Leakage current intelligent detection positioning method and system based on multi-modal modulation and space-time convolution

Abstract

The invention provides a leakage current intelligent detection positioning method and system based on multi-modal modulation and space-time convolution, and belongs to the technical field of power distribution network fault diagnosis. According to the fault detection and positioning model, based on a transformer area topological graph, space-time features are extracted by adopting a space-time diagram convolutional network, and an attention mechanism is fused with environment modal information to carry out fault positioning. The method can automatically and synchronously learn the propagation rule of the fault in the topological space of the power distribution area and the dynamic evolution characteristics of the fault in the time dimension from the data, thereby capturing a finer and more complex fault mode, achieving the more accurate fault positioning, and reducing the false alarm and missing report. The introduced multi-modal attention fusion mechanism can dynamically evaluate the importance of fault signals of each node under different environmental conditions, so that intelligent diagnosis of environmental sensitivity is realized, and performance reduction caused by using a fixed judgment criterion under different environments is avoided.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network fault diagnosis technology, specifically relating to a method and system for intelligent detection and location of leakage current based on multimodal modulation and spatiotemporal convolution. Background Technology

[0002] Leakage current faults are a common safety hazard in low-voltage distribution substations, potentially causing accidents such as electric shock and fire. Traditional detection methods primarily rely on residual current devices (RCDs), which determine operation based on a fixed threshold (e.g., 30mA). However, in actual operation, normal leakage current can reach hundreds of milliamperes, far exceeding the set threshold, leading to frequent false trips by the RCD. To avoid false trips, on-site maintenance personnel often resort to temporary measures such as short-circuiting the RCD, which inadvertently creates even greater safety hazards. Furthermore, traditional methods typically rely solely on current signals, lacking the fusion analysis of multi-modal information such as voltage, temperature, and humidity, making it difficult to comprehensively capture fault characteristics caused by the coupling effects of multiple factors, including insulation aging and humid environments. In structurally complex low-voltage distribution substations, radial and trunk-like topologies further increase the difficulty of detection. Traditional amplitude comparison methods and current pulse methods, due to the lack of spatial topology modeling, often result in branch line location errors exceeding 10 meters. On the other hand, existing detection methods suffer from poor real-time performance: centralized cloud processing suffers from communication delays, failing to meet the millisecond-level response requirements of portable devices; while clamp meters are inefficient for point-by-point detection, often taking several days for a single comprehensive inspection. At the model level, traditional machine learning methods such as Support Vector Machines (SVM) or single deep learning models such as Convolutional Neural Networks (CNN) are poorly adaptable to noise and different grounding methods (such as TN-S, TT, IT), and have insufficient generalization ability.

[0003] In recent years, machine learning technology has been gradually introduced into the field of leakage current detection, with some studies attempting to overcome the limitations of traditional methods. In signal processing and feature extraction, some studies have used Inherent Time Scale Decomposition (ITD) combined with Convolutional Deep Belief Networks (CDBN) to achieve fault segmentation and localization. However, this method does not consider the transformer substation topology and relies on power frequency signals, limiting its ability to capture transient features. Regarding multimodal data fusion, methods based on CNNs and Feature Linear Modulation (FiLM) have improved classification accuracy by introducing environmental parameters, but lack lightweight designs for edge computing, making it difficult to balance computational complexity and real-time requirements. Furthermore, deep learning models such as Spatiotemporal Graph Convolutional Networks (ST-GCN) have been used to extract spatiotemporal features, but related research is mostly focused on cloud analysis, and the edge-cloud collaboration mechanism is still imperfect, failing to achieve low-latency response. Although existing research has made some progress in signal processing and model accuracy, common challenges remain, including insufficient data fusion, lack of topological relationship modeling, insufficient real-time performance, and weak model generalization ability. For example, the Intrinsic Time-scale Decomposition – Convolutional Deep Belief Network (ITD-CDBN) method does not effectively utilize multimodal data, while the Convolutional Neural Network – Feature-wise Linear Modulation (CNN-FiLM) scheme has significant shortcomings in edge deployment.

