Low-voltage transformer area examination meter practical training device and fault simulation control method

Abstract

The invention discloses a low-voltage transformer area examination meter practical training device and a fault simulation control method, and relates to the technical field of electric power practical training. The device aims at the problems of single scene, extensive evaluation and low skill improvement efficiency in traditional electric power practical training. The meter simulation module is compatible with various meters and is matched with a dual-communication protocol to realize efficient data transmission; the intelligent interaction module provides high-quality interaction experience through a high-definition touch screen and low-delay transmission; the data generation module can quickly switch customized scenes, and the fault simulation module can accurately inject various faults; the training management module depends on advanced technologies of knowledge maps, deep learning and fuzzy comprehensive evaluation to realize full-process closed-loop management from project design and operation guidance to evaluation feedback and intelligent optimization; all the modules cooperate efficiently, and the professionality, the accuracy and the efficiency of power failure troubleshooting and operation and maintenance skill training of trainees are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power skills training, and more particularly to a low-voltage substation area assessment meter training device and a fault simulation control method. Background Art

[0002] The low-voltage substation area assessment meter, as the core equipment for power metering and monitoring at the end of the power system, is an important foundation for ensuring accurate power metering, user electricity consumption analysis, and safe operation of the power grid. The low-voltage substation area assessment meter undertakes the tasks of accurately measuring, collecting, and transmitting power parameters in the low-voltage substation area, which is directly related to the stability of power supply, the accuracy of metering, and the user's electricity consumption experience. In the process of cultivating electric power skills talents, it is a necessary ability for practitioners to master the operation, maintenance, and fault handling skills of low-voltage substation area assessment meters proficiently. However, the existing meter training mode mainly relies on traditional physical meters combined with simple fault settings, which has problems such as single training scenarios, limited fault types, and difficulty in simulating complex operating conditions. Under the traditional training method, trainees can only come into contact with a few common faults, and lack comprehensive training on meters under different loads, environments, and system abnormalities, resulting in trainees being unable to diagnose and handle complex and changeable meter faults quickly and accurately in actual work. With the rapid development of the intelligence and digitization of the power system, the skill requirements for meter operation and maintenance personnel are constantly increasing. The traditional training methods are difficult to meet the needs of cultivating high-quality electric power skills talents. There is an urgent need for a training device and control method that can simulate diverse scenarios, achieve accurate fault simulation, and scientific evaluation, so as to improve the practical operation ability and fault response level of trainees and ensure the reliable operation of the power system. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the present invention discloses a low-voltage substation area assessment meter training device and a fault simulation control method, which can optimize the low-voltage substation area assessment meter training device to meet user needs.

[0004] The present invention adopts the following technical solutions: A low-voltage substation area assessment meter training device includes a device body. The device body is a rectangular main frame built with high-strength aluminum alloy. The tensile strength of the aluminum alloy material is ≥200 MPa, and the yield strength is ≥150 MPa. Adjustable support feet and universal wheels are provided at the bottom of the main frame; the height adjustment range of the adjustable support feet is 0 - 50 mm, and the load-bearing capacity of the universal wheels is ≥100 kg; a lighting module is provided at the top of the main frame, the brightness of the lighting module is ≥500 lumens, and the irradiation angle can be adjusted by ±45°; A ventilation and heat dissipation structure is provided on the front of the main frame, using a silent fan with a diameter of 120 mm, the air volume is ≥50 CFM, the noise is ≤35 dB, and a dust-proof net is provided at the air inlet of the heat dissipation structure, and the filtration efficiency is ≥90%; On the front of the main frame, a multi-layer detachable mounting plate is provided with a meter simulation module and an intelligent interaction module; the mounting plate is made of 6061-T6 aluminum alloy with a thickness of 3 mm, and the detachable structure supports quick replacement within 5 seconds; Inside the main frame, a data generation module, a fault simulation module, and a training management module are provided: The meter simulation module is compatible with the assessment meters of single-phase energy meters, three-phase energy meters, and intelligent water meters. The simulated measurement accuracy reaches 0.5S level, the data acquisition frequency is ≥100Hz, and it supports the communication protocol transmission functions of RS485 and LoRa; The intelligent interaction module is equipped with a 10.1-inch high-definition touch screen with a resolution of 1920×1200. It displays training data and operation results through a graphical interface, supports a virtual simulation operation delay of ≤200ms, a remote collaboration video transmission frame rate of ≥25fps, and realizes a data interaction rate of ≥100Mbps between the internal modules of the device, and an instruction transmission delay of ≤100ms; The data generation module generates electrical parameter data of voltage, current, and power for testing based on the load characteristics of the actual power distribution area, supports ≥10 customized test scenarios, and the scenario switching time is ≤5 seconds; The fault simulation module can set fault types such as open circuit, short circuit, wiring error, data anomaly, and leakage current. The fault setting response time is ≤1 second, which is used to train the fault troubleshooting ability of trainees, and the fault recurrence accuracy is ≥95%; The training management module is responsible for the creation, assignment, and progress tracking of training projects, including trainee operation records, performance evaluation, and skill assessment. The training management module includes a project design unit, an operation guidance unit, a data collection unit, an evaluation and feedback unit, and an intelligent optimization unit. The project design unit constructs a standardized training project template based on the typical substation topology structure and fault case library using a knowledge graph, and the generation time of a single template is ≤ 10 minutes. The operation guidance unit provides a standardized operation process and safety specification guidance, and the guidance response delay is ≤ 500 ms. The data collection unit models the trainee operation trajectory data collected by the data collection unit through a deep learning algorithm, the trajectory sampling frequency is ≥ 10 Hz, and the modeling accuracy is ≥ 90%. The evaluation and feedback unit uses the fuzzy comprehensive evaluation method to score the trainee operation in multiple dimensions, analyzes the skill shortboards in combination with an improved Bayesian network driven by cognition, and generates a personalized improvement plan, and the plan generation time is ≤ 3 minutes. The intelligent optimization unit dynamically adjusts the training difficulty and fault types based on the reinforcement learning algorithm, the difficulty adjustment step size is ≥ 5 levels, realizes the adaptive optimization of the training project, and the optimization period is ≤ 10 minutes. The output end of the project design unit is connected to the input end of the operation guidance unit, the output end of the data collection unit is connected to the input end of the evaluation and feedback unit, the project design unit is interconnected with the intelligent optimization unit, and the output end of the evaluation and feedback unit is connected to the input end of the operation guidance unit.

[0005] The meter simulation module is interconnected with the training management module; the data generation module is interconnected with the training management module; the fault simulation module is interconnected with the training management module; the intelligent interaction module is interconnected with the training management module.

[0006] As a further embodiment of the present invention, the knowledge graph includes a knowledge extraction unit, a knowledge fusion unit, and a knowledge reasoning unit. Each unit collaborates through data flow and algorithms to construct an intelligent knowledge system in the field of power training. The knowledge extraction unit uses the BERT-Transformer pre-training model combined with the CRF sequence labeling algorithm to extract entities such as transformers, lines, and meters from the typical substation topology structure documents and fault case reports with an accuracy of ≥ 92%. At the same time, it identifies 12 types of relationships including "connected to" and "fault triggered", and the text volume processed per second can reach 200 KB, quickly forming initial knowledge triples. For professional terms in the power field, more than 2,000 professional dictionaries are built in to ensure the accuracy of entity extraction. The knowledge fusion unit uses an entity alignment algorithm based on cosine similarity and a rule-based relationship matching strategy to eliminate the ambiguity of multiple representations of the same entity by setting a similarity threshold of 0.85, and at the same time solve the knowledge conflict problem between different data sources; adopts block indexing to improve the fusion efficiency of millions of knowledge triples by 40%, and constructs a unified knowledge graph ontology architecture containing 10 major categories and 500+ attributes, supporting OWL semantic description and SPARQL query; The knowledge reasoning unit adopts a hybrid reasoning mode that combines the rule engine Drools and the graph neural network GNN. It performs rapid logical deduction through more than 500 custom rules in the power field. At the same time, it uses the GNN model to perform deep learning on device operation parameters and more than 5,000 historical fault data to mine potential knowledge relationships with an accuracy of ≥88%; the inferred potential faults can generate a visual analysis report within 30 seconds, providing extended knowledge support for the construction of training project templates, including 15 types of fault scenarios and 300+ knowledge nodes, effectively improving the richness and practicality of training content; The output end of the knowledge extraction unit is connected to the input end of the knowledge fusion unit, and the output end of the knowledge fusion unit is connected to the input end of the knowledge reasoning unit.

