Distributed fault detection system and method for power distribution control equipment

Abstract

The invention discloses a distributed fault detection system and method for power distribution control equipment, and belongs to the technical field of fault detection, and the method comprises the steps: collecting historical and real-time power operation data at a plurality of control nodes of a power distribution system; a current zero-sequence component change rate and an instantaneous harmonic phase disturbance index are extracted, a low-frequency high-risk sample identification function is constructed, and samples are scored; a sample feature set is constructed based on the scoring result, and a model is trained and generated and used for learning potential distribution of rare fault samples; simulated rare fault samples are generated and injected into the original training set, and an enhanced training set is constructed; training a decision tree model by using an enhanced training set, constructing a fault classification structure based on information gain, and deploying the fault classification structure at each control node to realize distributed recognition of rare faults; according to the method, the identification capability of the power distribution system on low-frequency and high-risk faults such as high-resistance grounding and arc disturbance is improved, and the comprehensiveness of fault classification and the safety of system operation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a distributed fault detection system and method for power distribution control equipment. Background Art

[0002] Distributed fault detection for distribution control equipment involves deploying intelligent modules with detection capabilities across multiple device nodes within the distribution system to achieve collaborative sensing, real-time analysis, and localized fault detection. Compared to centralized detection, distributed detection improves the system's fault response speed and accuracy, reduces information transmission latency, and enhances the system's self-healing capabilities in the face of localized faults or communication anomalies. This approach is particularly well-suited for the complex, multi-source, multi-node distribution network structures found in modern smart grids.

[0003] The existing technology has the following shortcomings:

[0004] In distribution network fault detection, high-risk faults like high-resistance ground faults occur very infrequently, leading to a severe shortage of training data samples. As a result, existing fault detection models, such as decision tree models, tend to favor common fault types when constructing classification paths, neglecting to model rare but potentially more damaging events, creating a fault blind spot. This data imbalance prevents the model from identifying critical anomalies, creating the risk of underreporting and directly threatening the safety of grid operations and system reliability. Summary of the Invention

[0005] The purpose of the present invention is to provide a distributed fault detection system and method for power distribution control equipment to address the deficiencies in the background technology.

[0006] In order to achieve the above objectives, the present invention provides the following technical solution: a distributed fault detection method for power distribution control equipment, comprising:

[0007] Collect historical and real-time power operation data at several control nodes in the power distribution system;

[0008] Extract low-frequency, high-risk fault samples from the collected data and construct a sample feature set. This includes obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, building a low-frequency, high-risk sample identification function, and performing sample scoring.

[0009] Training a generative model based on the sample feature set, wherein the generative model includes a variational autoencoder or a generative adversarial network for learning the potential feature distribution of rare fault samples;

[0010] Generate simulated rare fault samples using the generative model and inject them into the original training sample set to construct an enhanced training set;

[0011] The enhanced training set is trained to build a decision tree model based on information gain, forming a classification structure of fault classification paths and realizing distributed fault type recognition and classification.

[0012] Preferably, the collection of historical and real-time power operation data at several control nodes of the power distribution system includes: configuring power data collection devices at the feeder terminal FTU, switch station and intelligent ring network cabinet control nodes; collecting current, voltage, frequency and harmonic parameters in real time, and attaching timestamps and node position codes; the current data includes three-phase current signal I A (t), I B (t), I C (t), sampled at intervals of 20ms or less; the collected data is initially archived and denoised by the edge computing unit before being uploaded to the central training platform.

