Low-voltage cabinet fault prediction method and system based on artificial intelligence

Abstract

The invention discloses a low-voltage cabinet fault prediction method and system based on artificial intelligence, and relates to the technical field of low-voltage electrical equipment state monitoring, and the method comprises the steps: obtaining operation data, such as current and voltage, collected by a sensor in a low-voltage cabinet; a multi-dimensional tensor structure containing spatial topological codes is constructed through preprocessing such as normalization; inputting a fault prediction model to obtain fault type probability and trend evaluation parameters, and generating a dominant abnormal factor index set through an interpretable module; analyzing a root cause path by using a fault causal knowledge graph in combination with the data, and determining a fault evolution chain; and finally generating interpretable early warning prediction information. The technical problems that a traditional low-voltage cabinet fault detection mode is difficult to process multi-dimensional complex data, cannot efficiently mine key fault information and is difficult to accurately predict are solved, the prediction model is constructed by using the deep learning technology, and then efficient processing of the multi-dimensional complex data of the low-voltage cabinet and deep mining of the key fault information are achieved. The technical effect of accurate fault prediction is achieved.

Description

Technical Field

[0001] This invention relates to the field of low-voltage electrical equipment condition monitoring technology, and in particular to a method and system for predicting low-voltage switchgear faults based on artificial intelligence. Background Technology

[0002] As a critical piece of equipment in the power system, the failure of low-voltage switchgear can affect the stable supply of power. Fault prediction is of great significance for ensuring the reliable operation of the system. Existing technologies often rely on manual inspections and conventional sensor monitoring, which, while effective to some extent, have limitations. Traditional methods are unable to capture the potential fault characteristics of low-voltage switchgear in a timely and accurate manner under complex operating conditions.

[0003] Low-voltage switchgear operation data is abundant and complex. Traditional analysis methods cannot efficiently process and deeply mine the value of this data, easily overlooking key fault information and failing to meet the needs of accurate prediction. Deep learning, on the other hand, has advantages in processing multi-dimensional and massive amounts of data. Introducing deep learning technology to address the shortcomings of traditional methods can achieve more accurate low-voltage switchgear fault prediction. Summary of the Invention

[0004] This application provides a method and system for predicting low-voltage switchgear faults based on artificial intelligence, which is used to solve the technical problems that traditional low-voltage switchgear fault detection methods are unable to process multi-dimensional and complex data, cannot efficiently mine key fault information, and are difficult to predict accurately.

[0005] The first aspect of this application provides an artificial intelligence-based method for predicting faults in low-voltage switchgear. The method includes: acquiring operational data collected by multiple sensor nodes within the low-voltage switchgear, the operational data including at least current, voltage, temperature, partial discharge signals, acoustic signals, and switch status parameters; preprocessing the operational data to construct a multidimensional tensor structure with spatial topological coding, wherein the preprocessing includes normalization, anomaly removal, and time alignment; inputting the current tensor into a fault prediction model and outputting a prediction result, the prediction result including fault type probability and trend evaluation parameters; and parsing the prediction result using an interpretable module to generate a set of dominant anomaly factor indicators; based on the fault type probability, trend evaluation parameters, and the set of dominant anomaly factor indicators, using a fault causal knowledge graph to analyze the root cause path of the current prediction result and determine the fault evolution chain; and generating interpretable early warning prediction information based on the dominant anomaly factor indicators and the fault evolution chain.

[0006] A second aspect of this application provides an artificial intelligence-based low-voltage switchgear fault prediction system. The system includes: an operational data acquisition module for acquiring operational data collected by multiple sensor nodes within the low-voltage switchgear, the operational data including at least current, voltage, temperature, partial discharge signals, acoustic signals, and switch status parameters; an operational data preprocessing module for preprocessing the operational data to construct a multidimensional tensor structure with spatial topological coding, wherein the preprocessing includes normalization, anomaly removal, and time alignment; a prediction result acquisition module for inputting the current tensor into a fault prediction model and outputting prediction results, the prediction results including fault type probability and trend evaluation parameters, and performing result analysis through an interpretable module based on the prediction results to generate a set of dominant anomaly factor indicators; a fault evolution chain acquisition module for analyzing the root cause path of the current prediction result using a fault causal knowledge graph based on the fault type probability, trend evaluation parameters, and the set of dominant anomaly factor indicators to determine the fault evolution chain; and an early warning prediction information acquisition module for generating interpretable early warning prediction information based on the dominant anomaly factor indicators and the fault evolution chain.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] This application collects multi-sensor operation data from low-voltage switchgear, preprocesses it to construct a multi-dimensional tensor, inputs it into a deep learning model that integrates spatiotemporal convolution and graph attention networks, outputs fault prediction results, and combines interpretable modules and fault causal knowledge graphs to analyze root causes and generate interpretable early warning information. This achieves accurate fault prediction of low-voltage switchgear, improves the reliability of power system operation and maintenance, and realizes the technical effect of using deep learning technology to build a prediction model, thereby efficiently processing multi-dimensional complex data of low-voltage switchgear and deeply mining key fault information to achieve accurate fault prediction. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the low-voltage switchgear fault prediction method based on artificial intelligence provided in the embodiments of this application.

[0011] Figure 2 This is a schematic diagram of the structure of the low-voltage switchgear fault prediction system based on artificial intelligence provided in the embodiments of this application.

[0012] Figure labeling: Module 1 for running data acquisition, Module 2 for running data preprocessing, Module 3 for prediction result acquisition, Module 4 for fault evolution chain acquisition, and Module 5 for early warning and prediction information acquisition. Detailed Implementation

[0013] This application provides a method and system for predicting low-voltage switchgear faults based on artificial intelligence, which is used to solve the technical problems that traditional low-voltage switchgear fault detection methods are unable to process multi-dimensional and complex data, cannot efficiently mine key fault information, and are difficult to predict accurately.

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0016] Example 1, as Figure 1 As shown, an artificial intelligence-based low-voltage switchgear fault prediction method is provided, wherein the method includes:

[0017] Step A100: Acquire operating data collected by multiple sensor nodes in the low-voltage cabinet. The operating data includes at least current, voltage, temperature, partial discharge signal, acoustic signal, and switch status parameters.

[0018] In this embodiment of the application, the low-voltage switchgear is a key device in the power system used for power distribution and circuit control, and contains multiple electrical components.