[0004] Therefore, existing technologies have not yet systematically solved the problems of data uniformity, topology ignoring, poor real-time performance, and weak model adaptability. Faced with the urgent need for high-precision fault diagnosis in smart distribution networks, there is a pressing need to develop a new intelligent leakage current identification method that can integrate multimodal information, model spatiotemporal topology, and support edge-cloud collaboration, in order to achieve more accurate, faster, and adaptive leakage current detection and location. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of single data, topology neglect, poor real-time performance and insufficient model generalization ability in existing leakage current location technology, and to provide a method and system for intelligent leakage current detection and location based on multimodal modulation and spatiotemporal convolution.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for intelligent detection and localization of leakage current based on multimodal modulation and spatiotemporal convolution, comprising the following steps: Real-time acquisition of multimodal time-series data from distribution radio equipment and preprocessing of the acquired data; The preprocessed data is input into a pre-trained fault detection and localization model to generate the probability distribution results of the fault points; The fault detection and localization model is based on the transformer area topology map, uses a spatiotemporal graph convolutional network to extract spatiotemporal features, and uses an attention mechanism to fuse environmental modal information for fault localization.

[0007] The collected multimodal time-series data includes electrical data and environmental data. The electrical data includes three-phase current and voltage, and the environmental data includes temperature and humidity.

[0008] The method for preprocessing the collected data is as follows: Denoising of the acquired data is performed using wavelet packet transform:

[0009] in: Noise threshold; Where N is the noise standard deviation and N is the signal length; Normalize the denoised data:

[0010] in: For the normalized data, The data after noise reduction. and The mean and standard deviation are given. Environmental feature vectors are obtained by encoding environmental data.

[0011] The specific method for inputting the preprocessed data into the pre-trained fault detection and localization model to generate the probability distribution results of the fault points is as follows: A topology map of the distribution transformer area is constructed based on the electrical connection relationship of the distribution transformer area. The node feature matrix of the topology map is generated using preprocessed multimodal data. The adjacency matrix of the topology map is calculated by the electrical distance between the collection points. The obtained node feature matrix and adjacency matrix are input into the spatiotemporal graph convolutional network to extract spatiotemporal features. The extracted spatiotemporal features are then fused to obtain environment-adaptive fused features. Based on the obtained fusion features, the probability distribution of faulty nodes within the distribution radio area is generated.

[0012] In the step of constructing a topology map of a distribution substation based on its electrical connection relationships, generating a node feature matrix of the topology map using preprocessed multimodal data, and calculating the adjacency matrix of the topology map through the electrical distance between sampling points, the method for calculating the adjacency matrix of the topology map is as follows:

[0013] in: Let the connection weight between node i and node j be . Let i be the electrical distance between nodes i and j. This is the scale parameter.

[0014] The method for inputting the obtained node feature matrix and adjacency matrix into a spatiotemporal graph convolutional network to extract spatiotemporal features, and then fusing the extracted spatiotemporal features to obtain environment-adaptive fused features is as follows: The obtained node feature matrix and adjacency matrix are input into the spatiotemporal graph convolutional network, and then passed through spatial convolutional layers and temporal convolutional layers to capture spatial topological features and dynamic temporal features. The environmental feature vector and spatiotemporal features are input into the attention fusion module. The environmental feature vector is used as a query to perform weighted fusion of the spatiotemporal features to obtain environmentally adaptive fusion features.

[0015] The spatial convolutional layer employs a Chebyshev multinomial approximation graph convolution kernel to capture spatial topological features, as detailed below:

[0016]

[0017] in, For graph convolution kernel, For learnable parameter vectors, The characteristic signal of the input node, This is the scaled, normalized Laplace matrix; The maximum order of the Chebyshev polynomial; It is a Chebyshev polynomial; Let be the Laplace matrix of the graph. The largest eigenvalue of the Laplace matrix. To and Identity matrices of the same latitude are used to ensure the numerical stability of matrix operations; The temporal convolutional layer uses a one-dimensional convolutional kernel to perform sliding convolution on the feature sequence of each node along the time axis to extract local temporal features.