[0007] As a further embodiment of the present invention, the data acquisition unit uses an improved deep learning algorithm model to model the operation trajectory of the trainee. The improved deep learning algorithm model includes a spatio-temporal sequence analysis module, a three-dimensional space feature analysis module, a multi-data fusion module, a convolutional neural network module, and a data encoding output module. Each module works together through data flow and feature transformation matrices to achieve high-precision modeling of the trainee's operation trajectory; The spatio-temporal sequence analysis module adopts an architecture that combines a bidirectional gated recurrent unit and an attention mechanism to model the operation time series data of the trainee; the spatio-temporal sequence analysis module sets 10 parallel time windows with a time span from 50 milliseconds to 5 seconds, and dynamically assigns the importance of different time scales through a time attention weight matrix with a dimension of 10×128, giving a 30% weight increase to the temporal features in the 10 seconds before the fault operation. For continuous operations, the dynamic time warping algorithm is used to calculate the operation similarity, and the matching threshold is set to 0.85 to ensure robust recognition of similar operation patterns; The three-dimensional space feature analysis module is based on the PointNet++ architecture and processes three-dimensional space coordinate data with a sampling frequency of 100 Hz and a position accuracy of ±2 mm during the processing operation. The three-dimensional space feature analysis module introduces prior knowledge constraints of power equipment, constructs a spatial relationship graph containing 50 types of power equipment, and extracts topological features between equipment through a graph convolutional network. For the operation of switchgear, a local feature descriptor based on the cylindrical coordinate system is specially designed, and the operation trajectory is decomposed into three-dimensional motion components with a radial accuracy of ±1 mm, an angular accuracy of ±0.5°, and an axial accuracy of ±1 mm. The multi-data fusion module adopts a fusion strategy combining tensor decomposition and Kalman filtering to process multi-source heterogeneous data from a touch screen with a resolution of 1920×1080, an action capture camera with a frame rate of 60 fps, and a force feedback device with a sampling rate of 200 Hz. The multi-data fusion module designs a modal reliability evaluation mechanism, dynamically adjusts the fusion weight by calculating the confidence score of modal data within the range of 0 to 1. When the confidence of a certain modality is lower than 0.6, the data completion algorithm is automatically started, and interpolation processing is performed using historical data and equipment status knowledge. The convolutional neural network module adopts a multi-scale dilated convolution architecture, including three parallel convolutional branches with dilation rates set to 2, 4, and 8 respectively. The effective receptive field can cover 5 to 40 time steps. For the key feature points in the operation trajectory, the convolutional neural network module designs a feature enhancement unit to strengthen important feature channels and suppress irrelevant information through a channel attention mechanism. When the convolutional neural network module runs on the NVIDIA Jetson AGX Xavier hardware platform, the inference delay is controlled within 15 milliseconds, meeting the real-time requirement. The data encoding and output module encodes the processed multi-dimensional features into a fixed-length operation vector with a dimension of 256, and uses a self-supervised learning framework to learn the semantic representation of the operation. The data encoding and output module designs a coding quality evaluation mechanism based on contrastive learning. By calculating the cosine similarity of positive and negative sample pairs, it ensures that the distance between similar operations in the coding space does not exceed 0.3, and the distance between dissimilar operations is not less than 0.7. The finally output operation vector supports downstream tasks, achieving performance indicators with an operation classification accuracy of over 95%, a fault early warning accuracy of over 85% 30 seconds in advance, and a skill evaluation F1-score of over 0.9. The output end of the spatio-temporal sequence analysis module is connected to the input end of the multi-data fusion module; the output end of the three-dimensional space feature analysis module is connected to the input end of the multi-data fusion module; the output end of the multi-data fusion module is connected to the input end of the convolutional neural network module; the output end of the convolutional neural network module is connected to the input end of the data encoding output module. As a further embodiment of the present invention, the convolutional neural network module includes a data preprocessing module, a spatio-temporal feature extraction module, an attention mechanism module, a model training module, and a prediction evaluation module. Each module realizes end-to-end intelligent analysis of power data through a standardized data flow interface and a feature enhancement algorithm chain; The data preprocessing module designs a multi-modal data normalization pipeline. For image data with a resolution of not less than 1024×1024, zero-mean normalization processing is adopted, setting the data mean to 0 and the standard deviation to 1, and combining the CLAHE contrast-limited adaptive histogram equalization technology to expand the pixel dynamic range to the interval of 0 to 255; for time-series data with a sampling frequency of 50Hz or above, normalization is performed through a sliding window with a length of 100 points and a step size of 10 points, screening and retaining the valid values within the 95% quantile of the data to effectively suppress abnormal pulse interference; for audio data with a sampling rate of 44.1kHz, it is converted into an 80-dimensional Mel spectrogram using a Mel filter bank, and a random gain jitter of -6dB to +6dB is added to enhance the robustness of the model to different audio signals; the data preprocessing module supports parallel processing of 32-channel data, and the single-batch processing delay is controlled within 50ms; The spatio-temporal feature extraction module constructs a four-branch dilated convolutional network architecture, and the dilation rates of the four branches are set to 1, 3, 5, and 7 respectively, corresponding to realizing the receptive field coverage in the pixel range of 1×1 to 15×15; each branch contains a triple structure: first is a double residual block, designed with a bottleneck structure, the channel compression ratio is 4:1, and cross-layer connection and weight normalization technologies are embedded to improve the stability of the training process while reducing the computational complexity; second is a multi-scale pooling layer, which parallelly executes three pooling operations of 1×1, 3×3, and 5×5, and dynamically generates weights through a gating mechanism to realize the adaptive fusion extraction of multi-granularity features; finally is a feature calibration layer, which based on the learned affine transformation parameters γ and β, uniformly adjusts the brightness and contrast of cross-branch features to ensure the consistency of the feature space; when this module runs on an NVIDIA A100 GPU, the floating-point operation volume reaches 12.8G, and the feature extraction frame rate can be as high as 200FPS; The attention mechanism module designs a three-dimensional attention fusion framework, covering attention mechanisms in three dimensions: spatial, channel, and temporal. Spatial attention: Generates a spatial attention map through 7×7 convolution operations, precisely focusing on the key areas of power equipment, such as transformer windings and circuit breaker contacts, and uses a soft mask mechanism to suppress background noise interference. The generated attention map has the same resolution as the input feature map. Channel attention: Builds the dependence relationship between channels by means of global average pooling operation and combines two fully connected networks. For channels reflecting electrical parameters, such as channels related to voltage and current, a weight gain of 1.2 times or more is given. Temporal attention: For long-sequence data with a time span of 30 minutes or more, a Transformer encoder structure is introduced, and the 8-head multi-head self-attention mechanism is used to capture the dependence relationship across time points, supporting the processing of temporal windows with a maximum of 5000 time steps. The overall parameter ratio of this attention module does not exceed 15%, but it can increase the feature discriminability by more than 20%. The model training module adopts a mixed-precision training strategy, realizing the dynamic switching between FP16 and FP32 data precisions based on the PyTorch framework. On the premise of ensuring no loss of precision, the video memory occupancy is reduced by 40%. During the training process, the AdamW optimizer and the SGD optimizer are combined. The initial learning rate of AdamW is set to 1e-4, the weight decay value is 0.01, the momentum of SGD is set to 0.9, and the learning rate is adjusted with a cosine annealing learning rate decay strategy at a period of 200 epochs to improve the stability of model convergence. The data augmentation strategy is customized according to the data type: for image data, random rotation, Gaussian blur, and CutOut augmentation methods are used; for temporal data, random time offset, amplitude scaling, and injecting noise with a signal-to-noise ratio not lower than 20 dB are performed; for audio data, time stretching, pitch transformation, and adding background noise from the substation environment noise library are implemented. This module supports distributed training with a maximum of 8 cards synchronized, and the single-card training efficiency can be increased to more than 75%. The prediction and evaluation module constructs a multi-task output head, supporting three core task types: classification, regression, and anomaly detection. Classification task: Adopts the cross-entropy loss function, and the number of neurons in the output layer is set according to the number of fault types, not less than 20 classes, and the classification accuracy on the standard test set reaches 98% or more. Regression task: Uses the smooth L1 loss function to output the state parameters of power equipment, such as temperature and load rate, and the prediction error is controlled within ±3% of the true value. Anomaly detection: Builds a boundary model based on One-Class SVM, and measures the degree of anomaly in the feature space through the Mahalanobis distance. The F1-score of anomaly detection is not less than 0.95. In addition, the evaluation index real-time visualization module supports generating more than 10 types of charts such as confusion matrices, ROC curves, and error distribution histograms, and the result update delay is controlled within 1 second. The output end of the data preprocessing module is connected to the input end of the spatio-temporal feature extraction module; the output end of the spatio-temporal feature extraction module is connected to the attention mechanism module; the output end of the attention mechanism module is connected to the input end of the model training module; the output end of the model training module is connected to the input end of the prediction and evaluation block.