[0013] Preferably, by obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, a low-frequency high-risk sample identification function is constructed to perform sample scoring, specifically including:

[0014] Collect three-phase current I A (t), I B (t), I C (t), the zero-sequence component I0(t) is calculated by Clarke transform, that is: Then perform first-order difference or differential calculation on I0(t) to obtain its rate of change per unit time, which is expressed as: ΔI0(t) = |I0(t)-I0(t-Δt)|; I0(t) represents the zero-sequence current component, which measures the asymmetry of the three-phase current; ΔI0(t) represents the rate of change of the zero-sequence component, which indicates the degree of abnormal fluctuation; Δt represents the sampling time interval;

[0015] Perform fast Fourier transform on the collected voltage or current signal to extract the phase value φ of the nth harmonic n (t), where n∈{3,5,7}; calculate the phase mutation of each harmonic between adjacent sampling periods: Δφ n (t)=|φ n (t)-φ n (t-Δt)|; then the disturbance values ​​of all harmonics are weighted averaged to obtain the instantaneous harmonic phase disturbance index HPI.

[0016] Preferably, the zero-sequence component change rate and the instantaneous harmonic phase disturbance index are combined to construct a scoring function for identifying abnormal samples: R(t) = α·ΔI0(t) + β·HPI(t); wherein α and β are adjustment weights, and a scoring threshold Rth is set. When R(t)>Rth, the sample is judged to be a low-frequency high-risk fault sample and enters the training sample set for the subsequent generation model.

[0017] Preferably, the training generation model specifically includes:

[0018] Take the scored and labeled high-risk samples as input to train a variational autoencoder or a generative adversarial network;

[0019] The VAE model uses a multi-scale residual convolution structure to extract input features and introduces an attention mechanism to adaptively allocate feature weights.

[0020] The discriminator of the GAN model has multi-tasking capabilities. In addition to judging the authenticity of samples, it also outputs sample category labels and anomaly scores.

[0021] During the training process, the reconstruction error and KL divergence are optimized to enhance the stability of the generated samples.

[0022] Preferably, generating simulated rare fault samples and injecting them into the training set includes:

[0023] Use the trained generative model to sample latent variables z~N(0,1) from the latent space to generate simulated samples;

[0024] Apply the isolation forest algorithm to the generated samples to remove outliers;

[0025] Match the categories of the screened simulated samples with the original training samples, and use genetic algorithms to optimize the injection ratio of each category;

[0026] The injection weight is dynamically adjusted according to the category scarcity to construct an enhanced training set.

[0027] Preferably, the construction of a decision tree model based on information gain includes:

[0028] Input the features and class labels in the enhanced training set into the decision tree algorithm;

[0029] Using information gain as the node division criterion, recursively select the optimal features to construct the branch structure;

[0030] When the number of leaf node samples is insufficient or the categories are impure, pruning operations are triggered, including pre-pruning and post-pruning strategies;

[0031] After training is completed, the model structure is embedded into each control node terminal to achieve distributed fault judgment and synchronous reporting.

[0032] The present invention also provides a distributed fault detection system for power distribution control equipment, comprising a data acquisition module, a sample scoring module, a generation model training module, a sample enhancement and management module, and a distributed recognition module;

[0033] Data acquisition module: collects historical and real-time power operation data at several control nodes of the power distribution system;

[0034] Sample scoring module: Extracts low-frequency, high-risk fault samples from the collected data and constructs a sample feature set. This includes obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, building a low-frequency, high-risk sample identification function, and performing sample scoring.

[0035] Generative model training module: training a generative model based on the sample feature set, wherein the generative model includes a variational autoencoder or a generative adversarial network, and is used to learn the potential feature distribution of rare fault samples;

[0036] Sample enhancement and management module: uses the generative model to generate simulated rare fault samples and injects them into the original training sample set to construct an enhanced training set;

[0037] Distributed identification module: Trains the enhanced training set, builds a decision tree model based on information gain, forms a classification structure of the fault classification path, and realizes distributed fault type recognition and classification.

[0038] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0039] 1. This invention introduces deep generative models such as variational autoencoders and generative adversarial networks to model the underlying characteristic distribution of low-frequency, high-risk fault samples and expand the sample size. This overcomes the problem of sample scarcity leading to blind spots in traditional power distribution fault detection. Compared to existing classification methods that rely solely on real data for training, this invention can generate simulated samples with realistic structures and balanced categories, significantly improving the decision tree model's ability and accuracy in identifying hidden faults such as high-resistance grounding and arc discharge.