[0018] The method for inputting environmental feature vectors and spatiotemporal features into the attention fusion module, using the environmental feature vectors as a query, and weightedly fusing the spatiotemporal features to obtain environmentally adaptive fused features is as follows:

[0019] in: Let be the attention weight of the i-th node. For environmental feature vectors, Let W represent the spatiotemporal features of the i-th node, and W be the learnable weight matrix. Let be the multimodal feature vector of the j-th node. The multimodal feature vector includes environmental feature vector and spatiotemporal feature vector.

[0020] The method for generating the probability distribution of faulty nodes within the distribution area based on the obtained fusion features is as follows: The obtained fused features h are mapped to the final fault node probability distribution P using a fully connected network:

[0021] W0 and b0 are the learnable weights and biases of the fully connected layer. The softmax function normalizes the output values ​​of all nodes into a probability distribution, and the node with the highest probability is identified as the fault point by the model.

[0022] Secondly, the present invention provides a leakage current intelligent detection and location system based on multimodal modulation and spatiotemporal convolution, comprising: The data acquisition module is used to collect multimodal time-series data of the distribution radio station equipment in real time and to preprocess the collected data. The fault detection module is used to input the pre-processed data into the pre-trained fault detection and localization model to generate the probability distribution results of the fault points.

[0023] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method and system for intelligent leakage current detection and localization based on multimodal modulation and spatiotemporal convolution. By constructing a fault detection and localization model based on a distribution transformer area topology map, extracting spatiotemporal features using a spatiotemporal graph convolutional network, and fusing environmental modal information using an attention mechanism, the system can automatically and synchronously learn the propagation patterns of faults in the distribution transformer area topology and their dynamic evolution characteristics over time from data. This allows for the capture of more subtle and complex fault modes, achieving more accurate fault localization and reducing false alarms and missed alarms. The introduced multimodal attention fusion mechanism can dynamically evaluate the importance of fault signals at each node under different environmental conditions, thereby achieving environment-sensitive intelligent diagnosis and avoiding the performance degradation caused by using fixed judgment criteria in different environments.

[0024] Furthermore, by combining graph convolution and one-dimensional convolution through a spatiotemporal graph convolutional network, the model can directly process spatiotemporal graph data, characterize the spatiotemporal dynamic process of fault propagation along power lines, and make the localization logic more consistent with physical reality. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention; Figure 2This is a comparison diagram of the fault location effect in Embodiment 2 of the present invention. Detailed Implementation

[0026] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0027] Example 1 like Figure 1 As shown, a method for intelligent detection and localization of leakage current based on multimodal modulation and spatiotemporal convolution includes the following steps: S1: Real-time acquisition of multimodal time-series data from distribution radio station equipment and preprocessing of the acquired data; S2: Input the preprocessed data into the pre-trained fault detection and localization model to generate the probability distribution results of the fault points; The fault detection and localization model is based on the transformer area topology map, uses a spatiotemporal graph convolutional network to extract spatiotemporal features, and uses an attention mechanism to fuse environmental modal information for fault localization.

[0028] Specifically, in S1, multimodal time-series data of the distribution transformer area equipment is collected and preprocessed. Sensors are deployed at key nodes of the distribution transformer area equipment, such as transformers, branch boxes, and user terminals, and collect leakage current, voltage, temperature, and humidity data of the equipment at a preset sampling frequency, and transmit them through the Internet of Things (NB-IoT / 5G).

[0029] The collected multimodal time-series data includes electrical and environmental data, with the electrical data including three-phase current. ,Voltage Environmental data temperature and humidity Its data shape is C×T, where C is the number of channels and T is the length of the time series.