[0019] Specifically, multiple sensor nodes are deployed in key areas such as the low-voltage switchgear's busbar compartment, circuit breaker compartment, and cable compartment to ensure coverage of core components and fault-prone areas. These include current sensors that collect three-phase current data (range 0-1000A), voltage sensors that record line voltage values, temperature sensors that monitor temperatures of contacts and busbars (range 0-100℃), partial discharge sensors that capture pulse signals, acoustic sensors that collect abnormal operating noises (frequency 20-20000Hz), and status sensors that acquire binary status parameters of switch opening and closing. Each sensor node collects data in real time, transmits it to the data processing unit via industrial Ethernet, and after aggregation, sorts it by timestamp to form a raw operational dataset containing multiple types and nodes.

[0020] Through the above steps, comprehensive multi-dimensional operating data such as current and voltage in key areas of the low-voltage switchgear are collected, providing complete and accurate raw data support for subsequent preprocessing and feature extraction of deep learning models, and realizing a comprehensive perception of the operating status of the low-voltage switchgear.

[0021] Step A200: Preprocess the running data to construct a multidimensional tensor structure with spatial topological coding, wherein the preprocessing includes normalization, anomaly removal, and time alignment.

[0022] In this embodiment, spatial topology coding refers to coding based on the spatial positional relationship of sensor nodes within the low-voltage cabinet, which is the basis for constructing the first dimension of a multidimensional tensor structure with spatial topology coding.

[0023] Optionally, firstly, different sensor nodes have different sampling frequencies. For example, current and voltage sensors collect data every 10ms, temperature sensors every 50ms, acoustic signal sensors every 20ms, and switch state parameters are collected when the state changes. To ensure consistency of various data in the time dimension, time alignment processing is required. Using 10ms as a uniform time interval, low-frequency data from the temperature sensor is supplemented with values ​​at intermediate moments using linear interpolation; high-frequency data from the acoustic signal is downsampled by averaging adjacent 10ms intervals; and switch state parameters retain the timestamp of the state change moment, keeping the state value unchanged between adjacent change moments, ultimately forming a unified multi-parameter dataset with a consistent time axis.

[0024] Next, the time-aligned dataset may contain outliers, which need to be removed. For current data, the normal range is usually 0-1000A. A sudden increase of 1500A or a negative value of -200A is considered an anomaly. Voltage data normally fluctuates around 380V. Extreme values ​​of 0V or 1000V are considered anomalies. Temperature data normally ranges from 0-100℃. Jumps of -30℃ or 200℃ are considered anomalies. In partial discharge signals and acoustic signals, pulse signals with amplitudes exceeding the 99.9th percentile of similar historical data are also considered anomalies. Those skilled in the art calculate the mean and standard deviation of each parameter, mark values ​​exceeding the mean ± 3 times the standard deviation as anomalies and remove them. The gaps in the removed data are filled with the mean of the five adjacent normal data points to ensure the continuity of the data sequence.

[0025] After anomaly removal, the operational data is normalized. Due to significant differences in the dimensions and ranges of different data types (maximum current 1000A, maximum temperature 100℃, peak partial discharge signal 500mV, acoustic signal amplitude 200mV, and switch state parameters 0 or 1), a min-max normalization method is used. This maps the value of each parameter to the [0,1] interval. Specifically, for each data point, the minimum value of the parameter is subtracted, and then divided by the difference between the maximum and minimum values. For example, in voltage data, 380V corresponds to (380 - minimum voltage) / (maximum voltage - minimum voltage). This ensures that all types of operational data are on the same order of magnitude, avoiding the impact of numerical range differences on subsequent feature extraction.

[0026] Finally, constructing a multidimensional tensor structure with spatial topological coding refers to establishing first, second, and third dimension codes according to the sensor's spatial location, the time window of the monitoring timestamp of the operation data, and the multimodal feature channels, and then integrating these three dimension codes to form a three-dimensional tensor structure. The specific steps are explained in detail in A210-A240.

[0027] By unifying the time dimension of data through time alignment, eliminating invalid interference by removing anomalies, and normalizing the data volume, the quality of the running data is improved, providing a reliable foundation for constructing a multidimensional tensor structure with spatial topological coding.

[0028] Step A300: Input the current tensor into the fault prediction model and output the prediction results. The prediction results include the probability of fault type and trend evaluation parameters. Based on the prediction results, the results are analyzed through the interpretable module to generate a set of dominant anomaly factor indicators.

[0029] In this embodiment, the interpretable module is used to parse the output of the fault prediction model. It takes the fault type with the highest probability in the prediction results as the target probability, uses automatic differentiation technology to calculate the partial derivative of each running data feature with respect to the target probability, performs normalization on the contribution of each data feature based on the partial derivative to generate anomaly factor indicators, and then sets a dynamic threshold according to the fault type to filter the anomaly factor indicators, thereby generating a set of dominant anomaly factor indicators, thereby improving the interpretability of the fault prediction results.

[0030] In one embodiment of this application, the current tensor is input into the fault prediction model to output the prediction result, including extracting local spatial correlation features using a three-dimensional convolution kernel, extracting long-period dependency features using dilated convolution, and aggregating electrical correlation features using a graph attention network. After concatenation, the fault type probability and trend evaluation parameters are output through a fully connected layer, and then the prediction result is output. The specific steps are described in detail in A310-A340.

[0031] Before inputting the current tensor into the fault prediction model, a model architecture including spatiotemporal convolutional layers needs to be constructed. A sample set is constructed and divided using historical running data. The fault prediction model is obtained through training, validation, and testing. The specific steps are explained in detail in A350-A370.

[0032] Based on the prediction results, the dominant anomaly factor index set is generated through the interpretable module. This includes taking the highest probability fault type as the target probability, calculating the partial derivative of the feature with respect to it, normalizing the contribution to generate anomaly factors, and then using dynamic thresholds to filter and obtain the index set. The specific steps are explained in detail in A381-A384.

[0033] Step A400: Based on the fault type probability, trend evaluation parameters, and dominant anomaly factor index set, use the fault causal knowledge graph to analyze the root cause path of the current prediction result and determine the fault evolution chain.

[0034] Specifically, based on the probability of fault type, trend evaluation parameters, and dominant anomaly factor index set, a knowledge graph containing physical components, fault modes, and feature nodes is constructed. After mapping relevant nodes, the root cause path is traced back. The edge weights are adjusted according to the trend parameters and the scores are calculated. The path with the highest score is taken as the fault evolution chain and the evolution time window is marked. The specific steps are explained in detail in A410-A450.