[0040] 2. This invention utilizes an information-gain-driven fault classification path construction approach, combined with feature scoring, abnormal sample screening, multi-scale feature extraction, and an adaptive injection strategy, to enhance the model's generalization performance and engineering deployability. By embedding the trained model into each distribution control node, it achieves truly distributed intelligent fault detection with advantages such as rapid response, comprehensive identification, and structural interpretability, significantly enhancing the operational safety and fault self-healing capabilities of the distribution system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0042] Figure 1 Flow chart of the method of the present invention.

[0043] Figure 2 It is a flow chart of the system modules of the present invention. DETAILED DESCRIPTION

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0045] Example 1, please refer to Figure 1 As shown, the distributed fault detection method for power distribution control equipment described in this embodiment includes:

[0046] Collect historical and real-time power operation data at several control nodes in the power distribution system;

[0047] Extract low-frequency, high-risk fault samples from the collected data and construct a sample feature set. This includes obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, building a low-frequency, high-risk sample identification function, and performing sample scoring.

[0048] Training a generative model based on the sample feature set, wherein the generative model includes a variational autoencoder or a generative adversarial network for learning the potential feature distribution of rare fault samples;

[0049] Generate simulated rare fault samples using the generative model and inject them into the original training sample set to construct an enhanced training set;

[0050] The enhanced training set is trained to build a decision tree model based on information gain, forming a classification structure of fault classification paths and realizing distributed fault type recognition and classification.

[0051] Smart terminal devices with data collection and edge computing capabilities are deployed at multiple control nodes of the power distribution system (such as switch stations, distributed feeder terminals (FTUs), smart switches, ring main units, and cable branch boxes) to collect and upload the following power operation parameters in real time:

[0052] Voltage data acquisition: including three-phase voltage RMS, instantaneous voltage waveform, and phase-to-phase voltage; can be acquired in real time through voltage transformers (PTs) or voltage sensors; used to identify abnormal phenomena such as voltage drops, short-term overvoltages, and ground imbalance.

[0053] Current data acquisition: including three-phase current RMS, instantaneous current waveform, and current change rate; uses current transformers (CTs) or contactless current sensors; can identify events such as short-circuit current mutations, load surges, and line current interruptions.

[0054] Frequency measurement: Real-time acquisition of system frequency and frequency offset rate. Small frequency fluctuations often indicate power imbalance or heavy load switching. This is particularly important in island protection and distributed power source identification.

[0055] Harmonic and distortion information: Extracts harmonic components (such as 3rd, 5th, and 7th harmonics) from current and voltage; uses FFT or wavelet transform algorithms to calculate harmonic distortion (THD) online; can be used to identify harmonic pollution, nonlinear load anomalies, and arc faults.

[0056] All collected data is equipped with GPS or IEEE 1588 standard timestamps; synchronous collection ensures that data from different nodes have temporal correlation; the collected data of each node is packaged with a geographic location identifier to support geographically distributed model training.

[0057] Extract low-frequency, high-risk fault samples from the collected data and construct a sample feature set. This includes obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, building a low-frequency, high-risk sample identification function, and performing sample scoring. Specifically, this includes:

[0058] Collect three-phase current I A (t), I B (t), I C (t), the zero-sequence component I0(t) is calculated by Clarke transform, that is: Performing a first-order difference or differential calculation on I0(t) yields its rate of change per unit time, expressed as: ΔI0(t) = |I0(t) - I0(t - Δt)|. This metric reflects the fluctuation rate of the asymmetric current component in the system and is particularly responsive to weak current disturbances caused by high-resistance grounding or poor contact. I0(t) represents the zero-sequence current component, measuring the asymmetry of the three-phase currents; ΔI0(t) represents the rate of change of the zero-sequence component, indicating the degree of abnormal fluctuation; and Δt represents the sampling interval, typically 20ms or finer.