[0030] The collected equipment data is preprocessed as follows: First, the raw data is cleaned to remove or reduce random noise mixed in with the useful signal, thereby improving signal quality. The wavelet packet transform denoising threshold formula is as follows:

[0031] in: This is the noise threshold, used to determine whether wavelet coefficients should be retained or removed. All wavelet coefficients smaller than this threshold will be set to zero or shrunk, while wavelet coefficients larger than the threshold will be retained or weakened. Where is the noise standard deviation and N is the signal length.

[0032] After denoising, data normalization is performed to eliminate differences in the dimensions and numerical ranges of different physical quantities, bringing all features to the same scale. The specific formula is as follows:

[0033] in: For the normalized data, The data after noise reduction. and The mean and standard deviation of the training set.

[0034] Environmental feature vectors are obtained by encoding environmental data.

[0035] Specifically, in S2, the preprocessed data is input into a pre-trained fault detection and localization model to generate the probability distribution of fault points. The specific method is as follows: S21: Construct a topology map of the distribution area based on the electrical connection relationship of the distribution area, generate a node feature matrix of the topology map using preprocessed multimodal data, and calculate the adjacency matrix of the topology map by the electrical distance between the collection points; S22: Input the obtained node feature matrix and adjacency matrix into the spatiotemporal graph convolutional network to extract spatiotemporal features, and fuse the extracted spatiotemporal features to obtain environment-adaptive fused features; S23: Based on the obtained fusion features, generate the probability distribution of faulty nodes within the distribution area.

[0036] Specifically, in S21, a topology map of the transformer area is constructed based on the electrical connection relationship of the transformer area. The node feature matrix of the topology map is generated using preprocessed multimodal data, and the adjacency matrix of the topology map is calculated by the electrical distance between the collection points.

[0037] 1) A topology graph G=(V,E) for the transformer substations is constructed based on the electrical connection relationships. The node set V represents the measurement (acquisition) points, and the edge set E represents the electrical connection relationships, such as cable length and impedance. The originally isolated measurement point data is organized into a topological network through electrical connection relationships. This allows the model to not only see the data of each point, but also utilize the connection information between points.

[0038] 2) Generate a node feature matrix F of the topology graph based on the preprocessed multimodal data. Each measurement point is described by D features, including the effective value of three-phase current, voltage amplitude, temperature and humidity.

[0039] The feature matrix is Where N is the number of measurement points and D is the feature dimension.

[0040] 3) Calculate the adjacency matrix A of the topology graph based on the electrical distance between the collection points, as follows:

[0041] in: Let the connection weight between node i and node j be . Let i be the electrical distance between nodes i and j. This is the scale parameter.

[0042] Specifically, in S22, the obtained node feature matrix and adjacency matrix are input into the spatiotemporal graph convolutional network to extract spatiotemporal features. The extracted spatiotemporal features are then subjected to attention weight calculation and weighted fusion to obtain environment-adaptive fused features.

[0043] 1) Input the obtained node feature matrix and adjacency matrix into the spatiotemporal graph convolutional network, and capture spatial topological features and dynamic temporal features by passing them through spatial convolutional layers and temporal convolutional layers in sequence.

[0044] The spatial convolutional layer uses Chebyshev polynomial approximation graph kernels to capture spatial topological features, as detailed below:

[0045]

[0046] in, For graph convolution kernel, For learnable parameter vectors, The characteristic signal of the input node, This is the scaled, normalized Laplace matrix; The maximum order of the Chebyshev polynomial is the size of the receptive field of the graph convolution, which determines the information that each node can aggregate within the range of K-hop neighbors when performing convolution. It is a Chebyshev polynomial; Let be the Laplace matrix of the graph. The largest eigenvalue of the Laplace matrix. To and Identity matrices of the same latitude are used to ensure the numerical stability of matrix operations; The temporal convolutional layer uses a one-dimensional convolutional kernel with a kernel length of 5 and a stride of 1. It performs sliding convolution on the feature sequence of each node along the time axis to extract local temporal features.