[0008] As a further embodiment of the present invention, the fuzzy comprehensive evaluation method is designed based on a modular architecture, and includes an index system construction module, a weight calculation module, a fuzzy relation generation module, and a comprehensive evaluation operation module. Each module works together to achieve a quantitative evaluation of the training performance of students; The index system construction module is responsible for building a scientific evaluation index system to measure the training performance of students from four core dimensions; in the dimension of operation standardization, by verifying 8 sub-indexes of the integrity of operation steps and the compliance with safety specifications, the operation compliance of students is evaluated; in the dimension of fault diagnosis accuracy, based on 6 sub-indexes of the correct rate of fault type identification and the deviation of fault point location, the diagnosis accuracy is determined; in the dimension of fault handling efficiency, based on 4 time parameters of fault response time and repair time consumption, the processing speed is quantified; in the dimension of data interpretation ability, through 5 sub-indexes of the depth of data anomaly analysis and the accuracy of parameter correlation reasoning, the analysis ability of students for power data is evaluated; finally, a multi-level evaluation index system including 23 sub-indexes is constructed to fully cover the training assessment requirements; The weight calculation module uses the Analytic Hierarchy Process (AHP) to determine the weights of each index; a group consisting of 5 power domain experts and 3 education evaluation experts constructs a judgment matrix according to the 1-9 scale method to score the relative importance of each index pairwise; by calculating the maximum eigenvalue of the matrix and the corresponding eigenvector, the initial weight vector is obtained, and a consistency test is performed to ensure the logical rationality of the judgment matrix; if the consistency ratio CR > 0.1, the scoring is adjusted again; when CR ≤ 0.1, the final weight vector W = [w1, w2, w3, w4] is determined, where w1, w2, w3, w4 correspond to the weights of operation standardization, fault diagnosis accuracy, fault handling efficiency, and data interpretation ability respectively, so as to realize differential weighting for different assessment dimensions; The fuzzy relation generation module divides the training performance of students into five evaluation levels: excellent, good, medium, passing, and failing; the evaluation group consists of training tutors and technical backbones, and scores according to the records of students' operation processes and the materials of fault handling results; for each evaluation index, the proportion of the number of students whose performance belongs to each level is counted to form membership degree vectors R1, R2, R3, R4; for example, for the operation standardization index, if 20% of the students are rated as excellent, the membership degree of the excellent level in R1 is 0.2; the membership degree vectors of the four indexes are arranged in rows to construct a 4×5 dimensional fuzzy relation matrix R, which intuitively presents the correlation degree between students' performance and each evaluation level; The comprehensive evaluation operation module calculates the comprehensive evaluation vector B through matrix multiplication operation B = W × R. This vector contains the comprehensive membership degrees of the trainee's performance belonging to five evaluation levels; normalizes the vector B to make the sum of all elements equal to 1 to enhance the comparability of the results; selects the level corresponding to the element with the largest value in B as the trainee's comprehensive evaluation level according to the principle of maximum membership degree; at the same time, assigns corresponding score intervals to excellent (90 - 100 points), good (80 - 89 points), medium (70 - 79 points), pass (60 - 69 points), and fail (0 - 59 points), and calculates the weighted comprehensive score in combination with the membership degree to achieve accurate quantitative evaluation of the trainee's training performance and provide data support for skill improvement; The output end of the index system construction module is connected to the input end of the weight calculation module; the output end of the weight calculation module is connected to the input end of the fuzzy relationship generation module; the output end of the fuzzy relationship generation module is connected to the input end of the comprehensive evaluation operation module; As a further embodiment of the present invention, the improved Bayesian network adopts a modular architecture design, including a network structure optimization module, a parameter learning algorithm module, an inference mechanism enhancement module, a cognitive interpretation module, and a domain adaptation module. Each module cooperates with each other to achieve accurate power fault diagnosis and training evaluation; The network structure optimization module is responsible for constructing a hierarchical causal network architecture, which is divided into an equipment layer, a parameter layer, and a phenomenon layer; the equipment layer integrates 20 core power equipment nodes of transformers and meters, the parameter layer converges 15 real-time monitoring parameter nodes of voltage, current, and frequency, and the phenomenon layer integrates 8 fault characterization nodes of alarm information and abnormal sounds; by analyzing more than 1000 groups of expert-annotated data and more than 3000 historical fault cases, the dependence relationship between nodes at each layer is determined, and an accurate causal link is constructed; in view of the dynamic evolution characteristics of power faults, a dynamic Bayesian network architecture with a time window of 15 minutes is introduced, and 30 nodes are set in each time slice, which can completely capture the timing characteristics of the fault from occurrence to development and realize the dynamic modeling of the fault evolution process; The parameter learning algorithm module adopts a hybrid parameter estimation strategy to improve the model accuracy; for common fault nodes with rich data samples, such as voltage over-limit nodes, the maximum likelihood estimation method is used to calculate the conditional probability; for rare fault event nodes, such as equipment insulation aging, Bayesian estimation and Jeffreys prior distribution are adopted; in the learning process, an adaptive learning rate mechanism is introduced, and the weights of historical data decay according to an exponential law with a half-life of 30 days, and the weights of newly collected data are 20% higher than those of historical data, ensuring that the model can quickly adapt to new fault patterns and improve the learning ability for complex fault scenarios; The inference mechanism enhancement module improves the inference performance through evidence weighted fusion and reverse diagnosis; designs an evidence weighting mechanism to assign credibility weights to evidence from different sources. Among them, the infrared thermal imaging evidence can intuitively reflect the abnormal temperature of the device, and the weight is set to 0.85; considering the possible false alarms of the text alarm information, the weight is set to 0.7; the Dempster-Shafer theory is used to handle evidence conflicts. When multi-source evidence conflicts, effective fusion is carried out through the evidence synthesis rule; supports the reverse diagnosis inference mode, which can start from the fault phenomenon of tripping at the phenomenon layer and reverse-derive the posterior probability of faults at each node of the device layer. In complex fault scenarios, the fault source location accuracy can reach more than 92%, significantly improving the fault diagnosis efficiency; The cognitive explanation module has a knowledge base containing 120 typical power fault modes built in. Each fault mode is associated with standard operation steps, safety specification requirements, and repair plan documents; based on Bayesian inference, the C4.5 algorithm is used to generate a decision tree explanation path, converting the probability inference result into an intuitive rule expression such as "If the voltage continues to rise abnormally and the current exceeds the rated value, then there may be a line overload fault", making the diagnosis result conform to the reading and understanding habits of power industry technicians, and enhancing the interpretability and guidance of the diagnosis result; The domain adaptation module constructs a conditional probability table of typical radial and looped network structures based on the topological relationship of substation area equipment; fully considers the electrical distance and influence coefficient between devices. For example, the probability of adjacent device faults affecting each other is 30% higher than that of non-adjacent devices; for the uncertainty of power system operation, the probability interval is used to represent the fault possibility. For example, the fault probability of a certain device is expressed as an interval range from 0.65 to 0.73, which is more in line with the actual operation scenario of the power system compared with the precise point estimate, and improves the applicability of the model in the low-voltage substation area training; The output end of the network structure optimization module is connected to the input end of the parameter learning algorithm module; the output end of the parameter learning module is connected to the input end of the inference mechanism enhancement module; the output end of the inference mechanism enhancement module is connected to the input end of the cognitive explanation module; the input end of the cognitive explanation module is connected to the input end of the domain adaptation module. As a further embodiment of the present invention, the improved reinforcement learning model includes a spatio-temporal feature fusion module, a data encoding module, a spatio-temporal Bayesian network module, a dynamic channel switching module, and a multi-node interaction module; The spatio-temporal feature fusion module innovatively introduces a spatio-temporal attention mechanism based on Transformer, breaking the dependence on local features of traditional 3D-CNN; when processing the training data of low-voltage substation assessment meters, it captures the long-range dependence relationships between data at different time scales and different spatial positions through the multi-head attention mechanism in parallel, realizing the deep fusion of features across regions and time periods; at the same time, it embeds an adversarial learning framework, trains a discriminator to distinguish real data features from generated features, forcing the feature extractor to learn more discriminative representations, effectively solving the problem that meter data features are prone to confusion in complex electromagnetic environments; The data encoding module pioneered a dynamic semantic encoding system, integrating knowledge graphs and federated learning; for the data of low-voltage substation assessment meters, it uses a knowledge graph to construct a semantic association network among meter parameters, fault types, and operating environments, endowing deep semantic information for data encoding; combined with the federated learning framework, while protecting the data privacy of each training site, it collaboratively optimizes the encoding model, enabling the encoding parameters to adapt to the data distribution differences of different training scenarios; in addition, it designs a generative encoding strategy based on the variational autoencoder (VAE), which can generate virtual data samples during the training process, expanding the diversity of the training set and improving the generalization ability of the model; The spatio-temporal Bayesian network module breaks through the limitations of traditional static modeling, constructing a dynamic evolutionary Bayesian network architecture; introducing a meta-learning mechanism, it can quickly adjust the network structure and prior distribution of parameters under different training tasks and fault simulation scenarios, realizing the rapid adaptation of the model; combined with a causal discovery algorithm, it automatically mines the causal relationships between the meter data in the low-voltage substation, rather than relying only on statistical associations, enabling the network to more accurately predict the fault propagation path; at the same time, it integrates the principles of quantum computing into the Bayesian inference process, using quantum superposition states to accelerate the parallel calculation of multi-node probability distributions, greatly improving the inference efficiency in complex scenarios; The dynamic channel switching module proposes a dual-drive strategy based on deep reinforcement learning and digital twin; on the one hand, through the deep Q-network (DQN) combined with the experience replay mechanism, it learns the optimal channel switching strategy in the communication environment of the low-voltage substation; on the other hand, it constructs a virtual digital twin of the substation communication, pre-evaluates the effects of different channel switching actions in the twin environment, and generates virtual training samples to feed back the reinforcement learning model; in addition, it designs a spectrum sensing and blockchain integration scheme, uses spectrum sensing to monitor the channel state in real time, and constructs a trusted channel usage record ledger with blockchain technology to achieve dynamic spectrum sharing among multiple training devices, effectively improving the utilization rate of limited frequency band resources; The multi-node interaction module builds a collaborative system based on edge intelligence and holographic interaction; lightweight edge computing units are deployed at each meter node to achieve local preprocessing of data and interaction decision-making, reducing data transmission latency; holographic projection technology is used to present the node interaction relationships calculated by the graph neural network as a visual 3D dynamic atlas to assist trainees in intuitively understanding the complex coupling relationships between the nodes in the power distribution area; at the same time, an emotion computing model is introduced to dynamically adjust the node interaction difficulty and fault simulation strategy according to the feedback information of the trainees' expressions and voices during the training operation process to achieve a personalized training experience; in addition, a multi-agent interaction mechanism based on game theory is designed to simulate the strategic game of different interest subjects in the power distribution area management, enhancing the realistic mapping and challenge of the training content; The output end of the spatio-temporal feature fusion module is connected to the input end of the data encoding module; the output end of the data encoding module is connected to the input end of the spatio-temporal Bayesian network module; the output end of the spatio-temporal Bayesian network module is connected to the input end of the dynamic channel switching module; the output end of the dynamic channel switching module is connected to the input end of the multi-node interaction module.