[0059] Perform fast Fourier transform (FFT) on the collected voltage or current signal to extract the phase value φ of the nth harmonic n (t), where n∈{3,5,7} and other common odd harmonics. Calculate the phase mutation of each harmonic between adjacent sampling periods: Δφ n (t)=|φ n (t)-φ n(t-Δt)|; then the disturbance values ​​of all harmonics are weighted averaged to obtain the instantaneous harmonic phase disturbance index HPI.

[0060] The two characteristic indicators above are combined to construct a scoring function for identifying abnormal samples: R(t) = α·ΔI0(t) + β·HPI(t); where α and β are adjustment weights, configured based on different system characteristics. The system sets a scoring threshold, Rth. When R(t) > Rth, the sample is identified as a low-frequency, high-risk fault sample and is included in the training sample set for the subsequent generation model. Rth is the scoring threshold, which can be derived from historical fault sample statistics.

[0061] Through the above-mentioned feature extraction and scoring mechanism, the present invention can effectively identify low-frequency anomalies that are difficult to cover by traditional decision tree models. It is particularly suitable for fault scenarios with weak signal characteristics and unstable time distribution, such as high-resistance grounding and arc breakdown, and provides high-quality rare sample support for subsequent fault generation sample enhancement and classification model optimization.

[0062] In the present invention, in order to improve the detection coverage capability of rare fault types (such as high-resistance grounding, transient arc, etc.), a rare sample enhancement strategy based on a deep generative model is adopted to learn the potential distribution from the labeled low-frequency and high-risk fault sample feature set, and generate pseudo-fault samples with similar structure for use by subsequent decision tree classifiers.

[0063] The generative model can use a variational autoencoder (VAE) or a generative adversarial network (GAN). The training process is as follows:

[0064] Sample generation based on variational autoencoder (VAE) includes:

[0065] The characteristic parameters extracted from historical and real-time acquisition (such as the current zero-sequence component change rate ΔI0 and the harmonic phase disturbance index HPI) are composed of sample feature vectors as the input of the VAE model.

[0066] The encoder uses a multi-scale residual convolution structure to enhance its ability to extract weak fault features. Specifically, it includes:

[0067] Three convolution channels are connected in parallel, with convolution kernel sizes of 3×1, 5×1, and 7×1 respectively;

[0068] Each channel is connected to the ReLU activation function and residual connection;

[0069] Multi-scale features are fused and mapped to the latent space z~N(0,1).

[0070] The attention layer is introduced during the latent vector generation process. This dynamically adjusts the priority of input features in each dimension, improving the model's sensitivity to key features. The attention coefficient is calculated using the softmax function.

[0071] The decoder is a multi-layer fully connected network that reconstructs the latent vector z into feature samples from the original input space. During training, a joint loss function is optimized using the reconstruction error and the KL divergence.

[0072] After the training is completed, z~N(0,1) is randomly sampled, and a new rare fault sample feature vector is generated through the decoder for use by subsequent models.

[0073] Sample generation based on generative adversarial networks (GANs) includes:

[0074] The generator structure uses a latent variable input + full connection + one-dimensional convolution module, and outputs a feature vector with the same dimension as the original sample. The input latent variable is also sampled from a standard normal distribution.

[0075] The discriminator not only determines the authenticity of the sample, but also outputs it through a parallel sub-network:

[0076] The category label of the sample (e.g., “high-resistance grounding”, “transient arc”);

[0077] Sample anomaly intensity score;

[0078] Adversarial probability (real or fake).

[0079] The multi-task loss function of the discriminator can be expressed as: in, To combat losses, is the classification loss, is the anomaly scoring loss. To avoid instability in discriminator training, a gradient penalty term is added to the loss function to control the gradient norm close to 1, thereby improving generation quality and network convergence.

[0080] To further ensure the quality of generated samples, the Isolation Forest algorithm is used to perform outlier detection on the generated samples, eliminate potential false samples, and only retain fault data with stable structure and similarity to real features.

[0081] Through the training and enhancement process of the above-mentioned generative model, it is possible to automatically synthesize diverse and controllable high-risk fault samples when rare samples are insufficient, effectively improving the decision tree classifier's ability to identify abnormal working conditions, especially in multi-source complex power distribution scenarios, to improve system reliability and fault visibility.