[0047] 2) Input the environmental feature vector and spatiotemporal features into the attention fusion module, use the environmental feature vector as a query, and perform weighted fusion of spatiotemporal features to obtain environmentally adaptive fusion features;

[0048] in: Let be the attention weight of the i-th node. For environmental feature vectors, Let W represent the spatiotemporal features of the i-th node, and W be the learnable weight matrix. Let be the multimodal feature vector of the j-th node. The multimodal feature vector includes environmental feature vector and spatiotemporal feature vector.

[0049] Specifically, in S23, based on the obtained fusion features, the probability distribution of faulty nodes within the transformer area is generated.

[0050] The fused features h obtained from S3 are mapped to the final probability distribution P of the faulty nodes using a fully connected network:

[0051] W0 and b0 are the learnable weights and biases of the fully connected layer. The softmax function normalizes the output values ​​of all nodes into a probability distribution, where the node with the highest probability is identified by the model as the most likely fault point.

[0052] Furthermore, a weighted cross-entropy loss function is used to train the constructed fault detection and localization model to address the problem of extreme class imbalance in fault samples within the distribution transformer area, where normal nodes far outnumber faulty nodes. The specific formula is as follows:

[0053] in, These are the genuine labels for the equipment in the distribution area. Predictive labels for equipment in the distribution area. The class weights are calculated based on the frequency of the training set.

[0054] The optimizer uses AdamW, and the learning rate scheduling adopts the OneCycle strategy.

[0055] During the model training phase, a weighted cross-entropy loss function was used. By assigning larger loss weights to fault categories with smaller sample sizes, the model is forced to pay more attention to fault modes that are difficult to learn. This effectively avoids the tendency of model training to favor the majority class (normal state) and ensures that the model has excellent detection capabilities for all types of faults, especially rare faults.

[0056] Example 2 Taking a 10kV distribution substation as an application scenario, a hybrid radial and trunk topology is adopted, including 1 transformer, 5 branch boxes, and 14 user terminals, with a total of 20 measurement nodes. The total length of the substation lines is 3.5 kilometers, the grounding method is TN-S, and the load types include residential loads, power loads, and distributed photovoltaic access.

[0057] Multimodal sensor clusters are installed at transformer outlets, branch box inlets and outlets, and user meter boxes. Each sensor integrates a high-precision current transformer (accuracy ±0.5%), a voltage sampling circuit (accuracy ±1%), and a temperature and humidity sensor (temperature range -40° to 85°, humidity range 0 to 100%RH).

[0058] S1: Under the condition of a sampling frequency of 10KHz, synchronously collect three-phase current ( ),Voltage( ), temperature (T) and humidity (H); covering a complete power frequency cycle (20ms) before and after the fault, continuously collected for 30 days, with a cumulative data volume of approximately 1.2TB.

[0059] The collected equipment data is preprocessed as follows: Wavelet packet transform denoising: Sym4 wavelet basis is used, and the threshold calculation formula is as follows:

[0060] in The noise standard deviation is given (estimated from the signal tail segment), and N=200 is the window length. After denoising, the signal-to-noise ratio is improved to over 35dB.

[0061] Data normalization: Perform Z-score normalization on each channel:

[0062] in Calculations based on the training set (such as current signals) =-0.02A, =1.15A).

[0063] S21: Construct a diagram G=(V,E) based on the electrical drawings of the transformer substation. The node set V contains 20 measurement points, and the edge set E represents electrical connection relationships, such as cable length and impedance.

[0064] A node feature matrix F is generated based on the preprocessed multimodal data to create a topology graph. Each measurement point is described by six features: RMS three-phase current, voltage amplitude, temperature, and humidity. Feature matrix .