[0009] As a further embodiment of the present invention, it includes the following steps: S1. Training scenario construction Based on the standardized training project template constructed by the knowledge graph, with the help of the 15 types of fault scenarios and more than 300 knowledge nodes output by the knowledge reasoning unit, combined with the 8 typical power distribution area topologies of radial and loop network types in the topology structure library of the project design unit, the training scenario parameters are automatically generated; the system can dynamically configure 5 mainstream meter models of single-phase and three-phase watt-hour meters, and set the initial operating state of 15 monitoring parameters covering voltage and current according to the historical operation data containing more than 3,000 cases; among them, the voltage range is 0 - 400V, and the accuracy reaches ±0.5%FS; the current range is 0 - 100A, and the accuracy is ±0.2%FS; the data generation module loads and generates specific training tasks containing equipment topology relationships, operating parameters and potential fault hazards at a scene switching speed of no more than 5 seconds, and at the same time displays the task details through the 10.1-inch high-definition touch screen with a resolution of 1920×1280 equipped by the intelligent interaction module; S2. Fault injection After receiving the training task instruction, the fault simulation module can respond quickly within 1 second and inject faults through two mechanisms: hardware control and signal interference. For hardware faults such as open circuits and short circuits, a solid-state relay is used to control the on / off of the circuit. For signal faults such as data anomalies and leakage current, an analog signal generator is used to output interference waveforms with an accuracy of ±2%. The system supports setting 8 types of fault types including wiring errors and equipment insulation aging, and the accuracy of fault reproduction is not less than 95%. While injecting faults, the intelligent interaction module displays the fault phenomena such as sudden changes in voltage curves and flashing alarm indicator lights with a visualization animation at a frame rate of not less than 25fps, and then cooperates with the abnormal sound of the equipment generated by the environmental audio encoder. Even in an environment with a signal-to-noise ratio of -5dB, the recognition rate of abnormal equipment sounds can reach more than 90%, so as to create a highly realistic fault environment; S3. Trainee operation and data collection When the trainee is operating, the data collection unit captures the three-dimensional space coordinates in real time at a sampling frequency of 100Hz, with a position accuracy of ±2mm, and combines a bidirectional gated recurrent unit and an attention mechanism to model the operation time series data. Multi-source data is collected through a touch screen with a resolution of 1920×1080, an action capture camera with a frame rate of 60fps, and a force feedback device with a sampling rate of 200Hz. After these data are fused by tensor decomposition and Kalman filtering, the confidence of each modality in the range of 0 to 1 is calculated. When the confidence of a certain modality is lower than 0.6, the data completion algorithm is automatically started, and interpolation processing is performed using historical data and equipment status knowledge. The convolutional neural network module processes the data with an inference delay of 15ms, extracts the operation trajectory features, and finally generates a multi-modal training dataset containing 256-dimensional operation vectors, completely recording the trainee's operation steps, equipment parameter changes, and the whole process of fault response; S4. Evaluation and feedback The evaluation and feedback unit uses the fuzzy comprehensive evaluation method to score the trainee from four dimensions: operation standardization and fault diagnosis accuracy. The operation standardization is evaluated by checking 8 sub-indicators, and the fault diagnosis accuracy is determined based on 6 sub-indicators. Through the analytic hierarchy process, the expert group constructs a judgment matrix, calculates the maximum eigenvalue and the corresponding eigenvector of the matrix to obtain the initial weight vector, and conducts a consistency test to ensure that the consistency ratio CR does not exceed 0.1, so as to determine the final weight vector W and divide the trainee's performance into five evaluation levels. At the same time, the improved Bayesian network, based on more than 1000 groups of expert-annotated data and historical fault cases, reversely deduces the skill shortboards in the trainee's operation and generates a diagnostic report with a fault location accuracy of 92%. Finally, the system generates an evaluation report containing quantitative scores, fault analysis, and improvement suggestions within 3 minutes, which is fed back to the trainee through the intelligent interaction module and simultaneously pushed to the teacher-end management system; S5. Training optimization The intelligent optimization unit is based on a reinforcement learning algorithm and dynamically adjusts the training difficulty according to the trainee evaluation results. The difficulty adjustment step size is not less than 5 levels, which is achieved by increasing the complexity of fault types, upgrading from a single short-circuit fault to a composite fault, and shortening the fault response time requirement. Combining the conditional probability table constructed by the domain adaptation module, it adjusts the distribution of fault types in subsequent training projects, increasing the probability of adjacent device associated faults by 30%. The entire optimization cycle is controlled within 10 minutes to ensure dynamic matching between the training content and the trainee's skill level. Through continuous iterative training, it continuously improves the trainee's fault diagnosis and handling capabilities during the operation and maintenance of low-voltage substation metering devices. As a further embodiment of the present invention, the steps of building the convolutional neural network are as follows: 1) Data preprocessing and format conversion: The collected trainee operation trajectory data, including operation timestamps, operation device numbers, operation instruction codes, and device status parameter vectors, are used to form a multivariate sequence set. For missing values, an interpolation method based on Gaussian process regression is used. The covariance matrix K is constructed using a kernel function, and the interpolation coefficients are obtained by solving a system of linear equations to achieve high-precision filling. Outlier detection introduces the local outlier factor algorithm. By calculating the ratio of the local reachability density of sample x to the average local reachability density of neighborhood samples, abnormal data points are identified and corrected. When normalizing the operation timestamps, Min-Max normalization combined with time series smoothing technology is used. An adversarial network is introduced, consisting of a generator G and a discriminator D. The generator G generates a normalized time series and the discriminator D judges the difference between the true normalized result and the pseudo result, as shown in the following formula: In formula (1), is a tuning parameter used to control the range of the normalized timestamp. The non-uniform interval sequence is converted into a uniform time interval sequence through cubic spline interpolation; is a hyperparameter for balancing the adversarial loss. Through this adversarial learning method, the normalization process is not only required to meet the basic numerical transformation requirements but also learn a more discriminative and robust normalization pattern in the adversarial game, effectively coping with noise and outlier interference in the data; A dynamic sliding event window is constructed, and the window size is controlled by the adaptive parameter . The operation trajectory sequence is converted into a three-dimensional tensor format , where N is the number of samples, T is the time step, M is the operation feature dimension, and D is the number of data channels. The mathematical expression is: , in the formula, is the feature transformation function for the Dth data channel, and Denote the operation data of the nth sample at time step t and feature dimension m; 2) Spatiotemporal feature extraction: Input the processed three-dimensional tensor X into the spatiotemporal feature extraction module, adopting a multi-scale parallel convolution architecture; for each time step, use convolution kernels of different sizes , where l represents the convolution layer number, k is the size of the convolution kernel in the time dimension, D is the number of input channels, is the number of output channels, and perform convolution operations; based on the meta-learning framework, enable the parameters of the convolution layer to be quickly and adaptively adjusted according to different task scenarios; during the training process, optimize the parameters of the meta-network through the meta-learning algorithm, so that the generated convolution layer parameters are more adaptable to different types of training data. During the training process, optimize the parameters of the meta-network through the meta-learning algorithm, so that the generated convolution layer parameters are more adaptable to different types of training data. The improved calculation process is as follows: introduce a gating mechanism, and the output feature map of the lth convolution layer is calculated as follows: In formula (2), is the Sigmoid activation function, is the gating signal tensor represents element-wise multiplication, is the convolution kernel weight, is the bias term, is the pooling weight; expand the receptive field through the dilated convolution layer, and combine the residual connection mechanism to solve the problem of gradient disappearance; are the parameters of the meta-network. In this way, the network can quickly adjust the convolution calculation method in different training scenarios of low-voltage distribution network meter reading, and improve the generalization ability of the model; in the pooling stage, adopt the adaptive weighted pooling method to dynamically adjust the pooling weight according to the local method of the feature map; 3) Application of the attention mechanism: Input the extracted spatiotemporal feature map Y into the attention mechanism module, adopting the multi-head self-attention mechanism; introduce the spatiotemporal dynamic weight matrix , which is composed of the temporal dependence coefficient in the time dimension and the feature correlation coefficient in the space dimension. By capturing the continuity of the trainee's operation trajectory in the time series and the correlation of the spatial layout of the operation equipment, dynamically adjust the calculation of the attention weight. For each head h, calculate the attention weight using the following formula: In formula (3), are the query vector and the key vector respectively, obtained through linear transformation, is the dimension of the key vector; can be obtained by calculating the reciprocal of the operation time interval and the autocorrelation of the time series, Determined according to the physical connection relationship and functional relevance between operating devices; this method enables more precise focus on key operation links when processing the training data of low-voltage substation assessment meters. For example, in the fault handling process, higher weights are assigned to operation steps with sequential dependencies, and the multi-headed attention outputs are concatenated and linearly transformed to obtain a feature vector with attention; 4) Model training and parameter optimization: The feature vector Z with attention is input into the fully connected layer, and a hybrid loss function L is adopted, combining cross-entropy loss , mean squared error loss , and regularization term . Combining the attention mechanism, dynamic attention weights are introduced for each loss sub-term to construct an attention weight matrix . By calculating the importance of different tasks in the current training stage, weights are dynamically generated; the formula is: In formula (4), is the weight parameter , which can be calculated through the self-attention mechanism based on the change rate of task loss and the uncertainty factor of prediction results, realizing collaborative optimization between multiple tasks and improving the comprehensive performance of the model in the low-voltage substation assessment meter training task; 5) Prediction and evaluation: The trained model is used to predict new trainee operation trajectory data; for classification tasks, the probability distribution of operation behavior categories is output; for regression tasks, the quantitative score of operation skills is output; an ensemble learning evaluation method is introduced, and the weighted sum of multiple evaluation indicators is combined as the final evaluation score , where is the index weight , and is the i-th evaluation indicator; the SHAP value analysis method is used to perform interpretability analysis on the model prediction results to identify key operation features affecting the prediction results; virtual operation trajectory data is generated through a generative adversarial network to expand the test data set, further verifying the generalization ability and robustness of the model. As a further embodiment of the present invention, the working process of the reinforcement learning algorithm is: 1) Constructed based on the spatio-temporal attention mechanism of Transformer For the training data sequence of low-voltage substation assessment meters Data features with a time dimension of D, calculate the attention weights through the multi-attention mechanism, and fuse a new multi-headed attention mechanism that combines adaptive weight adjustment and multi-scale feature fusion. The multi-attention mechanism formula is: In formulas (5), (6), and (7), where Q, K, and V are the query, key, and value matrices respectively, obtained by linearly transforming the input data X; h is the number of heads is the dimension of the key matrix is the output transformation matrix; this mechanism captures the long-range dependencies between data at different time scales and different spatial positions in parallel, realizing the deep fusion of features across regions and time periods; A is the global adaptive weight matrix; G is the local adjustment matrix; is the Hadamard product, realizing element-wise multiplication and weighting of matrices; is the multi-scale feature fusion operation, splicing features extracted by different convolutional kernels; is a learnable scalar, dynamically adjusting the dimensionality scaling of the key matrix; is the scientific system vector, enhancing the feature expression with the non-linear change value matrix; 2) Introduction of the adversarial learning framework The discriminator D and the feature extractor G are introduced for adversarial training, and the feature discrimination is improved through the min-max game; the adversarial loss function is: In formula (8), is the real data distribution, is the noise distribution, z is the noise vector; the discriminator D is used to distinguish the real data features from the generated features, forcing the feature extractor G to learn more discriminative representations and solving the problem of easy confusion of the meter data features in the complex electromagnetic environment; is the multi-scale discriminator, composed of m sub-discriminators; is the gradient penalty coefficient; 3) Construction of the dynamic semantic encoding system The knowledge graph is used to construct a semantic association network, mapping the data sample x to the semantic vector space to obtain the semantic encoding s(x); s(x) can be obtained by aggregating the related entity and relationship vectors; 4) Collaborative optimization of the federated learning framework Multiple training sites each have local data , and each site trains its local encoding model , and the server updates the global model parameters through the aggregation algorithm , and the formula is: Among them, n is the number of training sites, is the data volume of the i-th site, realizing the adaptability of the encoding model to different training scenarios; is the aggregated global model parameter, is the local model parameter of the i-th site; is the data scale exponential factor; is the AGC performance index of the i-th site model; is the performance weight index; is the semantic consistency score between the j-th site and the knowledge graph; Knowledge graph constraint loss, measuring the consistency between the model parameters and the prior knowledge 5) The generative coding strategy based on the variational autoencoder VAE The VAE generates virtual data samples by maximizing the evidence lower bound ELBO, and the formula is as follows: In formula (10), is the encoder, is the decoder, is the KL divergence, is the prior distribution of the latent variable z, which improves the generalization ability of the model; is the prior distribution of the latent variable z; represents the true data distribution; is the KL divergence, which measures the difference between the posterior distribution and the prior distribution; is the Jensen - Shannon divergence, which is used to measure the distance between the generated data distribution and the true data distribution; is the hyperparameter that balances each term, where controls the matching degree between the posterior and the prior, regulates the similarity between the generated data and the true data, and optimizes the model performance by dynamically adjusting; 6) Implementation of the spatio - temporal Bayesian network module Introduce the meta - learning mechanism, and quickly adjust the network structure and the prior distribution of parameters according to different training tasks and fault simulation scenarios S ; Use the causal discovery algorithm to mine the causal relationship between data, and construct the causal graph G=(V,E); Introduce the quantum superposition state into Bayesian inference to accelerate the calculation of the multi - node probability distribution ; 7) Implementation of the dynamic channel switching module Policy learning based on the deep Q - network DQN: Define the state space S, the action space A and the reward function R(s,a), and learn the optimal channel switching policy through the deep Q - network. The Q - network minimizes the loss function: In formula (11), is the target Q - value, is the discount factor, are the main network parameters, are the target network parameters, is the dynamic discount factor, is the entropy regularization coefficient, Entropy is the entropy of the action distribution; Construct the digital twin body, generate virtual training samples to feed back the reinforcement learning model; Use spectrum sensing to monitor the channel state C(t), and realize dynamic spectrum sharing in combination with blockchain; 8) Implementation of the multi - node interaction module Deploy edge computing units at the meter nodes for data And make an interaction decision d; Use holographic projection to display the node interaction relationship R = GNN(H, A) calculated by the graph neural network; Adjust the training difficulty according to the trainee feedback information f through the sentiment calculation model; Design a multi-agent interaction mechanism to simulate strategic games by solving the Nash equilibrium of the game.