[0082] In this paper, a trained generative model (such as a variational autoencoder (VAE) or a generative adversarial network (GAN)) is used to sample from the latent variable space z to N (0, 1) and decode it to generate a set of simulated samples with the characteristics of real rare faults. Each generated sample has the same feature dimension structure as the original sample and is classified and stored according to the fault type.

[0083] To construct an enhanced training set, the system mixes the generated samples with the original training samples in a fixed ratio, such as a 1:1, 2:1, or rare class weighted ratio scheme (see subsequent examples). This enhanced training set is used as input for subsequent decision tree model training, which can effectively improve the model's recognition ability on rare classes.

[0084] Considering that the generated model may output structural anomalies or noise samples, resulting in distortion of model training, a density-based unsupervised clustering algorithm, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), is introduced in this embodiment to screen the generated samples.

[0085] The specific process includes:

[0086] Extract the main feature dimensions of all generated samples;

[0087] Perform DBSCAN clustering on the samples, set the minimum number of neighbor points MinPts and the distance threshold ε;

[0088] In the clustering results, the “core samples” and “boundary samples” are retained, and the abnormal samples identified as “noise points” are eliminated.

[0089] The filtered sample set is more representative, reducing the risk of the model being interfered with by pseudo samples during the training phase.

[0090] In order to achieve the optimal data injection ratio configuration between different categories, this embodiment uses a genetic algorithm to perform global optimization on the simulated sample injection ratio. The specific steps are as follows:

[0091] Construct the initial population, where each individual is encoded as a set of injection ratio vectors r = [r1, r2, ..., rn], which represents the injection ratio of generated samples of each category;

[0092] Define the fitness function: F(r) = p·Acc+q·Recall rare -γ·C; where:

[0093] Acc represents the overall accuracy of the model;

[0094] Recall rare represents the recall rate of rare categories;

[0095] C represents the model complexity (such as tree depth or number of nodes);

[0096] p, q, and γ represent weight factors.

[0097] Iteratively perform operations such as selection, crossover, and mutation, and eventually converge to the global optimal proportion configuration.

[0098] This approach ensures that the generated sample injection strategy is data-driven and objective.

[0099] In view of the data scarcity of different fault categories, this embodiment introduces a “category scarcity weight” to dynamically adjust the injection ratio of generated samples of each category.

[0100] The specific calculation method is as follows:

[0101] For each type of fault, calculate the number of original samples N real and the number of generated samples N gen ;

[0102] Define category scarcity parameters:

[0103] Where c represents the category number, ∈ is a small constant to avoid division by zero;

[0104] According to ρ c The normalized result redistributes the generated sample injection ratio, so that the high-scarcity categories are preferentially filled with more pseudo samples. This method dynamically controls the density of injected samples and can effectively alleviate the problem of class imbalance.

[0105] To further improve the "pseudo-realism" of injected samples, this embodiment introduces an adversarial sample fusion ranking mechanism to score and screen the generated samples for realism. The specific process is as follows:

[0106] Build a lightweight two-class adversarial discriminator D, with the original sample and the generated sample as input;

[0107] The discriminator outputs the true probability of the sample Preal(x)∈[0,1]. The closer the value is to 1, the closer it is to the original sample distribution.

[0108] All generated samples are ranked by score, and the top k% pseudo samples are retained for injection into the training set.

[0109] Through this mechanism, generated samples with high degree of forgery and abnormal distribution are effectively filtered out, further improving the training value of the enhanced data and the stability of the model.

[0110] The present invention constructs a decision tree classification model with strong rare fault identification capabilities based on the generated and screened enhanced training samples. The process specifically includes the following steps:

[0111] All samples in the augmented training set are standardized with unified labels, including common fault types (such as short circuit, overload, and ground fault); rare fault types (such as high-resistance ground fault, arc discharge, and nonlinear disturbance); and normal operation status (as a control category). This labeling system is used for supervised learning modeling.