[0035] Step A500: Generate interpretable early warning and prediction information based on the dominant anomaly factor indicators and the fault evolution chain.

[0036] Specifically, the dominant anomaly indicators include key features such as temperature and partial discharge signals. For example, the indicator values ​​for the three nodes in the temperature feature are 1.5, 1.3, and 1.4, respectively, while the indicator values ​​for the two nodes in the partial discharge signal are 1.3 and 1.2. For these indicators, the corresponding sensor location information is extracted, such as the temperature sensor in region A and the partial discharge sensor in region B. Specific numerical changes, such as a 5°C increase in temperature in region A within one hour, are used to form a detailed list of anomalies, clearly indicating that significant anomalies have occurred in the corresponding features of the relevant locations.

[0037] The fault evolution chain presents the highest-scoring path from the root cause node to the fault mode node, assumed to be: poor contactor contact → increased contact resistance → increased local temperature → insulation aging. Evolution time windows are marked, with the root cause triggering period being the past 6 hours and the fault propagation period being the next 4 hours. Based on this evolution chain, the relationships and time nodes of each link are analyzed. For example, poor contactor contact began to appear 6 hours ago, causing the contact resistance to increase 3 hours ago, which in turn caused the local temperature to rise continuously from 2 hours ago, and signs of insulation aging are expected to appear 4 hours later.

[0038] By combining detailed anomaly characteristics with the fault evolution chain, early warning and prediction information is generated. For temperature anomalies, the correlation between the current temperature rise and poor contact of the contactor is explained in conjunction with the local temperature rise stage in the evolution chain, predicting that the temperature may rise to the critical value within the next 4 hours and indicating the potential for insulation aging. For partial discharge signal anomalies, the synergistic effect of signal enhancement and temperature rise is explained in conjunction with the early characteristics of insulation aging, warning of the risk of decreased insulation performance. Simultaneously, the dominant anomaly factor and evolution chain stage corresponding to each early warning information are clearly marked, enabling those skilled in the art to understand the basis of the early warning and the fault development trend.

[0039] By integrating information from dominant anomaly factor indicators and fault evolution chains, the generated interpretable early warning and prediction information clearly presents the anomaly characteristics, fault development path, and future trends, providing technicians with clear judgment criteria and action guidelines, and improving the practicality and reliability of low-voltage switchgear fault early warning.

[0040] Furthermore, step A200 in the method provided in this application embodiment includes:

[0041] A210: Establish the first dimension encoding according to the spatial positional relationship of the sensor nodes.

[0042] A220: Set a time window based on the monitoring timestamp of the running data, and establish a second-dimensional code for the data sequence within the time window.

[0043] A230: Establish a third-dimensional encoding based on the multimodal feature channels of the running data.

[0044] A240: Integrate the first dimension encoding, the second dimension encoding, and the third dimension encoding to construct a three-dimensional tensor structure.

[0045] In this embodiment of the application, the multimodal feature channel refers to the channel corresponding to different types of features in the low-voltage switchgear operation data, specifically including current, voltage, temperature, partial discharge signal, acoustic signal, switch status parameters, etc. These different types of operation data reflect the operation status of the low-voltage switchgear from different dimensions, and each type of data constitutes an independent feature channel.

[0046] Specifically, inside the low-voltage switchgear, sensor nodes are distributed according to functional areas. For example, the busbar compartment has two current sensors and one temperature sensor, the circuit breaker compartment has one voltage sensor, one partial discharge sensor, and one switch status sensor, and the cable compartment has one acoustic sensor and one temperature sensor, for a total of eight sensor nodes. Based on the spatial relationship of these nodes, the three-dimensional coordinates of each node are first mapped. Then, according to their horizontal distance from the low-voltage switchgear door (from near to far) and vertical height (from top to bottom), each node is assigned a unique code. For example, the current sensor in the upper part of the busbar compartment is assigned code 1, the temperature sensor in the lower part of the busbar compartment is assigned code 2, and so on until the acoustic sensor in the cable compartment is assigned code 8. This establishes the first dimension code, with a dimension size of 8, used to characterize the spatial distribution and association of the sensors.

[0047] The monitoring timestamps of the operational data are recorded continuously at 10ms intervals. The time window length is set to 10 seconds, meaning each time window contains 10 seconds / 0.01 seconds = 1000 time steps. For each time window, continuous data sequences from all sensors within that time period are extracted and arranged in chronological order by timestamp. For example, if the window starts at time t0, it contains 1000 data points at timestamps t0, t0+10ms, ..., t0+9.99 seconds. A second-dimensional encoding with a dimension size of 1000 is established to capture the temporal variation characteristics of the data.

[0048] The multimodal feature channels of the operational data include six types: current, voltage, temperature, partial discharge signal, acoustic signal, and switch state parameters. Each feature channel is assigned a unique identifier: current corresponds to the first type, voltage to the second, temperature to the third, partial discharge signal to the fourth, acoustic signal to the fifth, and switch state parameters to the sixth. A third-dimensional encoding with a size of 6 is established to distinguish between different types of operational data features.

[0049] The first dimension is encoded, for example, containing 8 sensor nodes; the second dimension is encoded, containing 1000 time steps; and the third dimension is encoded, containing 6 feature channels. These are integrated, with the indices of each dimension corresponding to each other, forming an 8×1000×6 three-dimensional tensor structure. Each element value in the tensor corresponds to a specific type of data from a specific sensor node at a specific time step. For example, tensor [1,50,2] represents the voltage data value of the first sensor node at the 50th time step.

[0050] By establishing and integrating dimensional codes based on sensor spatial location, time window data sequence, and multimodal feature channels, a three-dimensional tensor structure containing spatial topology, temporal variation, and multiple feature types is constructed, laying the foundation for extracting comprehensive feature information for fault prediction models.

[0051] Furthermore, step A300 in the method provided in this application embodiment includes:

[0052] A310: It uses a three-dimensional convolution kernel to slide along the first dimension of the tensor to extract the local spatial correlation features of sensor nodes.

[0053] A320: Dilated convolution is used in the second dimension of the tensor to expand the temporal receptive field and extract long-period dependent features.

[0054] A330: Construct a node adjacency matrix based on electrical connection relationships, and aggregate the electrical association features between nodes through a graph attention network.

[0055] A340: Combine the local spatial correlation features, long-period dependency features, and electrical correlation features, and output the fault type probability and trend evaluation parameters through the fully connected layer.