[0065] Among them, there are 3 effective values ​​of the three-phase current, which are used to reflect the load condition and current balance state at that point; Voltage amplitude: used to reflect the voltage level at that point; Temperature: Used to reflect the operating temperature of the equipment; Humidity: Used to reflect environmental conditions, which may affect the insulation performance of equipment.

[0066] Adjacency matrix calculation: Defining connection weights based on electrical distance:

[0067] in The distance between electrical appliances is calculated using line parameters. =0.1 is the scale parameter, and the matrix is ​​normalized after adding self-loops.

[0068] S22: Spatial Graph Convolution: Chebyshev polynomial approximation (order K=3), hidden layer dimension 64, Laplacian matrix normalization parameters. .

[0069] Temporal convolution: One-dimensional convolution kernel length 5, stride 1, output channels 128, activation function is GELU.

[0070] Attention Fusion: Environmental Feature Vector Attention weights are generated using LMP (32-dimensional hidden layer):

[0071] Where W is the learnable weight matrix.

[0072] Training parameters: Loss function: Weighted cross-entropy Category weight The weight is calculated based on the distribution of the training set, such as 1.2 for unidirectional grounding and 1.5 for electrical arc.

[0073] Optimizer: AdamW (learning rate 0.001, weight decay 0.01), learning rate adjustment adopts OneCycle strategy, maximum learning rate 0.005.

[0074] Training cycle: 100 epochs, batch_size=32, early stop patience value: 10 epochs.

[0075] Hardware environment: The edge uses an ARM Cortex-A53 processor with 1GB of memory, and the cloud uses an NVIDIA V100 GPU for training.

[0076] S23: By employing the method described in this invention, such as Figure 2 As shown, on a test set containing 200 fault samples, the localization accuracy reached 97.8%, with an average error of 2.8 meters (compared to >10 meters for traditional error methods). The edge inference latency from data acquisition to result output was 45ms, meeting the requirements of portable devices. Furthermore, tests were conducted on a TT-connected ground station with 15 nodes, achieving an accuracy of 96.1%, demonstrating the generalization ability of the fault detection and localization model.

[0077] Example 3 A smart leakage current detection and location system based on multimodal modulation and spatiotemporal convolution, comprising: The data acquisition module is used to collect multimodal time-series data of the distribution radio station equipment in real time and to preprocess the collected data. The fault detection module is used to input pre-processed data into a pre-trained fault detection and localization model, generate the probability distribution results of fault points, and display them through a visualization interface.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent detection and localization of leakage current based on multimodal modulation and spatiotemporal convolution, characterized in that, Includes the following steps: Real-time acquisition of multimodal time-series data from distribution radio equipment and preprocessing of the acquired data; The preprocessed data is input into a pre-trained fault detection and localization model to generate the probability distribution results of the fault points; The fault detection and localization model is based on the transformer area topology map, uses a spatiotemporal graph convolutional network to extract spatiotemporal features, and uses an attention mechanism to fuse environmental modal information for fault localization.

2. The intelligent leakage current detection and localization method based on multimodal modulation and spatiotemporal convolution according to claim 1, characterized in that, The collected multimodal time-series data includes electrical data and environmental data. The electrical data includes three-phase current and voltage, and the environmental data includes temperature and humidity.

3. The intelligent leakage current detection and localization method based on multimodal modulation and spatiotemporal convolution according to claim 2, characterized in that, The method for preprocessing the collected data is as follows: Denoising of the acquired data is performed using wavelet packet transform: in: Noise threshold; Where N is the noise standard deviation and N is the signal length; Normalize the denoised data: in: For the normalized data, The data after denoising. and The mean and standard deviation are given. Environmental feature vectors are obtained by encoding environmental data.

