP+FN. The experimental results are shown in Table 3.

[0170] Table 3 Data flow scenario test results

[0171]

[0172] From the results in Table 3, we can see that when the training samples cannot be obtained immediately and completely, the model cannot fully learn all the knowledge of the current data, resulting in poor recognition ability of the model, and the recognition accuracy of the test set needs to be improved. Under the incremental learning framework, the model can adjust the training samples in real time according to the current environment for incremental learning. It can learn the knowledge of new data while resisting the forgetting of old data knowledge, so that the model can still accurately judge the data that was accurately recognized in the test set, that is, the old data, and accurately judge the data that was incorrectly recognized in the test set, that is, the new data. Therefore, the recognition ability of the model is continuously improved in the incremental learning process. It can also be seen from the indicators of model label accuracy and human-computer collaboration accuracy that based on the incremental learning framework, the model can select more representative training samples for model update according to the current scenario, and the HIF recognition ability is gradually improved. The human-computer collaboration method plays a major role in guiding the evolution of the model in the early stage, so that the model evolves in the right direction, and plays an auxiliary role in the evolution of the model in the later stage. With the improvement of the model recognition performance, the model inevitably assigns high-confidence erroneous labels to individual data in the simulated data stream, resulting in the inability of the human-computer collaboration method to correct the labels, but it does not affect the overall recognition performance of the model, and has a certain fault tolerance. In summary, under the incremental learning framework, the BP neural network based on the regularized incremental learning paradigm has good robustness and anti-forgetting ability for old data.

[0173] In an exemplary embodiment, based on the above embodiment, the method of the embodiment of the present application further includes:

[0174] Step 1: Construct a penalty term; determine the target loss function based on the penalty term and the pre-established cross entropy loss function;

[0175] Step 2: Acquire zero-sequence current data; perform confidence assessment processing on the zero-sequence current data to obtain a confidence assessment result of the zero-sequence current data; when the confidence assessment result is greater than a preset verification threshold and the total number of zero-sequence current data is less than a preset storage threshold, determine an additional training set based on the zero-sequence current data; update the training data set based on the additional training set to obtain an updated training data set.

[0176] Step 3: Input the updated training data set into the fault identification model trained in the historical period to perform ground fault detection processing, and obtain multiple historical detection results;

[0177] Step 4: Perform importance evaluation on multiple neurons in the hidden layer network of the fault recognition model trained in the historical period to obtain multiple neuron weights and second norms; determine multiple important neurons according to the second norm of each neuron weight and the preset norm; adjust the model parameters of the fault recognition model trained in the historical period according to the neuron parameters of each important neuron and the target loss function to obtain the fault recognition model after incremental learning;

[0178] Step 5: Perform wavelet decomposition on the zero-sequence current data of the distribution network feeder line to obtain detail coefficients of multiple levels; perform standard deviation calculation on the detail coefficients of each level to obtain fault characteristics;

[0179] Step 6: Input the fault feature into the input layer of the fault recognition model after incremental learning for data conversion processing to obtain a fault vector; input the fault vector into the hidden layer network of the fault recognition model after incremental learning for multi-layer nonlinear transformation processing to obtain a nonlinear transformation result; input the nonlinear transformation result into the output layer of the fault recognition model after incremental learning for classification decision processing to obtain a target detection result.

[0180] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0181] Based on the same inventive concept, the embodiment of the present application also provides a ground fault detection device for implementing the ground fault detection method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more ground fault detection device embodiments provided below can refer to the limitations of the ground fault detection method above, and will not be repeated here.

[0182] In one embodiment, Fig.13 As shown, a ground fault detection device 1000 is provided, comprising:

[0183] The feature extraction module 1001 is used to perform feature extraction processing on the pre-acquired zero-sequence current data of the feeder line of the distribution network to obtain fault features;

[0184] The fault detection module 1002 is used to input the fault characteristics into a pre-built fault identification model to perform ground fault detection processing and obtain a target detection result; the model parameters of the fault identification model are based on the zero-sequence current data to update the training data set, and are dynamically adjusted by incremental learning according to the updated training data set; the target detection result is used to determine the ground fault type corresponding to the fault characteristics.