[0056] In this embodiment, the three-dimensional convolutional kernel is used to slide along the first dimension of the tensor to extract local spatial correlation features of sensor nodes. Dilated convolution is used in the second dimension of the tensor to expand the temporal receptive field, thereby extracting long-period dependent features. The temporal receptive field refers to the range of the time series of running data that the model can perceive in the second dimension of the tensor.

[0057] Optionally, the first dimension of the current tensor corresponds to the spatial location encoding of 8 sensor nodes. A 3×1×1 three-dimensional convolutional kernel is used, sliding along the first dimension with a stride of 1, covering 3 adjacent sensor nodes at a time. For example, when the convolutional kernel slides over nodes 1-3, it calculates the correlation of data between these 3 nodes at the same time step and feature channel, such as the numerical change relationship between adjacent current sensors and temperature sensors, outputting a 6×1000×6 feature map (8-3+1=6), thereby extracting the local spatial correlation features of the sensor nodes.

[0058] The second dimension of the tensor is a time window containing 1000 time steps. A dilated convolution with a dilation rate of 2 is used in this dimension, with a kernel size of 1×3×1. By setting intervals between kernel elements, the receptive field is expanded to 1 + (3-1)×2 = 5 time steps. This method can capture long-period changes that are discontinuous in time but correlated, such as the slow upward trend of a temperature sensor at time steps 100, 200, and 300, extracting long-period dependent features while keeping the time dimension of the output feature map constant at 1000.

[0059] Based on the electrical connections within the low-voltage switchgear, such as circuit breakers connected to busbars and contactors connected to cables, an 8×8 node adjacency matrix is ​​constructed. An element of 1 in the matrix indicates a direct electrical connection between two nodes, while 0 indicates no direct connection. The extracted features are then input into a graph attention network. The network calculates attention weights, assigning higher weights to nodes with close electrical connections (such as sensors corresponding to busbars and circuit breakers), aggregating these to form the electrical association features of each node, and outputting an 8×1000×32 feature matrix.

[0060] Finally, the features of local spatial correlation, long-period dependence and electrical correlation are spliced ​​together, and after being spliced ​​by node dimension, they are compressed by convolutional layer and expanded into vector input dual-branch fully connected layer to generate fault type probability and trend evaluation parameters and output prediction results. The specific steps are explained in detail in A341-A343.

[0061] By using 3D convolution, dilated convolution, and graph attention network to extract local spatial correlation, long-period dependence, and node electrical correlation features of the sensor, multi-dimensional and effective feature support is provided for subsequent feature stitching and prediction output.

[0062] Furthermore, step A300 in the method provided in this application embodiment includes:

[0063] A350: Constructs the model architecture, including spatiotemporal convolutional layers, graph structure generators, graph attention networks, feature concatenation layers, and fully connected layers.

[0064] A360: Collect historical operating data to construct a sample set. Each sample includes a three-dimensional tensor, a fault type label, and a trend change label. The sample set is divided according to a preset ratio to obtain a training set, a validation set, and a test set.

[0065] A370: The model architecture is trained, validated, and tested using the training set, validation set, and test set to obtain the fault prediction model.

[0066] Specifically, when constructing the model architecture, the spatiotemporal convolutional layer includes a 3D convolutional kernel and a dilated convolutional component. The 3D convolutional kernel size is set to 3×1×1, which is used to slide and extract local spatial correlation features of the sensor in the first dimension. The dilation rate of the dilated convolution is set to 2, which is used to expand the temporal receptive field in the second dimension. The graph structure generator generates a node adjacency matrix based on the connection relationship between the circuit breaker and the busbar and the connection relationship between the contactor and the cable in the low-voltage cabinet. The matrix element value of 1 indicates that there is an electrical connection between two nodes, and 0 indicates no connection. The graph attention network contains 32 hidden units and aggregates the electrical correlation features between nodes by calculating attention weights. The feature splicing layer uses splicing operations to fuse three types of features. The fully connected layer contains two branches, which output the fault type probability and trend evaluation parameters, respectively.

[0067] Next, low-voltage switchgear operation data from the past five years were collected, covering data under normal operating conditions and various fault states such as short circuits, overheating, and insulation aging, for example, 12 types, to construct a sample set. The three-dimensional tensor in each sample was constructed with an 8×1000×6 structure. Fault type labels used integers from 0 to 11 to represent different faults, and trend change labels used floating-point numbers from -1 to 1 to represent the fault development trend, with positive values ​​indicating deterioration and negative values ​​indicating mitigation. The sample set was divided according to a preset ratio of 7:2:1, with 70% used as the training set for model parameter updates, 20% as the validation set for adjusting hyperparameters such as the learning rate and number of network layers, and 10% as the test set for evaluating model performance.

[0068] Finally, the model architecture was trained using the training set with the Adam optimizer. The initial learning rate was set to 0.001, decaying by 10% every 5 epochs, and the batch size was set to 32. The iterations were performed for 100 epochs. The validation accuracy of the model was monitored using the validation set. Training was stopped when the accuracy did not improve for 10 consecutive epochs to determine the optimal hyperparameters. The model was evaluated using the test set, and the accuracy of fault type prediction and the mean square error of trend evaluation parameters were calculated. When the accuracy reached above 92% and the mean square error was below 0.05, the model was considered to have completed training, and a fault prediction model was obtained. The input of the fault prediction model is a three-dimensional tensor with spatial topological encoding. This tensor is composed of the encoding of the spatial location of sensor nodes, the time window of running data, and the multimodal feature channels. The output is the prediction result, including the probability of fault type and the trend evaluation parameters.

[0069] By constructing a model with a specific architecture and training and validating it using historical data samples, a fault prediction model that can accurately output fault type probabilities and trend evaluation parameters was obtained, providing a reliable tool for subsequent fault prediction.

[0070] Furthermore, step A340 in the method provided in this application embodiment includes:

[0071] A341: The local spatial correlation features, long-period dependency features, and electrical correlation features are concatenated according to the node dimension and compressed through the convolutional layer compression channel.

[0072] A342: Expands compressed features into vectors.

[0073] A343: Input the expanded vector into the dual-branch fully connected layer, generate fault type probability parameters through the fault type probability branch, generate trend evaluation parameters through the trend evaluation parameter branch, and output the prediction results.