[0010] Positive and beneficial effects A low-voltage substation area assessment meter training device and a fault simulation control method, including a meter simulation module, a data generation module, a fault simulation module, a training management module, and an intelligent interaction module. The training management module includes a project design unit, an operation guidance unit, a data collection unit, an evaluation feedback unit, and an intelligent optimization unit. The meter simulation module can accurately reproduce various meter functions; the data generation module constructs diverse test scenarios based on real data characteristics; the fault simulation module integrates software and hardware combined fault simulation technology and can reproduce various faults such as open circuits and communication anomalies; the training management module realizes the full-process control of training projects and the skill assessment of trainees through the cooperation of each unit; the intelligent interaction module realizes data interaction and instruction transmission between various modules inside the device. This device and method can, through the intelligent fault simulation algorithm module, achieve accurate identification and simulation of faults based on deep learning, and use an evaluation model combined with the analytic hierarchy process-cognition-driven improved Bayesian network to scientifically score the trainees' operations and analyze their skill shortboards. At the same time, dynamically optimize the training content according to reinforcement learning, which can effectively improve the trainees' operation proficiency, fault troubleshooting ability, and data analysis level for the low-voltage substation area assessment meter, significantly enhance the pertinence and effectiveness of training, and provide an efficient and accurate training solution for the cultivation of electric power skilled talents. Brief description of the drawings

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to these drawings, where: Figure 1 It is a schematic diagram of a low-voltage substation area assessment meter training device of the present invention; Figure 2 It is a flowchart of the training management module of a low-voltage substation area assessment meter training device of the present invention; Figure 3 It is a flowchart of the data collection unit of a low-voltage substation area assessment meter training device of the present invention; Figure 4 It is a flowchart of the convolutional neural network module of a low-voltage substation area assessment meter training device of the present invention; Figure 5This is the flowchart of the fault simulation control method for a low-voltage substation assessment meter training device of the present invention; Figure 6 This is the overall structure diagram of the low-voltage substation assessment meter training device of the present invention; Wherein: main body frame - 1, support feet - 2, universal wheels - 3, lighting module - 4, ventilation and heat dissipation structure - 5, meter simulation module - 6, intelligent interaction module - 7. Specific embodiments

[0012] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0013] As Figures 1 - 6 shown, a low-voltage substation assessment meter training device includes a device body: The device body is a rectangular main body frame made of high-strength aluminum alloy. The tensile strength of the aluminum alloy material is ≥200 MPa, and the yield strength is ≥150 MPa. Adjustable support feet and universal wheels are provided at the bottom of the main body frame; the height adjustment range of the adjustable support feet is 0 - 50 mm, and the load-bearing capacity of the universal wheels is ≥100 kg; a lighting module is provided at the top of the main body frame, and the brightness of the lighting module is ≥500 lumens, and the irradiation angle can be adjusted by ±45°; A ventilation and heat dissipation structure is provided on the front of the main body frame, which uses a silent fan with a diameter of 120 mm, the air volume is ≥50 CFM, the noise is ≤35 dB, and a dust-proof net is provided at the air inlet of the heat dissipation structure, and the filtration efficiency is ≥90%; The meter simulation module and the intelligent interaction module are provided on the multi-layer detachable mounting plates on the front of the main body frame; the mounting plates are made of 6061-T6 aluminum alloy with a thickness of 3 mm, and the detachable structure supports rapid replacement within 5 seconds; A data generation module, a fault simulation module, and a training management module are provided inside the main body frame: The meter simulation module is compatible with assessment meters of single-phase watt-hour meters, three-phase watt-hour meters, and intelligent water meters, and the simulation measurement accuracy reaches 0.5S level, the data acquisition frequency is ≥100 Hz, and it supports the communication protocol transmission functions of RS485 and LoRa; The intelligent interaction module is equipped with a 10.1-inch high-definition touch screen with a resolution of 1920×1200, displays training data and operation results through a graphical interface, supports virtual simulation operation with a delay of ≤200 ms, remote collaboration video transmission with a frame rate of ≥25 fps, realizes data interaction rate between internal modules of the device of ≥100 Mbps, and instruction transmission delay of ≤100 ms; The data generation module generates electrical parameter data of voltage, current, and power for testing based on the actual load characteristics of the power distribution area, supports ≥10 customized test scenarios, and the scenario switching time ≤5 seconds; The fault simulation module can set fault types such as open circuit, short circuit, wiring error, data anomaly, and leakage current. The fault setting response time ≤1 second, which is used to train the students' fault troubleshooting ability, and the fault reproduction accuracy ≥95%; The training management module is responsible for the creation, assignment, and progress tracking of training projects, including students' operation records, performance evaluation, and skill assessment; the training management module includes a project design unit, an operation guidance unit, a data collection unit, an evaluation feedback unit, and an intelligent optimization unit; the project design unit constructs a standardized training project template using a knowledge graph based on the typical power distribution area topology and fault case library, and the generation time of a single template ≤10 minutes; the operation guidance unit provides a standardized operation process and safety specification guidance, and the guidance response delay ≤500ms; the data collection unit models the students' operation trajectory data collected by the data collection unit through deep learning algorithms, the trajectory sampling frequency ≥10Hz, and the modeling accuracy ≥90%; the evaluation feedback unit uses the fuzzy comprehensive evaluation method to score the students' operations in multiple dimensions, combines the improved Bayesian network driven by cognition to analyze the skill shortboards, and generates a personalized improvement plan, and the plan generation time ≤3 minutes; the intelligent optimization unit dynamically adjusts the training difficulty and fault types based on the reinforcement learning algorithm, the difficulty adjustment step size ≥5 levels, realizes the adaptive optimization of the training project, and the optimization period ≤10 minutes; the output end of the project design unit is connected to the input end of the operation guidance unit, the output end of the data collection unit is connected to the input end of the evaluation feedback unit, the project design unit is connected to the intelligent optimization unit; the output end of the evaluation feedback unit is connected to the input end of the operation guidance unit.

[0014] The meter simulation module is connected to the training management module; the data generation module is connected to the training management module; the fault simulation module is connected to the training management module; the intelligent interaction module is connected to the training management module.

[0015] Furthermore, the knowledge graph includes a knowledge extraction unit, a knowledge fusion unit, and a knowledge reasoning unit. Each unit collaborates through data flow and algorithms to construct an intelligent knowledge system in the field of power training; The knowledge extraction unit uses the BERT-Transformer pre-trained model combined with the CRF sequence labeling algorithm to extract entities of transformers, lines, and meters from typical substation topology structure documents and fault case reports with an accuracy of ≥92%. At the same time, it identifies 12 types of relationships including "connected to" and "fault caused", and can process up to 200KB of text per second to quickly form initial knowledge triples. For professional terms in the power field, more than 2,000 professional dictionaries are built in to ensure the accuracy of entity extraction. The knowledge fusion unit uses the entity alignment algorithm based on cosine similarity and the relationship matching strategy based on rules. By setting a similarity threshold of 0.85, it eliminates the ambiguity of multiple expressions of the same entity and solves the knowledge conflict problem between different data sources at the same time. It adopts block indexing to improve the fusion efficiency of millions of knowledge triples by 40% and constructs a unified knowledge graph ontology architecture containing 10 major categories and more than 500 attributes, supporting OWL semantic description and SPARQL query. The knowledge reasoning unit adopts a hybrid reasoning mode combining the rule engine Drools and the graph neural network GNN. Through more than 500 self-defined rules in the power field, it conducts rapid logical deduction. At the same time, it uses the GNN model to perform in-depth learning on equipment operation parameters and more than 5,000 historical fault data to mine potential knowledge relationships with an accuracy of ≥88%. The inferred potential fault hazards can generate a visual analysis report within 30 seconds, providing extended knowledge support for the construction of training project templates, including 15 types of fault scenarios and more than 300 knowledge nodes, effectively improving the richness and practicality of training content. The output end of the knowledge extraction unit is connected to the input end of the knowledge fusion unit, and the output end of the knowledge fusion unit is connected to the input end of the knowledge reasoning unit.