[0185] In one embodiment, the fault identification model includes an input layer, a hidden layer network and an output layer; the fault detection module includes:

[0186] A vector determination unit, used for inputting the fault feature into the input layer for data conversion processing to obtain a fault vector;

[0187] A nonlinear transformation unit is used to input the fault vector into the hidden layer network for multi-layer nonlinear transformation processing to obtain a nonlinear transformation result;

[0188] The decision unit is used to input the nonlinear transformation results into the output layer for classification decision processing to obtain the target detection results.

[0189] In one embodiment, the above device further comprises:

[0190] A historical detection module, used to input the updated training data set into the fault recognition model trained in the historical period to perform ground fault detection processing, and obtain multiple historical detection results;

[0191] The parameter adjustment module is used to adjust the model parameters of the fault recognition model trained in the historical period according to the historical detection results and the pre-built target loss function to obtain the fault recognition model after incremental learning.

[0192] In one embodiment, the above device further comprises:

[0193] A confidence evaluation module is used to perform confidence evaluation processing on zero-sequence current data to obtain a confidence evaluation result of the zero-sequence current data;

[0194] An additional training set determination module is used to determine an additional training set according to the zero-sequence current data when the confidence evaluation result is greater than a preset verification threshold and the total number of zero-sequence current data is less than a preset storage threshold;

[0195] The training set determination module is used to update the training data set according to the additional training set to obtain an updated training data set.

[0196] In one embodiment, the above device further comprises:

[0197] A penalty item determination module is used to construct a penalty item; the penalty item includes a corresponding relationship between a training data set, an updated training data set and a maximum mean difference;

[0198] The objective function determination module is used to determine the objective loss function based on the penalty term and the pre-established cross entropy loss function.

[0199] In one embodiment, the parameter adjustment module according to each historical detection result and the pre-constructed target loss function includes:

[0200] A binary norm determination unit is used to perform importance evaluation processing on multiple neurons in a hidden layer network of a fault recognition model that has been trained in a historical period to obtain binary norms of multiple neuron weights;

[0201] An important neuron determination unit, used for determining a plurality of important neurons according to the second norm of each neuron weight and a preset norm;

[0202] The target model determination unit is used to adjust the model parameters of the fault recognition model trained in the historical period according to the neuron parameters of each important neuron and the target loss function to obtain the fault recognition model after incremental learning.

[0203] In one embodiment, the above device further comprises:

[0204] A weight acquisition module, used to acquire multiple historical network layer weights in the hidden layer network of the fault recognition model trained in the historical period and multiple target network layer weights of the fault recognition model after incremental learning;

[0205] A two-norm calculation module is used to perform weight two-norm calculation processing on each historical network layer weight and each target network layer weight, respectively, to obtain multiple historical weight two-norms and multiple target weight two-norms;

[0206] The mean calculation module is used to input each historical weighted two-norm and each target weighted two-norm into the average value function to perform average value calculation, and obtain the historical two-norm average value and the target two-norm average value;

[0207] The ratio calculation module is used to perform ratio processing on the historical two-norm average value and the target two-norm average value, and multiply the ratio result by the weight of each target network layer to obtain the optimized target network layer weight.

[0208] In one embodiment, the fault detection module includes:

[0209] A wavelet decomposition unit is used to perform wavelet decomposition processing on zero-sequence current data of the feeder line of the distribution network to obtain detail coefficients of multiple levels;

[0210] The feature determination unit is used to calculate the standard deviation of the detail coefficients of each level to obtain the fault feature.

[0211] Each module in the above-mentioned ground fault detection device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0212] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.14 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a ground fault detection method is implemented.