[0074] Specifically, the dimensions of the local spatial correlation feature are 6×1000×6, the long-period dependency feature is 8×1000×6, and the electrical correlation feature is 8×1000×32, all of which use the sensor node as the first dimension. When splicing according to the node dimension, the local spatial correlation feature is zero-padded to make its first dimension consistent with the other two features (8), and then the three features are spliced ​​in the third dimension to form a spliced ​​feature of 8×1000×44, 6+6+32=44.

[0075] Subsequently, convolutional layers are used to compress the channels. The convolutional layers employ 1×1×44 kernels with a stride of 1 and 16 output channels. This compresses the concatenated features, resulting in 8×1000×16 compressed features. These 8×1000×16 compressed features are then sequentially unfolded into a one-dimensional vector. The unfolded vector has dimensions of 8×1000×16 = 128000, with each element corresponding to a specific node, time step, and compressed feature channel value, forming a flat feature vector that can be input into fully connected layers.

[0076] The 128,000-dimensional expanded vector is input into a two-branch fully connected layer. The fault type probability branch contains two fully connected layers: the first layer reduces the vector from 128,000 dimensions to 1024 dimensions, and the second layer reduces it from 1024 dimensions to 12 dimensions, corresponding to 12 fault types, ultimately generating fault type probability parameters. The trend evaluation parameter branch also contains two fully connected layers: the first layer reduces the vector from 128,000 dimensions to 512 dimensions, and the second layer reduces it from 512 dimensions to 1 dimension, generating trend evaluation parameters. The final output is a prediction result containing both of these parameters.

[0077] The fault type probability branch obtains the fault type probability parameters by outputting the probability distribution through the Softmax activation function, and the trend evaluation parameter branch obtains the trend evaluation parameters by outputting the scalar parameters through the ELU activation function and the linear layer. The specific steps are explained in detail in A343-1-A343-2.

[0078] By concatenating features along the node dimension, performing convolutional compression, expanding vectors, and processing with a dual-branch fully connected layer, the effective fusion and transformation of multi-dimensional features were achieved, outputting quantitative prediction results that can be used for fault analysis, laying the foundation for subsequent interpretable module analysis.

[0079] Furthermore, step A343 in the method provided in this application embodiment includes:

[0080] A343-1: The fault type probability branch outputs the fault type probability distribution through the Softmax activation function to obtain the fault type probability parameter.

[0081] A343-2: The trend evaluation parameter branch obtains the trend evaluation parameters by using the ELU activation function and the linear layer to output the trend evaluation scalar parameters.

[0082] In this embodiment, the Softmax activation function is used in the fault type probability branch to output the fault type probability distribution and obtain the fault type probability parameters. The ELU activation function is used in the trend evaluation parameter branch in conjunction with the linear layer to output the trend evaluation scalar parameters and obtain the trend evaluation parameters.

[0083] Specifically, the input to the fault type probability branch is the expanded feature vector, which corresponds to the 12 fault types of the low-voltage switchgear. After the vector is input into this branch, it is first transformed through a fully connected layer containing 64 neurons, and then connected to the Softmax activation function. The Softmax function normalizes the 12 output values, making each value between 0 and 1, with a sum of 1, forming the fault type probability distribution. For example, after calculation, the output value corresponding to short-circuit fault is 0.82, overheating fault is 0.13, insulation aging is 0.03, and the remaining fault types are all below 0.01, thus obtaining the fault type probability parameters.

[0084] The trend evaluation parameter branch also receives this multi-dimensional feature vector. It is first processed by the ELU activation function, which effectively alleviates the gradient vanishing problem. The ELU function fits the nonlinear features reflecting the fault development trend (such as slow temperature rise, enhanced partial discharge signal, etc.) in the vector, outputting a feature vector with unchanged dimensions. This vector is then input into a linear layer with one neuron. The linear layer converts the feature vector into a single scalar parameter through weighted summation. The scalar range is set from -1 to 1, where positive values ​​indicate a worsening fault trend and negative values ​​indicate a mitigating trend. For example, an output scalar of 0.72 indicates a significant worsening trend in the current fault, thus obtaining the trend evaluation parameter.

[0085] By processing the Softmax activation function, the ELU activation function, and the linear layer respectively, the fault type probability branch and the trend evaluation parameter branch each generate corresponding parameters, providing quantitative prediction results to support the subsequent interpretation of the module and the determination of the fault evolution chain.

[0086] Furthermore, step A300 in the method provided in this application embodiment includes:

[0087] A381: Use the fault type with the highest probability in the prediction results as the target probability.

[0088] A382: Calculate the partial derivative of each running data feature with respect to the target probability using automatic differentiation technology. The partial derivative represents the change in target probability caused by a unit change in feature.

[0089] A383: Based on the partial derivatives of each data feature, normalize the contribution of each data feature to generate an anomaly factor index.

[0090] A384: Set dynamic thresholds according to the fault type, and use the dynamic thresholds to filter the abnormal factor indicators to obtain the set of dominant abnormal factor indicators.

[0091] In this embodiment of the application, automatic differentiation is a technique used to calculate the partial derivative of each running data feature with respect to the target probability, wherein the partial derivative represents the change in target probability caused by a unit change in feature.

[0092] In one embodiment, the prediction results include probability distributions for 12 fault types. For example, the probability of a short-circuit fault is 0.82, overheating fault is 0.13, insulation aging is 0.03, and the probability of the remaining types is all below 0.01. The probability value of 0.82 corresponding to the short-circuit fault, which has the highest probability, is taken as the target probability, thereby focusing on analyzing the most likely fault type.

[0093] For six operational data features, including current, voltage, and temperature, each feature contains time-series data from eight sensor nodes. Automatic differentiation techniques are used to calculate the partial derivative of each feature with respect to the target probability. For example, the partial derivative of a certain node's data in the temperature feature is 0.05, meaning that for every unit change in temperature at that node, the target probability of a short-circuit fault increases by 0.05. Similarly, the partial derivative of a certain node's data in the voltage feature is -0.02, meaning that for every unit change in voltage at that node, the target probability decreases by 0.02. This quantifies the impact of each feature on the target probability.

[0094] Next, based on the partial derivatives of each data feature, normalization contribution is performed to generate anomaly factor indicators. This includes taking the absolute value of the operational data feature value, calculating the L2 norm of the probability of all fault types, dividing the absolute value of the feature by the L2 norm, and then multiplying it by the absolute value of the partial derivative. The specific steps are explained in detail in A383-1-A383-3.