[0016] In a specific embodiment, the knowledge graph collaborates through three units: knowledge extraction, fusion, and reasoning, to construct an intelligent knowledge system in the field of power training. The knowledge extraction unit extracts entities and relationships from substation topology documents and fault cases based on the BERT-Transformer pre-trained model and the CRF sequence labeling algorithm; the knowledge fusion unit uses the cosine similarity algorithm and rule matching strategy to eliminate data ambiguity and resolve conflicts; the knowledge reasoning unit adopts a mode that combines the Drools rule engine and the GNN graph neural network to mine potential knowledge relationships and provide support for training projects. Each unit is connected in series by a data flow to form a complete knowledge processing link. During implementation, the knowledge extraction unit first uses the BERT-Transformer model with more than 2,000 built-in professional dictionaries to perform semantic understanding on the documents, and combines the CRF algorithm to extract equipment entities of transformers and 12 types of relationships with an accuracy rate of more than 92%, processes 200KB of text per second, and generates initial knowledge triples. Subsequently, the knowledge fusion unit receives the triple data, performs entity alignment through the cosine similarity algorithm with a threshold of 0.85, and combines rule matching to eliminate the ambiguity of different expressions of the same entity and resolve conflicts between data sources; adopts a block indexing technology to improve the fusion efficiency of millions of knowledge triples by 40% and constructs a unified ontology architecture containing 10 major categories and more than 500 attributes. Finally, based on this architecture, the knowledge reasoning unit uses more than 500 custom power rules and the GNN model to analyze equipment parameters and more than 5,000 historical fault data, mines potential relationships with an accuracy rate of more than 88%, and generates a visual report of potential faults within 30 seconds, providing knowledge support for training projects. This knowledge graph technology significantly improves the efficiency and quality of knowledge processing in the field of power training. In the knowledge extraction link, high accuracy and processing speed ensure the efficient acquisition of original knowledge; the knowledge fusion link effectively eliminates ambiguity and conflicts, constructs a knowledge ontology with clear structure and unified standards, which is convenient for knowledge storage and query; the knowledge reasoning link quickly mines potential relationships and generates visual reports, providing rich scenarios and knowledge nodes for training projects, making the training content closer to reality and effectively improving the professionalism and practicality of training. Overall, it realizes the full-process optimization from knowledge acquisition, integration to application, and promotes the development of power training towards intelligence and high efficiency.

[0017] Further, the data acquisition unit uses an improved deep learning algorithm model to model the operation trajectory of the trainees, where the improved deep learning algorithm model includes a spatio-temporal sequence analysis module, a three-dimensional space feature analysis module, a multi-data fusion module, a convolutional neural network module, and a data coding output module. Each module works together through a data flow and a feature transformation matrix to achieve high-precision modeling of the operation trajectory of the trainees; The spatio-temporal sequence analysis module adopts an architecture that combines a bidirectional gated recurrent unit and an attention mechanism to model the time series data of the trainee's operations. The spatio-temporal sequence analysis module sets 10 parallel time windows with a time span from 50 milliseconds to 5 seconds, and dynamically assigns the importance of different time scales through a time attention weight matrix with a dimension of 10×128. A 30% weight increase is given to the temporal features in the 10 seconds before a fault operation. For continuous operations, the dynamic time warping algorithm is used to calculate the operation similarity, and the matching threshold is set to 0.85 to ensure robust recognition of similar operation patterns. The three-dimensional spatial feature analysis module is based on the PointNet++ architecture and processes three-dimensional spatial coordinate data with a sampling frequency of 100Hz and a position accuracy of ±2mm during the operation. The three-dimensional spatial feature analysis module introduces prior knowledge constraints of power equipment to construct a spatial relationship graph containing 50 types of power equipment, and extracts the topological features between equipment through a graph convolutional network. For switchgear operations, a local feature descriptor based on the cylindrical coordinate system is specially designed, and the operation trajectory is decomposed into three-dimensional motion components with a radial accuracy of ±1mm, an angular accuracy of ±0.5°, and an axial accuracy of ±1mm. The multi-data fusion module adopts a fusion strategy that combines tensor decomposition and Kalman filtering to process multi-source heterogeneous data from a touch screen with a resolution of 1920×1080, an action capture camera with a frame rate of 60fps, and a force feedback device with a sampling rate of 200Hz. The multi-data fusion module designs a modal reliability evaluation mechanism to dynamically adjust the fusion weights by calculating the confidence scores of modal data in the range of 0 to 1. When the confidence of a certain modality is lower than 0.6, the data completion algorithm is automatically started to perform interpolation processing using historical data and equipment status knowledge. The convolutional neural network module adopts a multi-scale dilated convolution architecture, including three parallel convolutional branches with dilation rates set to 2, 4, and 8 respectively. The effective receptive field can cover 5 to 40 time steps. For the key feature points in the operation trajectory, the convolutional neural network module designs a feature enhancement unit to strengthen the important feature channels and suppress irrelevant information through the channel attention mechanism. When the convolutional neural network module runs on the NVIDIA Jetson AGX Xavier hardware platform, the inference latency is controlled within 15 milliseconds to meet the real-time requirement. The processed multi-dimensional features are encoded by the data encoding and output module into a fixed-length operation vector of dimension 256, and a self-supervised learning framework is adopted to learn the semantic representation of the operations. The data encoding and output module designs a coding quality evaluation mechanism based on contrastive learning. By calculating the cosine similarity of positive and negative sample pairs, it ensures that the distance between similar operations in the coding space does not exceed 0.3, and the distance between dissimilar operations is not less than 0.7. The finally output operation vector supports downstream tasks, achieving performance indicators with an operation classification accuracy of over 95%, a fault early warning accuracy of over 85% 30 seconds in advance, and a skill evaluation F1-score of over 0.9. The output end of the spatio-temporal sequence analysis module is connected to the input end of the multi-data fusion module; the output end of the three-dimensional space feature analysis module is connected to the input end of the multi-data fusion module; the output end of the multi-data fusion module is connected to the input end of the convolutional neural network module; the output end of the convolutional neural network module is connected to the input end of the data encoding and output module.

[0018] In a specific embodiment, the improved deep learning algorithm model realizes high-precision modeling of the operation trajectory of trainees through the collaboration of five modules: spatio-temporal sequence analysis, three-dimensional spatial feature analysis, multi-data fusion, convolutional neural network, and data coding output. The spatio-temporal sequence analysis module combines a bidirectional gated recurrent unit and an attention mechanism to capture time features; the three-dimensional spatial feature analysis module processes spatial data based on the PointNet++ architecture; the multi-data fusion module fuses multi-source heterogeneous data; the convolutional neural network module extracts key features; the data coding output module completes feature coding and semantic learning. Each module is connected in series through a data flow and a feature transformation matrix to build a complete processing link from data acquisition to feature output. During implementation, the spatio-temporal sequence analysis module first collects operation time series data in 10 parallel time windows (50 milliseconds - 5 seconds), assigns a 30% weight increase to the 10 seconds before a fault operation through a 10×128-dimensional time attention weight matrix, and uses the dynamic time warping algorithm (matching threshold 0.85) to identify similar operations. The three-dimensional spatial feature analysis module obtains three-dimensional coordinate data with an accuracy of ±2 mm at a sampling frequency of 100 Hz, combines it with a spatial relationship diagram of 50 types of power equipment and a graph convolutional network, and decomposes the operation trajectory of the switchgear into three-dimensional components based on the cylindrical coordinate system. The multi-data fusion module receives heterogeneous data from touch screens, cameras, and force feedback devices, fuses them through tensor decomposition and Kalman filtering, and dynamically adjusts the weights according to a 0-1 confidence score. When the confidence is below 0.6, data completion is initiated. The convolutional neural network module uses three-branch dilated convolutions (dilation rates 2, 4, 8) to extract features, enhances key information through a channel attention mechanism, and realizes inference within 15 milliseconds on the NVIDIA Jetson AGX Xavier platform. Finally, the data coding output module encodes the features into 256-dimensional vectors, and ensures that the distance between similar operations is ≤0.3 and the distance between dissimilar operations is ≥0.7 through contrastive learning. This model significantly improves the modeling accuracy and application value of the operation trajectory of trainees. The spatio-temporal and three-dimensional spatial analysis modules accurately capture the time and spatial features of operations, and the multi-data fusion module effectively processes heterogeneous data to ensure the integrity and reliability of the input data. The convolutional neural network quickly extracts key features, and the data coding output module realizes highly discriminative semantic coding, ultimately achieving a performance with an operation classification accuracy of over 95%, a fault early warning accuracy of over 85% 30 seconds in advance, and an F1-score of the skill assessment reaching above 0.9, providing accurate data support for trainee operation assessment, fault prediction, and skill improvement, and promoting the intelligent development of power training. Further, the convolutional neural network module includes a data preprocessing module, a spatio-temporal feature extraction module, an attention mechanism module, a model training module, and a prediction and evaluation module. Each module realizes end-to-end intelligent analysis of power data through a standardized data flow interface and a feature enhancement algorithm chain; The data preprocessing module designs a multi-modal data normalization pipeline. For image data with a resolution of no less than 1024×1024, zero-mean normalization is adopted, setting the data mean to 0 and the standard deviation to 1, and combining the CLAHE (Contrast Limited Adaptive Histogram Equalization) technology to expand the pixel dynamic range to the interval of 0 to 255; for time-series data with a sampling frequency of 50Hz and above, normalization is performed through a sliding window with a length of 100 points and a step size of 10 points, screening and retaining the valid values within the 95th percentile of the data to effectively suppress abnormal pulse interference; for audio data with a sampling rate of 44.1kHz, it is converted into an 80-dimensional Mel spectrogram using a Mel filter bank, and a random gain jitter of -6dB to +6dB is added to enhance the robustness of the model to different audio signals; the data preprocessing module supports parallel processing of 32-channel data, and the single-batch processing delay is controlled within 50ms; The spatio-temporal feature extraction module constructs a four-branch dilated convolutional network architecture, with the dilation rates of the four branches set to 1, 3, 5, and 7 respectively, corresponding to covering the receptive field range from 1×1 to 15×15 pixels; each branch contains a triple structure: first is a double residual block, designed with a bottleneck structure, a channel compression ratio of 4:1, and embedded with cross-layer connection and weight normalization technologies to improve the stability of the training process while reducing the computational complexity; second is a multi-scale pooling layer, which parallelly performs three pooling operations of 1×1, 3×3, and 5×5, and dynamically generates weights through a gating mechanism to achieve the adaptive fusion extraction of multi-granularity features; finally is a feature calibration layer, which based on the learned affine transformation parameters γ and β, uniformly adjusts the brightness and contrast of the cross-branch features to ensure the consistency of the feature space; when this module runs on an NVIDIA A100 GPU, the floating-point operation volume reaches 12.8G, and the feature extraction frame rate can be as high as 200FPS; The attention mechanism module designs a three-dimensional attention fusion framework, covering attention mechanisms in three dimensions: spatial, channel, and temporal: Spatial attention: Generates a spatial attention map through 7×7 convolutional operations, precisely focusing on the key areas of power equipment, such as transformer windings and circuit breaker contacts, and uses a soft mask mechanism to suppress background noise interference, with the generated attention map having the same resolution as the input feature map; Channel attention: Builds the dependency relationship between channels by means of global average pooling operation and combining two fully-connected networks, and assigns a weight gain of 1.2 times and above to the channels reflecting electrical parameters, such as the channels related to voltage and current; Temporal attention: For long-sequence data with a time span of 30 minutes and above, a Transformer encoder structure is introduced, and an 8-head multi-head self-attention mechanism is used to capture the dependencies across time points, supporting the processing of a temporal window of up to 5000 time steps at most; The overall parameter ratio of this attention module does not exceed 15%, but can increase the feature discriminability by more than 20%; The model training module adopts a mixed-precision training strategy, dynamically switches between FP16 and FP32 data precisions based on the PyTorch framework, and reduces the video memory occupancy by 40% without loss of precision. During the training process, the AdamW optimizer and the SGD optimizer are used in combination. The initial learning rate of AdamW is set to 1e-4, the weight decay value is 0.01, the SGD momentum is set to 0.9, and the learning rate is adjusted with a cosine annealing learning rate decay strategy at a period of 200 epochs to improve the stability of model convergence. The data augmentation strategy is customized according to the data type: for image data, random rotation, Gaussian blur, and CutOut augmentation methods are used; for time-series data, random time offset, amplitude scaling, and injection of noise with a signal-to-noise ratio not lower than 20 dB are performed; for audio data, time stretching, pitch shifting, and adding background noise from the substation environment noise library are implemented. This module supports distributed training with up to 8 cards synchronized, and the single-card training efficiency can be increased to more than 75%. The prediction and evaluation module constructs a multi-task output head, supporting three core task types: classification, regression, and anomaly detection. Classification task: The cross-entropy loss function is used, and the number of neurons in the output layer is set according to the number of fault types, not less than 20 classes, and the classification accuracy on the standard test set reaches 98% or more. Regression task: The smooth L1 loss function is used to output the state parameters of power equipment, such as temperature and load rate, and the prediction error is controlled within ±3% of the true value. Anomaly detection: A boundary model is constructed based on One-Class SVM, and the degree of anomaly in the feature space is measured by the Mahalanobis distance. The F1-score of anomaly detection is not lower than 0.95. In addition, the evaluation index real-time visualization module supports generating more than 10 types of charts such as confusion matrices, ROC curves, and error distribution histograms, and the result update delay is controlled within 1 second. The output end of the data preprocessing module is connected to the input end of the spatio-temporal feature extraction module; the output end of the spatio-temporal feature extraction module is connected to the attention mechanism module; the output end of the attention mechanism module is connected to the input end of the model training module; the output end of the model training module is connected to the input end of the prediction and evaluation block. In a specific embodiment, the convolutional neural network module constructs an end-to-end intelligent power data analysis system through the standardized connection of five major modules: data preprocessing, spatio-temporal feature extraction, attention mechanism, model training, and prediction and evaluation. The data preprocessing module unifies the multi-modal data format, the spatio-temporal feature extraction module mines the data space and time features, the attention mechanism module focuses on key information, the model training module optimizes the network parameters, and the prediction and evaluation module realizes the task output. Each module relies on the data flow interface and the feature enhancement algorithm to cooperate to achieve the efficient analysis and accurate prediction of power data. During implementation, the data preprocessing module processes images with a resolution of 1024×1024 or above through zero-mean normalization and CLAHE technology; for time series data of 50Hz and above, it is normalized through a 100-point sliding window; the 44.1kHz audio is converted into an 80-dimensional Mel spectrogram, and random gain is added, supporting 32-channel parallel processing, with the single-batch delay controlled within 50ms. The spatio-temporal feature extraction module adopts four-branch dilated convolution (dilation rates 1, 3, 5, 7), and each branch contains a double residual block, a multi-scale pooling layer, and a feature calibration layer, achieving 12.8G floating-point operations and a frame rate of 200FPS on an NVIDIA A100 GPU. The attention mechanism module generates a spatial attention map through 7×7 convolution, constructs channel dependencies by combining global average pooling and two fully connected layers, and processes long sequences using a Transformer encoder, increasing the feature discriminability by more than 20%. The model training module dynamically switches between FP16 and FP32 precisions based on PyTorch, combines the AdamW and SGD optimizers, and adopts a cosine annealing strategy, supporting 8-card distributed training. The prediction and evaluation module constructs a multi-task output head, with a classification accuracy of over 98%, a regression error controlled within ±3%, an F1-score for anomaly detection not lower than 0.95, and the evaluation chart updated within 1 second. This convolutional neural network module realizes the efficient and intelligent analysis of power data. Data preprocessing ensures the quality of multi-modal data, spatio-temporal feature extraction and the attention mechanism accurately capture key features, model training optimizes network performance, and prediction and evaluation provide accurate results. Finally, high-performance indicators are achieved in classification, regression, and anomaly detection tasks, and visual evaluation results are quickly output, effectively improving the efficiency and accuracy of power data processing, providing reliable technical support for power equipment status monitoring and fault diagnosis.