[0213] Those skilled in the art will understand that Fig.14 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0214] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0215] Perform feature extraction processing on the zero-sequence current data of the feeder line of the distribution network acquired in advance to obtain fault features;

[0216] The fault characteristics are input into a pre-built fault identification model for ground fault detection processing to obtain a target detection result; the model parameters of the fault identification model are updated based on the zero-sequence current data training data set, and are dynamically adjusted by incremental learning according to the updated training data set; the target detection result is used to determine the ground fault type corresponding to the fault characteristics.

[0217] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0218] Input the fault features into the input layer for data conversion processing to obtain the fault vector;

[0219] Input the fault vector into the hidden layer network for multi-layer nonlinear transformation processing to obtain a nonlinear transformation result;

[0220] The nonlinear transformation results are input into the output layer for classification decision processing to obtain the target detection results.

[0221] In one embodiment, the method further comprises:

[0222] The updated training data set is input into the fault identification model trained in the historical period to perform ground fault detection processing, and multiple historical detection results are obtained;

[0223] According to each historical detection result and the pre-constructed target loss function, the model parameters of the fault recognition model trained in the historical period are adjusted to obtain the fault recognition model after incremental learning.

[0224] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0225] Performing confidence assessment processing on the zero-sequence current data to obtain a confidence assessment result of the zero-sequence current data;

[0226] When the confidence evaluation result is greater than a preset verification threshold and the total number of zero-sequence current data is less than a preset storage threshold, determining an additional training set according to the zero-sequence current data;

[0227] The training data set is updated according to the additional training set to obtain an updated training data set.

[0228] In one embodiment, the process of constructing the above objective loss function includes:

[0229] Construct a penalty term; the penalty term includes the corresponding relationship between the training data set, the updated training data set and the maximum mean difference;

[0230] Based on the penalty term and the pre-established cross entropy loss function, the target loss function is determined.

[0231] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0232] Performing importance evaluation processing on multiple neurons in the hidden layer network of the fault recognition model that has been trained in the historical period, and obtaining the second norms of multiple neuron weights;

[0233] Determine multiple important neurons according to the second norm of each neuron weight and the preset norm;

[0234] According to the neuron parameters of each important neuron and the target loss function, the model parameters of the fault recognition model trained in the historical period are adjusted to obtain the fault recognition model after incremental learning.

[0235] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0236] Obtain multiple historical network layer weights in the hidden layer network of the fault recognition model trained in the historical period and multiple target network layer weights of the fault recognition model after incremental learning;

[0237] The weighted two-norm calculation process is performed on each historical network layer weight and each target network layer weight respectively, so as to obtain a plurality of historical weighted two-norms and a plurality of target weighted two-norms;

[0238] Input each historical weighted two-norm and each target weighted two-norm into the average value function to calculate the average value, and obtain the historical two-norm average value and the target two-norm average value;

[0239] The historical two-norm average value and the target two-norm average value are ratio processed, and the ratio result is multiplied by the weight of each target network layer to obtain the optimized target network layer weight.

[0240] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0241] The zero-sequence current data of the distribution network feeder lines are processed by wavelet decomposition to obtain multiple levels of detail coefficients;

[0242] The standard deviation of detail coefficients at each level is calculated to obtain fault characteristics.

[0243] According to some embodiments of the present application, a computer program product is also provided, and when the computer program is executed by a processor, the above method can be implemented. The computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, part or all of the above method can be implemented in whole or in part according to the process or function described in the embodiment of the present application.

[0244] According to some embodiments of the present application, there is also provided a non-transitory computer-readable storage medium including instructions, such as a memory including instructions, and the above instructions can be executed by a processor of an electronic device to complete the above method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0245] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0246] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0247] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0248] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A ground fault detection method, characterized in that: The method comprises: Perform feature extraction processing on the zero-sequence current data of the feeder line of the distribution network acquired in advance to obtain fault features; The fault feature is input into a pre-built fault identification model for ground fault detection processing to obtain a target detection result; the model parameters of the fault identification model are updated based on the zero-sequence current data training data set, and are dynamically adjusted by incremental learning according to the updated training data set; the target detection result is used to determine the ground fault type corresponding to the fault feature.