[0095] Finally, dynamic thresholds are set according to the fault type. These thresholds are adjusted based on the fault type, and the specific values ​​are determined by those skilled in the art based on the fault type and actual conditions. For example, for short-circuit faults, the dynamic threshold is set to 1.2; for overheating faults, based on their fault characteristics and historical data patterns, the dynamic threshold is set to 1.5; for insulation aging faults, the dynamic threshold can be set to 0.9, and so on. Different fault types correspond to different thresholds to adapt to the differences in the characteristics of various faults and achieve targeted screening.

[0096] By determining the target probability, calculating the partial derivative, generating and screening anomaly factor indicators, a set of dominant anomaly factor indicators was generated, clarifying the key factors affecting the fault prediction results and improving the interpretability of fault prediction.

[0097] Furthermore, step A383 in the method provided in this application embodiment includes:

[0098] A383-1: Take the absolute value of the characteristic value of the running data.

[0099] A383-2: Calculate the L2 norm of the probabilities of all fault types, where the L2 norm is the square root of the sum of the squares of the probabilities.

[0100] A383-3: Divide the absolute value of the feature by the L2 norm, and then multiply it by the absolute value of the partial derivative to generate the anomaly factor index.

[0101] In this embodiment, the operational data feature value refers to the specific numerical value of the operational data collected by multiple sensor nodes within the low-voltage switchgear. The L2 norm is the square root of the sum of squares of all probabilities, used to calculate the L2 norm of the probabilities of all fault types.

[0102] Optionally, the operating data features include current, voltage, temperature, partial discharge signals, etc., and each feature has a specific value. For example, a temperature sensor may collect a feature value of 28, and a partial discharge signal may have a feature value of -15. The absolute values ​​of these feature values ​​are taken to obtain 28 and 15, thereby eliminating the influence of the positive and negative directions of the feature values ​​and unifying them as non-negative values ​​for subsequent calculations.

[0103] Next, the probabilities of all fault types form a set of data. Assume there are 12 fault types, including short circuit, overheating, and insulation aging, with probabilities of 0.78, 0.12, 0.05, 0.02, 0.01, ..., 0.001, respectively. The L2 norm of these probabilities is calculated by first squareding each probability, then summing the squares, and finally taking the square root. This is used to quantify the overall distribution intensity of the probability of all failure types.

[0104] Finally, for each data feature, the absolute value of the previously obtained feature is divided by the L2 norm, and then multiplied by the absolute value of the partial derivative of that feature with respect to the target probability. For example, the absolute value of the temperature feature, 28, divided by the L2 norm of 0.79, is approximately 35.44. If the absolute value of its partial derivative is 0.04, then 35.44 × 0.04 ≈ 1.42, which is the anomaly factor index of this temperature feature. Similarly, the absolute value of the partial discharge signal feature, 15, divided by 0.79, is approximately 18.99. If the absolute value of its partial derivative is 0.06, then 18.99 × 0.06 ≈ 1.14, which is the anomaly factor index of this partial discharge signal feature.

[0105] By taking the absolute value of the eigenvalues, calculating the L2 norm, and multiplying both by the absolute value of the partial derivatives, the generated anomaly factor index normalizes the contribution of different features, objectively reflects the degree of influence of each data feature on the probability of target failure, and provides a standardized basis for screening dominant anomaly factors.

[0106] Furthermore, step A400 in the method provided in this application embodiment includes:

[0107] A410: Construct a knowledge graph based on the induction and propagation relationships of physical component nodes, fault mode nodes, and feature nodes.

[0108] A420: Map the dominant anomaly factor indicator set to feature nodes in the knowledge graph, and map the fault type with the highest probability to the fault mode node in the knowledge graph.

[0109] A430: Starting from the feature node, trace back along the causal relationship edge of the knowledge graph to the root cause node to obtain the root cause path.

[0110] A440: Dynamically adjust the weight values ​​of each relation edge in the root cause path according to the trend evaluation parameters, and calculate the path score based on the adjusted edge weights.

[0111] A450: The highest-scoring path from the root cause node to the fault mode node is taken as the fault evolution chain, and the evolution time window is marked.

[0112] In one embodiment, firstly, a knowledge graph is constructed based on the induction and transmission relationships between the physical components, fault modes and characteristic nodes of the low-voltage switchgear, clarifying the causal relationships such as corresponding heat dissipation device failure (physical component) → temperature abnormality (characteristic) → insulation aging (fault mode).

[0113] Subsequently, the dominant abnormal factor index set is extracted from the collected operational data. If the temperature is continuously exceeded by 20%, it is mapped to the temperature abnormal feature node of the knowledge graph. At the same time, if the probability of insulation aging failure reaches 0.75 (the highest value) according to the fault prediction model, it is mapped to the fault mode node.

[0114] Next, starting from the temperature anomaly feature node, the root cause is traced backward along the causal relationship edges of the knowledge graph: following the inverse logic of inducement-conduction, the deduction leads to root cause nodes such as heat dissipation device failure. During this process, a trend evaluation parameter ΔR is introduced to dynamically adjust the edge weights: if ΔR indicates an increase in risk, such as a faster rate of temperature increase, and ΔR is positive and exceeds the threshold, then the induced edge weight of heat dissipation device failure → temperature anomaly increases, while the conductive edge weight of temperature anomaly → insulation aging decreases; if the risk decreases (ΔR is negative), then the induced edge weight decreases and the conductive edge weight increases, conforming to the influence of risk changes on edge weights.

[0115] When calculating the path score, for each traced path, the adjusted edge weights are multiplied by a time decay factor. The closer to the current time, the closer the decay factor is to 1, reflecting the time series effect. The total score is obtained by accumulating these factors.

[0116] Finally, the highest-scoring path from the root cause node (heat dissipation device failure) to the failure mode node (insulation aging) is selected as the failure evolution chain, and the time windows are marked: stage 1 is the root cause triggering period, that is, the current time is pushed forward by T1, such as 12 hours, corresponding to the development of heat dissipation device failure; stage 2 is the failure propagation period, the current time is pushed forward by T2, such as 8 hours, corresponding to the propagation of temperature abnormality to insulation aging.

[0117] By real-time correction of edge weights, embedding trend parameters into formulas, and amplifying emergency paths with gain coefficients, the problems of real-time performance, lag, and neglect of deterioration acceleration in static graphs are solved, enabling accurate analysis and prediction of fault evolution processes.