[0112] Key features are extracted from the sample set and fed into the model. These typically include, but are not limited to, the rate of change of the zero-sequence current component; the instantaneous harmonic phase disturbance index; voltage sag amplitude, frequency fluctuation rate, and harmonic distortion (THD). Redundant features are eliminated based on feature importance evaluation (e.g., information gain). All features are normalized or standardized to eliminate dimensional differences.

[0113] Using information gain as the node division criterion, the decision tree classification structure is iteratively constructed: Select the root node feature: Calculate the information gain of all features and select the largest one as the root node; Recursively divide the child nodes: Repeat the above process on the child data set until: the sample purity reaches the set threshold; the tree depth reaches the preset maximum depth; the number of samples is less than the splitting lower limit; Leaf node classification results: Each leaf node corresponds to a specific fault category or normal state.

[0114] To avoid overfitting of the model, the following pruning strategy is implemented:

[0115] Pre-pruning: Determine the split benefit before the node splits, and stop if it is below the threshold; Post-pruning: Simplify the structure based on the validation set error and merge subtrees with small error changes; Regularization control: Introduce tree depth penalty terms or minimum sample number limits to constrain model complexity.

[0116] The final trained decision tree model is deployed on multiple distribution control nodes to achieve distributed recognition, as follows:

[0117] Each smart terminal embeds the model locally and receives updates from the master station regularly;

[0118] The control node can independently process the collected data and complete local fault diagnosis;

[0119] Report and link the identification results, such as circuit breaker tripping, area isolation and other control measures.

[0120] Example 2, please refer to Figure 2 As shown, the distributed fault detection system for power distribution control equipment described in this embodiment includes a data acquisition module, a sample scoring module, a generation model training module, a sample enhancement and management module, and a distributed recognition module;

[0121] Data acquisition module: collects historical and real-time power operation data at several control nodes of the power distribution system;

[0122] Sample scoring module: Extracts low-frequency, high-risk fault samples from the collected data and constructs a sample feature set. This includes obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, building a low-frequency, high-risk sample identification function, and performing sample scoring.

[0123] Generative model training module: training a generative model based on the sample feature set, wherein the generative model includes a variational autoencoder or a generative adversarial network, and is used to learn the potential feature distribution of rare fault samples;

[0124] Sample enhancement and management module: uses the generative model to generate simulated rare fault samples and injects them into the original training sample set to construct an enhanced training set;

[0125] Distributed identification module: Trains the enhanced training set, builds a decision tree model based on information gain, forms a classification structure of the fault classification path, and realizes distributed fault type recognition and classification.

[0126] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0127] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. A and B can be singular or plural. Furthermore, the character " / " as used herein generally indicates an "or" relationship between the associated objects, but it may also indicate an "and / or" relationship. For specific understanding, please refer to the context.

[0128] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0129] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A distributed fault detection method for power distribution control equipment, characterized by: include: Collect historical and real-time power operation data at several control nodes in the power distribution system; Extract low-frequency, high-risk fault samples from the collected data and construct a sample feature set. This includes obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, building a low-frequency, high-risk sample identification function, and performing sample scoring. Training a generative model based on the sample feature set, wherein the generative model includes a variational autoencoder or a generative adversarial network for learning the potential feature distribution of rare fault samples; Generate simulated rare fault samples using the generative model and inject them into the original training sample set to construct an enhanced training set; The enhanced training set is trained to build a decision tree model based on information gain, forming a classification structure of fault classification paths and realizing distributed fault type recognition and classification.

2. A distributed fault detection method for power distribution control equipment according to claim 1, characterized in that: The collection of historical and real-time power operation data at several control nodes of the power distribution system includes: configuring power data collection devices at the feeder terminal FTU, switch station and intelligent ring network cabinet control nodes; collecting current, voltage, frequency and harmonic parameters in real time, and attaching timestamps and node location codes; the current data includes three-phase current signals , sampling at time intervals of 20ms or less; the collected data is initially archived and denoised by the edge computing unit and then uploaded to the central training platform.