4. The intelligent leakage current detection and localization method based on multimodal modulation and spatiotemporal convolution according to claim 1, characterized in that, The specific method for inputting the preprocessed data into the pre-trained fault detection and localization model to generate the probability distribution results of the fault points is as follows: A topology map of the distribution transformer area is constructed based on the electrical connection relationship of the distribution transformer area. The node feature matrix of the topology map is generated using preprocessed multimodal data. The adjacency matrix of the topology map is calculated by the electrical distance between the collection points. The obtained node feature matrix and adjacency matrix are input into the spatiotemporal graph convolutional network to extract spatiotemporal features. The extracted spatiotemporal features are then fused to obtain environment-adaptive fused features. Based on the obtained fusion features, the probability distribution of faulty nodes within the distribution radio area is generated.

5. The intelligent leakage current detection and localization method based on multimodal modulation and spatiotemporal convolution according to claim 4, characterized in that, In the step of constructing a topology map of a distribution substation based on its electrical connection relationships, generating a node feature matrix of the topology map using preprocessed multimodal data, and calculating the adjacency matrix of the topology map through the electrical distance between sampling points, the method for calculating the adjacency matrix of the topology map is as follows: in: Let be the connection weight between node i and node j. Let i be the electrical distance between nodes i and j. This is the scale parameter.

6. A method for intelligent detection and localization of leakage current based on multimodal modulation and spatiotemporal convolution according to claim 3 or 4, characterized in that, The method for inputting the obtained node feature matrix and adjacency matrix into a spatiotemporal graph convolutional network to extract spatiotemporal features, and then fusing the extracted spatiotemporal features to obtain environment-adaptive fused features is as follows: The obtained node feature matrix and adjacency matrix are input into the spatiotemporal graph convolutional network, and then passed through spatial convolutional layers and temporal convolutional layers to capture spatial topological features and dynamic temporal features. The environmental feature vector and spatiotemporal features are input into the attention fusion module. The environmental feature vector is used as a query to perform weighted fusion of the spatiotemporal features to obtain environmentally adaptive fusion features.

7. The intelligent leakage current detection and localization method based on multimodal modulation and spatiotemporal convolution according to claim 6, characterized in that, The spatial convolutional layer employs a Chebyshev multinomial approximation graph convolution kernel to capture spatial topological features, as detailed below: in, For graph convolution kernel, For learnable parameter vectors, The characteristic signal of the input node, This is the scaled, normalized Laplace matrix; The maximum order of the Chebyshev polynomial; It is a Chebyshev polynomial; Let be the Laplace matrix of the graph. The largest eigenvalue of the Laplace matrix. To and Identity matrices at the same latitude; The temporal convolutional layer uses a one-dimensional convolutional kernel to perform sliding convolution on the feature sequence of each node along the time axis to extract local temporal features.

8. The intelligent leakage current detection and localization method based on multimodal modulation and spatiotemporal convolution according to claim 7, characterized in that, The method for inputting environmental feature vectors and spatiotemporal features into the attention fusion module, using the environmental feature vectors as a query, and weightedly fusing the spatiotemporal features to obtain environmentally adaptive fused features is as follows: in: Let be the attention weight of the i-th node. For environmental feature vectors, Let W represent the spatiotemporal features of the i-th node, and W be the learnable weight matrix. Meaning: The multimodal feature vector of the j-th node.

9. The intelligent leakage current detection and localization method based on multimodal modulation and spatiotemporal convolution according to claim 8, characterized in that, The method for generating the probability distribution of faulty nodes within the distribution area based on the obtained fusion features is as follows: The obtained fused features h are mapped to the final fault node probability distribution P using a fully connected network: W0 and b0 are the learnable weights and biases of the fully connected layer. The softmax function normalizes the output values ​​of all nodes into a probability distribution, and the node with the highest probability is identified as the fault point by the model.

10. A smart leakage current detection and location system based on multimodal modulation and spatiotemporal convolution, characterized in that, include: The data acquisition module is used to collect multimodal time-series data of the distribution radio station equipment in real time and to preprocess the collected data. The fault detection module is used to input the pre-processed data into the pre-trained fault detection and localization model to generate the probability distribution results of the fault points.

Images