2. The method according to claim 1, characterized in that The fault identification model includes an input layer, a hidden layer network and an output layer; the fault characteristics are input into the pre-built fault identification model to perform ground fault detection processing to obtain a target detection result, including: Inputting the fault feature into the input layer for data conversion processing to obtain a fault vector; Inputting the fault vector into the hidden layer network for multi-layer nonlinear transformation processing to obtain a nonlinear transformation result; The nonlinear transformation result is input into the output layer for classification decision processing to obtain the target detection result.

3. The method according to claim 1, characterized in that The method further comprises: Inputting the updated training data set into the fault identification model trained in the historical period to perform ground fault detection processing to obtain multiple historical detection results; According to each of the historical detection results and the pre-constructed target loss function, the model parameters of the fault recognition model that has been trained in the historical period are adjusted to obtain a fault recognition model after incremental learning.

4. The method according to claim 3, characterized in that The process of determining the updated training data set includes: Performing confidence assessment processing on the zero-sequence current data to obtain a confidence assessment result of the zero-sequence current data; In a case where the confidence evaluation result is greater than a preset verification threshold and the total number of the zero-sequence current data is less than a preset storage threshold, determining an additional training set according to the zero-sequence current data; The training data set is updated according to the additional training set to obtain the updated training data set.

5. The method according to claim 4, characterized in that The process of constructing the target loss function includes: Constructing a penalty term; the penalty term includes a corresponding relationship between the training data set, the updated training data set and the maximum mean difference; The target loss function is determined according to the penalty term and a pre-established cross entropy loss function.

6. The method according to claim 5, characterized in that The adjusting process of the model parameters of the fault identification model trained in the historical period according to each of the historical detection results and the pre-constructed target loss function to obtain the fault identification model after incremental learning includes: Performing importance evaluation processing on multiple neurons in the hidden layer network of the fault recognition model trained in the historical period to obtain multiple neuron weight bi-norms; Determining a plurality of important neurons according to the second norm of each neuron weight and a preset norm; The model parameters of the fault recognition model trained in the historical period are adjusted according to the neuron parameters of each of the important neurons and the target loss function to obtain the fault recognition model after incremental learning.

7. The method according to claim 6, characterized in that The method further comprises: Acquire multiple historical network layer weights in the hidden layer network of the fault recognition model trained in the historical period and multiple target network layer weights of the fault recognition model after incremental learning; Performing weighted second norm calculation processing on each of the historical network layer weights and each of the target network layer weights respectively to obtain a plurality of historical weighted second norms and a plurality of target weighted second norms; Inputting each of the historical weighted two norms and each of the target weighted two norms into the average value function to calculate the average value, thereby obtaining the historical two norm average value and the target two norm average value; The historical two-norm average value and the target two-norm average value are ratio-processed, and the ratio result and each target network layer weight are multiplied to obtain the optimized target network layer weight.

8. The method according to any one of claims 1 to 7, characterized in that: The feature extraction process is performed on the pre-acquired zero-sequence current data of the distribution network feeder line to obtain the fault features, including: Performing wavelet decomposition processing on zero-sequence current data of the feeder line of the distribution network to obtain detail coefficients of multiple levels; The standard deviation of the detail coefficients of each level is calculated to obtain the fault characteristics.

9. A ground fault detection device, characterized in that: The device comprises: A feature extraction module is used to perform feature extraction processing on the pre-acquired zero-sequence current data of the distribution network feeder line to obtain fault features; A fault detection module is used to input the fault characteristics into a pre-built fault identification model to perform ground fault detection processing and obtain a target detection result; the model parameters of the fault identification model are based on updating the training data set based on zero-sequence current data, and are dynamically adjusted by incremental learning according to the updated training data set; the target detection result is used to determine the ground fault type corresponding to the fault characteristics.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images