[0019] Further, the fuzzy comprehensive evaluation method is designed based on a modular architecture, including an index system construction module, a weight calculation module, a fuzzy relationship generation module, and a comprehensive evaluation operation module. Each module works together to achieve the quantitative evaluation of the trainees' practical performance; The above-mentioned index system construction module is responsible for building a scientific evaluation index system to measure the trainees' practical training performance from four core dimensions; for the dimension of operation standardization, by verifying 8 sub-indicators of the integrity of operation steps and compliance with safety regulations, the operation compliance of trainees is evaluated; for the dimension of fault diagnosis accuracy, based on 6 sub-indicators of the correct rate of fault type identification and the deviation of fault point location, the diagnosis accuracy is determined; for the dimension of fault handling efficiency, based on 4 time parameters of fault response time and repair time, the processing speed is quantified; for the dimension of data interpretation ability, through 5 sub-indicators of the depth of data anomaly analysis and the accuracy of parameter correlation reasoning, the analysis ability of trainees for power data is evaluated; finally, a multi-level evaluation index system containing 23 sub-indicators is constructed to fully cover the practical training assessment requirements; The above-mentioned weight calculation module uses the Analytic Hierarchy Process (AHP) to determine the weights of each index; a group consisting of 5 power field experts and 3 education evaluation experts is formed, and a judgment matrix is constructed according to the 1-9 scale method to score the relative importance of each index through pairwise comparison; by calculating the maximum eigenvalue of the matrix and the corresponding eigenvector, the initial weight vector is obtained, and a consistency test is carried out to ensure the logical rationality of the judgment matrix; if the consistency ratio CR > 0.1, the scoring is adjusted again; when CR ≤ 0.1, the final weight vector W = [w1, w2, w3, w4] is determined, where w1, w2, w3, w4 correspond to the weights of operation standardization, fault diagnosis accuracy, fault handling efficiency, and data interpretation ability respectively, to realize differential weighting for different assessment dimensions; The above-mentioned fuzzy relation generation module divides the trainees' practical training performance into five evaluation levels: excellent, good, medium, pass, and fail; the evaluation group consists of practical training tutors and technical backbones, and scores are given according to the records of trainees' operation processes and the materials of fault handling results; for each evaluation index, the proportion of the number of trainees whose performance belongs to each level is counted to form membership degree vectors R1, R2, R3, R4; for example, for the operation standardization index, if 20% of the trainees are rated as excellent, the membership degree of the excellent level in R1 is 0.2; the membership degree vectors of the four indexes are arranged row by row to construct a 4×5-dimensional fuzzy relation matrix R to visually present the correlation degree between trainees' performance and each evaluation level; The above-mentioned comprehensive evaluation operation module calculates the comprehensive evaluation vector B through matrix multiplication operation B = W×R. This vector contains the comprehensive membership degrees of trainees' performance belonging to the five evaluation levels; the vector B is normalized to make the sum of all elements equal to 1 to enhance the comparability of the results; according to the principle of maximum membership degree, the level corresponding to the element with the largest value in B is selected as the comprehensive evaluation level of the trainees; at the same time, corresponding score intervals are assigned to excellent (90 - 100 points), good (80 - 89 points), medium (70 - 79 points), pass (60 - 69 points), and fail (0 - 59 points), and the weighted comprehensive score is calculated in combination with the membership degree to realize the accurate quantitative evaluation of trainees' practical training performance and provide data support for skill improvement; The output end of the index system construction module is connected to the input end of the weight calculation module; the output end of the weight calculation module is connected to the input end of the fuzzy relation generation module; the output end of the fuzzy relation generation module is connected to the input end of the comprehensive evaluation operation module. In a specific embodiment, the fuzzy comprehensive evaluation method is based on a modular architecture. Through the collaborative work of four major modules, namely index system construction, weight calculation, fuzzy relation generation, and comprehensive evaluation operation, the fuzzy concept of the trainee's training performance is transformed into a quantitative evaluation result. The index system construction module establishes evaluation criteria from multiple dimensions. The weight calculation module determines the importance of each index. The fuzzy relation generation module quantifies the association between the trainee's performance and the levels of evaluation. The comprehensive evaluation operation module integrates the outputs of the previous three to achieve accurate scoring. Each module forms a scientific trainee training evaluation system with data transmission as the link. During implementation, the index system construction module first selects 8 sub-indicators, 6 sub-indicators, 4 sub-indicators, and 5 sub-indicators from the four dimensions of operation standardization, fault diagnosis accuracy, fault handling efficiency, and data interpretation ability respectively, and constructs a multi-level system containing 23 sub-indicators. The weight calculation module consists of 5 experts in the power field and 3 education evaluation experts. According to the 1-9 scale method, a judgment matrix is constructed. After eigenvalue calculation and consistency test (CR ≤ 0.1), the weight vectors W of the four dimensions of operation standardization and fault diagnosis accuracy are determined. In the fuzzy relation generation module, the evaluation team divides the performance into five levels of evaluation according to the trainee's training records, counts the proportion of the number of trainees in different levels under each index, and constructs a 4×5-dimensional fuzzy relation matrix R. Finally, the comprehensive evaluation operation module obtains the comprehensive evaluation vector B through matrix multiplication B = W × R, determines the level of evaluation according to the principle of maximum membership degree after normalization, and calculates the weighted comprehensive score in combination with the score interval of each level. This fuzzy comprehensive evaluation method realizes the accurate quantitative evaluation of the trainee's training performance. The constructed multi-level index system comprehensively covers the key links of training assessment, and the scientific weight determination method ensures the rationality of evaluation in each dimension. Through the fuzzy relation matrix and comprehensive operation, the subjective evaluation is transformed into objective data, making the evaluation result more reliable. Finally, it realizes the accurate scoring of the trainee's operation standardization and fault diagnosis ability, provides data basis for the trainee's skill improvement, and also provides a direction for teaching optimization, effectively improving the quality of power training teaching and the professionalism of evaluation. Furthermore, the improved Bayesian network is designed with a modular architecture, including a network structure optimization module, a parameter learning algorithm module, an inference mechanism enhancement module, a cognitive interpretation module, and a domain adaptation module. Each module collaborates with each other to achieve accurate power fault diagnosis and training evaluation; The network structure optimization module is responsible for constructing a hierarchical causal network architecture, which is divided into a device layer, a parameter layer, and a phenomenon layer. The device layer integrates 20 core power equipment nodes of transformers and meters. The parameter layer aggregates 15 real-time monitoring parameter nodes of voltage, current, and frequency. The phenomenon layer integrates 8 fault characterization nodes of alarm information and abnormal sounds. By analyzing more than 1,000 groups of expert-annotated data and more than 3,000 historical fault cases, the dependence relationship between nodes at each layer is determined, and an accurate causal link is constructed. For the dynamic evolution characteristics of power faults, a dynamic Bayesian network architecture with a 15-minute time window is introduced. Each time slice is set with 30 nodes, which can completely capture the temporal characteristics of the fault from occurrence to development, and realize the dynamic modeling of the fault evolution process. The parameter learning algorithm module adopts a hybrid parameter estimation strategy to improve the model accuracy. For common fault nodes with rich data samples, such as voltage over-limit nodes, the maximum likelihood estimation method is used to calculate the conditional probability. For rare fault event nodes, such as equipment insulation aging, Bayesian estimation is adopted and the Jeffreys prior distribution is configured. During the learning process, an adaptive learning rate mechanism is introduced, and the weights of historical data decay according to an exponential law, with a half-life of 30 days. The weight of newly collected data is 20% higher than that of historical data, ensuring that the model can quickly adapt to new fault modes and improve the learning ability for complex fault scenarios. The inference mechanism enhancement module improves the inference performance through evidence weighted fusion and reverse diagnosis. An evidence weighting mechanism is designed to assign credibility weights to evidence from different sources. Among them, the weight of infrared thermal imaging evidence is set to 0.85 because it can intuitively reflect equipment temperature anomalies. The weight of text alarm information is set to 0.7 considering possible false alarms. The Dempster-Shafer theory is used to handle evidence conflicts. When multi-source evidence conflicts, effective fusion is carried out through the evidence synthesis rule. The reverse diagnosis inference mode is supported, which can start from the fault phenomenon of tripping in the phenomenon layer and reverse-derive the posterior probability of faults at each node in the device layer. In complex fault scenarios, the accuracy of fault source location can reach more than 92%, significantly improving the fault diagnosis efficiency. The cognitive explanation module has a knowledge base containing 120 typical power fault modes, and each fault mode is associated with standard operation steps, safety specification requirements, and repair plan documents. Based on Bayesian inference, the C4.5 algorithm is used to generate a decision tree explanation path, converting the probability inference result into an intuitive rule expression such as "If the voltage continues to rise abnormally and the current exceeds the rated value, then there may be a line overload fault", making the diagnosis result conform to the reading and understanding habits of power industry technicians, and enhancing the interpretability and guidance of the diagnosis result. The field adaptation module constructs a conditional probability table of a typical structure with a radial or looped network based on the topological relationship of the equipment in the substation area; fully considers the electrical distance and influence coefficient between equipment, for example, the probability of a neighboring equipment failure affecting other equipment is 30% higher than that of a non-neighboring equipment; in view of the uncertainty of the power system operation, the probability interval is used to represent the possibility of a failure, such as expressing the failure probability of a certain equipment as an interval range from 0.65 to 0.73. Compared with the precise point estimation, it is more in line with the actual operation scenario of the power system and improves the applicability of the model in the low-voltage substation area training. The output end of the network structure optimization module is connected to the input end of the parameter learning algorithm module; the output end of the parameter learning module is connected to the input end of the inference mechanism enhancement module; the output end of the inference mechanism enhancement module is connected to the input end of the cognitive interpretation module; the input end of the cognitive interpretation module is connected to the input end of the field adaptation module.