3. A distributed fault detection method for power distribution control equipment according to claim 2, characterized in that: By obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, a low-frequency high-risk sample identification function is constructed to perform sample scoring, including: Collect three-phase current , calculate the zero sequence component by Clarke transform ,Right now: ; then Perform first-order difference or differential calculation to obtain its rate of change per unit time, which is expressed as: ; Represents the zero-sequence current component, which measures the asymmetry of the three-phase current; It represents the rate of change of the zero-sequence component and the degree of abnormal fluctuation; Δt represents the sampling time interval; Perform fast Fourier transform on the collected voltage or current signal to extract the phase value of the nth harmonic , where n∈{3,5,7}; calculate the phase mutation of each harmonic between adjacent sampling periods: ; Then the disturbance values ​​of all harmonics are weighted averaged to obtain the instantaneous harmonic phase disturbance index HPI.

4. A distributed fault detection method for power distribution control equipment according to claim 3, characterized in that: The zero-sequence component change rate and the instantaneous harmonic phase disturbance index are combined to construct a scoring function for identifying abnormal samples: ; Among them, α and β are adjustment weights, and the scoring threshold Rth is set. When R(t)>Rth, the sample is judged to be a low-frequency high-risk fault sample and enters the training sample set of the subsequent generation model.

5. A distributed fault detection method for power distribution control equipment according to claim 4, characterized in that: The training generation model specifically includes: Take the scored and labeled high-risk samples as input to train a variational autoencoder or a generative adversarial network; The VAE model uses a multi-scale residual convolution structure to extract input features and introduces an attention mechanism to adaptively allocate feature weights. The discriminator of the GAN model has multi-tasking capabilities. In addition to judging the authenticity of samples, it also outputs sample category labels and anomaly scores. During the training process, the reconstruction error and KL divergence are optimized to enhance the stability of the generated samples.

6. A distributed fault detection method for power distribution control equipment according to claim 5, characterized in that: The generating of simulated rare fault samples and injecting them into the training set includes: Use the trained generative model to sample latent variables from the latent space , generate simulation samples; Apply the isolation forest algorithm to the generated samples to remove outliers; Match the categories of the screened simulated samples with the original training samples, and use genetic algorithms to optimize the injection ratio of each category; The injection weight is dynamically adjusted according to the category scarcity to construct an enhanced training set.

7. A distributed fault detection method for power distribution control equipment according to claim 6, characterized in that: The construction of a decision tree model based on information gain includes: Input the features and class labels in the enhanced training set into the decision tree algorithm; Using information gain as the node division criterion, recursively select the optimal features to construct the branch structure; When the number of leaf node samples is insufficient or the categories are impure, pruning operations are triggered, including pre-pruning and post-pruning strategies; After training is completed, the model structure is embedded into each control node terminal to achieve distributed fault judgment and synchronous reporting.

8. A distributed fault detection system for power distribution control equipment, used to implement the distributed fault detection method for power distribution control equipment according to any one of claims 1 to 7, characterized in that: It includes data collection module, sample scoring module, generation model training module, sample enhancement and management module, and distributed recognition module; Data acquisition module: collects historical and real-time power operation data at several control nodes of the power distribution system; Sample scoring module: Extracts low-frequency, high-risk fault samples from the collected data and constructs a sample feature set. This includes obtaining the current zero-sequence component change rate and the instantaneous harmonic phase disturbance index, building a low-frequency, high-risk sample identification function, and performing sample scoring. Generative model training module: training a generative model based on the sample feature set, wherein the generative model includes a variational autoencoder or a generative adversarial network, and is used to learn the potential feature distribution of rare fault samples; Sample enhancement and management module: uses the generative model to generate simulated rare fault samples and injects them into the original training sample set to construct an enhanced training set; Distributed identification module: Trains the enhanced training set, builds a decision tree model based on information gain, forms a classification structure of the fault classification path, and realizes distributed fault type recognition and classification.

Images