[0118] In summary, the low-voltage switchgear fault prediction method based on artificial intelligence provided in this application has the following technical effects:

[0119] This application acquires operational data from multiple sensor nodes within a low-voltage switchgear, preprocesses it to construct a multidimensional tensor structure with spatial topological coding, inputs the current tensor into a fault prediction model to obtain prediction results, combines an interpretable module to generate a set of dominant anomaly factor indicators, utilizes a fault causal knowledge graph to analyze the root cause path to determine the fault evolution chain, and finally generates interpretable early warning prediction information, thereby achieving accurate prediction of low-voltage switchgear faults and making the low-voltage switchgear fault prediction results more accurate and reliable. It achieves the technical effect of using deep learning technology to build a prediction model, and then efficiently processing multidimensional complex data of low-voltage switchgear and deeply mining key fault information to achieve accurate fault prediction.

[0120] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides an artificial intelligence-based low-voltage switchgear fault prediction system, the system comprising:

[0121] The operation data acquisition module 1 is used to acquire operation data collected by multiple sensor nodes in the low-voltage cabinet. The operation data includes at least current, voltage, temperature, partial discharge signal, acoustic signal, and switch status parameters.

[0122] Runtime data preprocessing module 2 is used to preprocess the running data and construct a multidimensional tensor structure with spatial topological coding. The preprocessing includes normalization, anomaly removal, and time alignment.

[0123] The prediction result acquisition module 3 is used to input the current tensor into the fault prediction model and output the prediction result. The prediction result includes the probability of fault type and trend evaluation parameters. Based on the prediction result, the result is analyzed by the interpretable module to generate a set of dominant anomaly factor indicators.

[0124] The fault evolution chain acquisition module 4 determines the fault evolution chain by analyzing the root cause path of the current prediction result based on the fault type probability, trend evaluation parameters, and dominant abnormal factor index set.

[0125] The early warning and prediction information acquisition module 5 is used to generate interpretable early warning and prediction information based on the dominant abnormal factor index and the fault evolution chain.

[0126] Furthermore, the runtime data preprocessing module 2 is used to perform the following steps:

[0127] Based on the spatial relationship of sensor nodes, a first-dimensional encoding is established; based on the monitoring timestamps of the operational data, a time window is set, and a second-dimensional encoding is established for the data sequence within the time window; based on the multimodal feature channels of the operational data, a third-dimensional encoding is established; the first-dimensional encoding, the second-dimensional encoding, and the third-dimensional encoding are integrated to construct a three-dimensional tensor structure.

[0128] Furthermore, the prediction result acquisition module 3 is used to perform the following steps:

[0129] A three-dimensional convolutional kernel is used to slide along the first dimension of the tensor to extract local spatial correlation features of sensor nodes; a dilated convolution is used in the second dimension of the tensor to expand the temporal receptive field and extract long-period dependency features; a node adjacency matrix is ​​constructed based on electrical connection relationships, and the electrical correlation features between nodes are aggregated through a graph attention network; the local spatial correlation features, long-period dependency features, and electrical correlation features are concatenated, and the fault type probability and trend evaluation parameters are output through a fully connected layer.

[0130] Furthermore, the prediction result acquisition module 3 is used to perform the following steps:

[0131] A model architecture is constructed, including a spatiotemporal convolutional layer, a graph structure generator, a graph attention network, a feature concatenation layer, and a fully connected layer. Historical operating data is collected to construct a sample set, each sample including a three-dimensional tensor, a fault type label, and a trend change label. The sample set is divided according to a preset ratio to obtain a training set, a validation set, and a test set. The model architecture is trained, validated, and tested using the training set, validation set, and test set to obtain the fault prediction model.

[0132] Furthermore, the prediction result acquisition module 3 is used to perform the following steps:

[0133] The local spatial correlation features, long-period dependency features, and electrical correlation features are concatenated according to the node dimension and compressed through the convolutional layer compression channel. The compressed features are then expanded into vectors. The expanded vectors are input into a two-branch fully connected layer, which generates fault type probability parameters through the fault type probability branch and trend evaluation parameters through the trend evaluation parameter branch, and outputs the prediction results.

[0134] Furthermore, the prediction result acquisition module 3 is used to perform the following steps:

[0135] The fault type probability branch outputs the fault type probability distribution through the Softmax activation function to obtain the fault type probability parameter; the trend evaluation parameter branch outputs the trend evaluation scalar parameter through the ELU activation function and the linear layer to obtain the trend evaluation parameter.

[0136] Furthermore, the prediction result acquisition module 3 is used to perform the following steps:

[0137] The fault type with the highest probability in the prediction results is taken as the target probability; the partial derivative of each running data feature with respect to the target probability is calculated using automatic differentiation technology, where the partial derivative represents the change in target probability caused by a unit change in feature; based on the partial derivatives of each data feature, the contribution of each data feature is normalized to generate anomaly factor indicators; a dynamic threshold is set according to the fault type, and the anomaly factor indicators are screened using the dynamic threshold to obtain the set of dominant anomaly factor indicators.

[0138] Furthermore, the prediction result acquisition module 3 is used to perform the following steps:

[0139] Take the absolute value of the feature value of the running data; calculate the L2 norm of the probability of all fault types, where the L2 norm is the square root of the sum of squares of each probability; divide the absolute value of the feature by the L2 norm, and then multiply it by the absolute value of the partial derivative to generate the anomaly factor index.

[0140] Furthermore, the fault evolution chain acquisition module 4 is used to perform the following steps:

[0141] A knowledge graph is constructed based on the induced transmission relationships of physical component nodes, fault mode nodes, and feature nodes. The dominant anomaly factor index set is mapped to the feature nodes in the knowledge graph, and the fault type with the highest probability is mapped to the fault mode node in the knowledge graph. Starting from the feature node, the root cause path is obtained by tracing back along the causal relationship edges of the knowledge graph. The weight values ​​of each relation edge in the root cause path are dynamically adjusted according to the trend evaluation parameters, and the path score is calculated based on the adjusted edge weights. The highest-scoring path from the root cause node to the fault mode node is taken as the fault evolution chain, and the evolution time window is marked.

[0142] The low-voltage switchgear fault prediction system based on artificial intelligence provided in this embodiment of the invention can execute the low-voltage switchgear fault prediction method based on artificial intelligence provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0143] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0144] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A low-voltage switchgear fault prediction method based on artificial intelligence, characterized in that, include: The system acquires operational data collected by multiple sensor nodes within the low-voltage switchgear, including at least current, voltage, temperature, partial discharge signal, acoustic signal, and switch status parameters. The running data is preprocessed to construct a multidimensional tensor structure with spatial topological coding, wherein the preprocessing includes normalization, anomaly removal, and time alignment. The current tensor is input into the fault prediction model, and the prediction results are output. The prediction results include fault type probability and trend evaluation parameters. The results are then analyzed by an interpretable module to generate a set of dominant anomaly factor indicators. Based on the fault type probability, trend evaluation parameters, and dominant anomaly factor index set, the root cause path of the current prediction result is analyzed using the fault causal knowledge graph to determine the fault evolution chain. Based on the dominant anomaly factor indicators and the fault evolution chain, interpretable early warning and prediction information is generated.

2. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 1, characterized in that, The construction of a multidimensional tensor structure with spatial topological encoding includes: Establish the first dimension encoding based on the spatial relationship of the sensor nodes; Based on the monitoring timestamps of the operational data, a time window is set, and a second-dimensional code is established for the data sequence within the time window; A third-dimensional encoding is established based on the multimodal feature channels of the runtime data; By integrating the first-dimensional encoding, the second-dimensional encoding, and the third-dimensional encoding, a three-dimensional tensor structure is constructed.

3. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 2, characterized in that, Input the current tensor into the fault prediction model and output the prediction results, including: A three-dimensional convolution kernel is used to slide along the first dimension of the tensor to extract local spatial correlation features of sensor nodes; Dilated convolution is used in the second dimension of the tensor to expand the temporal receptive field and extract long-period dependent features. A node adjacency matrix is ​​constructed based on electrical connection relationships, and the electrical association features between nodes are aggregated through a graph attention network; The local spatial correlation features, long-period dependency features, and electrical correlation features are spliced ​​together, and the fault type probability and trend evaluation parameters are output through the fully connected layer.

4. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 3, characterized in that, Input the current tensor into the fault prediction model, which previously included: The model architecture is constructed, including spatiotemporal convolutional layers, graph structure generators, graph attention networks, feature concatenation layers, and fully connected layers; Historical operational data is collected to construct a sample set. Each sample includes a three-dimensional tensor, a fault type label, and a trend change label. The sample set is then divided according to a preset ratio to obtain a training set, a validation set, and a test set. The model architecture is trained, validated, and tested using the training set, validation set, and test set to obtain the fault prediction model.

5. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 3, characterized in that, By splicing together the local spatial correlation features, long-period dependency features, and electrical correlation features, the fault type probability and trend evaluation parameters are output through the fully connected layer, including: The local spatial correlation features, long-period dependency features, and electrical correlation features are concatenated according to the node dimension and compressed through the convolutional layer compression channel. Expand the compressed features into vectors; The expanded vector is input into a two-branch fully connected layer. Fault type probability parameters are generated through the fault type probability branch, and trend evaluation parameters are generated through the trend evaluation parameter branch. The prediction results are then output.

6. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 5, characterized in that, Fault type probability parameters are generated through the fault type probability branch, and trend evaluation parameters are generated through the trend evaluation parameter branch, including: The fault type probability branch outputs the fault type probability distribution through the Softmax activation function to obtain the fault type probability parameters; The trend evaluation parameter branch obtains the trend evaluation parameters by using the ELU activation function and the linear layer to output the trend evaluation scalar parameters.

7. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 5, characterized in that, Based on the prediction results, the results are analyzed through an interpretable module to generate a set of dominant anomaly factor indicators, including: The fault type with the highest probability in the prediction results is taken as the target probability; The partial derivative of each running data feature with respect to the target probability is calculated using automatic differentiation technology. The partial derivative represents the change in target probability caused by a unit change in feature. Based on the partial derivatives of each data feature, a normalized contribution is performed on each data feature to generate an anomaly factor index. A dynamic threshold is set according to the fault type, and the abnormal factor indicators are filtered using the dynamic threshold to obtain the set of dominant abnormal factor indicators.

8. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 7, characterized in that, Based on the partial derivatives of each data feature, a normalized contribution is performed on each data feature to generate anomaly factor indices, including: Take the absolute value of the feature value of the running data; Calculate the L2 norm of the probabilities of all fault types, where the L2 norm is the square root of the sum of the squares of the probabilities; The absolute value of the feature is divided by the L2 norm, and then multiplied by the absolute value of the partial derivative to generate the anomaly factor index.

9. The low-voltage switchgear fault prediction method based on artificial intelligence according to claim 1, characterized in that, Based on the aforementioned fault type probability, trend evaluation parameters, and dominant anomaly factor indicator set, the root cause path of the current prediction result is analyzed using a fault causal knowledge graph to determine the fault evolution chain, including: A knowledge graph is constructed based on the induction and transmission relationships of physical component nodes, failure mode nodes, and feature nodes; Map the dominant anomaly factor index set to feature nodes in the knowledge graph, and map the fault type with the highest probability of fault type to fault mode nodes in the knowledge graph. Starting from the feature node, trace back along the causal relationship edge of the knowledge graph to the root cause node to obtain the root cause path; The weight values ​​of each relation edge in the root cause path are dynamically adjusted according to the trend evaluation parameters, and the path score is calculated based on the adjusted edge weights. The highest-scoring path from the root cause node to the fault mode node is taken as the fault evolution chain, and the evolution time window is marked.

10. A low-voltage switchgear fault prediction system based on artificial intelligence, characterized in that, The system is used to implement the artificial intelligence-based low-voltage switchgear fault prediction method according to any one of claims 1-9, the system comprising: The operation data acquisition module is used to acquire operation data collected by multiple sensor nodes in the low-voltage cabinet. The operation data includes at least current, voltage, temperature, partial discharge signal, acoustic signal, and switch status parameters. The runtime data preprocessing module is used to preprocess the runtime data and construct a multidimensional tensor structure with spatial topological coding. The preprocessing includes normalization, anomaly removal, and time alignment. The prediction result acquisition module is used to input the current tensor into the fault prediction model and output the prediction result. The prediction result includes the probability of fault type and trend evaluation parameters. Based on the prediction result, the result is analyzed by the interpretable module to generate a set of dominant anomaly factor indicators. The fault evolution chain acquisition module, based on the fault type probability, trend evaluation parameters, and dominant anomaly factor index set, uses a fault causal knowledge graph to analyze the root cause path of the current prediction result and determine the fault evolution chain. The early warning and prediction information acquisition module is used to generate interpretable early warning and prediction information based on the dominant abnormal factor indicators and the fault evolution chain.

Images