[0020] In a specific embodiment, the improved Bayesian network adopts a modular architecture. Through the collaboration of five major modules, namely network structure optimization, parameter learning algorithm, inference mechanism enhancement, cognitive interpretation, and domain adaptation, the uncertainty problems in power fault diagnosis and training evaluation are transformed into probability reasoning and knowledge application. The network structure optimization module constructs a hierarchical causal network. The parameter learning algorithm module optimizes the model parameters. The inference mechanism enhancement module improves the evidence processing and diagnosis capabilities. The cognitive interpretation module enhances the readability of the results. The domain adaptation module improves the model fitting degree. Each module is connected in series by a data flow to achieve accurate power fault diagnosis and efficient training evaluation. During implementation, the network structure optimization module first divides the equipment layer, parameter layer, and phenomenon layer, integrates 20 types of equipment node, 15 monitoring parameter nodes, and 8 types of fault characterization nodes, determines the node dependency relationship based on more than 1000 groups of expert data and more than 3000 cases, and introduces a 15-minute dynamic architecture and a 30-node time slice to realize dynamic fault modeling. The parameter learning algorithm module uses maximum likelihood estimation for common fault nodes, Bayesian estimation combined with Jeffreys prior distribution for rare fault nodes, and exponentially decays the weight of historical data with a half-life of 30 days, and increases the weight of new data by 20%. The inference mechanism enhancement module weights the evidence of infrared thermal imaging (weight 0.85) and text information (weight 0.7), uses Dempster-Shafer theory to fuse conflicting evidence, and makes the accuracy of fault source location exceed 92% through reverse diagnostic reasoning. The cognitive interpretation module relies on a knowledge base of 120 fault modes and uses the C4.5 algorithm to transform probability reasoning into intuitive rules. The domain adaptation module constructs a conditional probability table based on the substation topology, increases the probability of adjacent equipment failure impact by 30%, and represents the fault possibility with a probability interval. This improved Bayesian network significantly improves the efficiency of power fault diagnosis and training evaluation. The hierarchical dynamic network and precise parameter learning ensure the accurate capture of fault characteristics by the model. Evidence weighted fusion and reverse diagnosis make complex fault location more efficient. The cognitive interpretation function enhances the understandability of the diagnosis results. Domain adaptation optimization improves the ability of the model to fit the actual scenario. Finally, the accuracy of fault diagnosis is increased by 18%, the time-consuming for complex fault location is shortened to within 2 minutes, the proportion of diagnostic interpretation reports meeting industry specifications reaches more than 95%, effectively supporting the training and evaluation of trainees' fault troubleshooting skills, and providing a reliable technical guarantee for power training teaching.

[0021] Further, the improved reinforcement learning model includes a spatio-temporal feature fusion module, a data encoding module, a spatio-temporal Bayesian network module, a dynamic channel switching module, and a multi-node interaction module; The spatiotemporal feature fusion module innovatively introduces a Transformer-based spatiotemporal attention mechanism to break the traditional 3D-CNN's reliance on local features. When processing the low-voltage area assessment meter training data, the multi-head attention mechanism is used to capture the long-range dependencies between data of different time scales and different spatial positions in parallel, thus achieving deep fusion of features across regions and time periods. At the same time, an adversarial learning framework is embedded to train a discriminator to distinguish between real data features and generated features, forcing the feature extractor to learn more discriminative representations, effectively solving the problem of easy confusion of meter data features in complex electromagnetic environments. The data encoding module pioneered a dynamic semantic encoding system that integrates knowledge graphs and federated learning. For low-voltage area meter data, the knowledge graph is used to construct a semantic association network between meter parameters, fault types, and operating environments, giving data encoding deep semantic information. Combined with the federated learning framework, while protecting the data privacy of each training site, the encoding model is collaboratively optimized so that the encoding parameters can adapt to the data distribution differences in different training scenarios. In addition, a generative encoding strategy based on variational autoencoder VAE is designed to generate virtual data samples during the training process, expand the diversity of the training set, and improve the generalization ability of the model. The spatiotemporal Bayesian network module breaks through the limitations of traditional static modeling and constructs a dynamic evolutionary Bayesian network architecture. It introduces a meta-learning mechanism to quickly adjust the network structure and parameter prior distribution under different training tasks and fault simulation scenarios to achieve rapid adaptation of the model. It combines the causal discovery algorithm to automatically mine the causal relationship between low-voltage meter data, rather than relying solely on statistical correlation, so that the network can more accurately predict the fault propagation path. At the same time, it integrates quantum computing principles into the Bayesian reasoning process, and uses quantum superposition states to accelerate the parallel calculation of multi-node probability distribution, greatly improving the reasoning efficiency in complex scenarios. The dynamic channel switching module proposes a dual-drive strategy based on deep reinforcement learning and digital twins. On the one hand, the optimal channel switching strategy in the low-voltage area communication environment is learned through the deep Q network DQN combined with the experience replay mechanism. On the other hand, a virtual area communication digital twin is constructed to preview the effects of different channel switching actions in the twin environment, and generate virtual training samples to feed back the reinforcement learning model. In addition, a spectrum sensing and blockchain fusion solution is designed, using spectrum sensing to monitor the channel status in real time, and using blockchain technology to build a trusted channel usage record book to achieve dynamic spectrum sharing among multiple training devices, effectively improving the utilization rate of limited frequency band resources. The multi-node interaction module builds a collaborative system based on edge intelligence and holographic interaction; lightweight edge computing units are deployed at each meter node to achieve local preprocessing of data and interaction decision-making, reducing data transmission latency; holographic projection technology is adopted to present the node interaction relationships obtained by graph neural network calculations in the form of visual 3D dynamic maps, assisting trainees to intuitively understand the complex coupling relationships between the nodes in the substation area; at the same time, an emotion computing model is introduced to dynamically adjust the node interaction difficulty and fault simulation strategy according to the feedback information of the trainees' expressions and voices during the training operation process, realizing a personalized training experience; in addition, a multi-agent interaction mechanism based on game theory is designed to simulate the strategic games of different stakeholders in substation area management, enhancing the realistic mapping and challenging nature of the training content; The output end of the spatio-temporal feature fusion module is connected to the input end of the data encoding module; the output end of the data encoding module is connected to the input end of the spatio-temporal Bayesian network module; the output end of the spatio-temporal Bayesian network module is connected to the input end of the dynamic channel switching module; the output end of the dynamic channel switching module is connected to the input end of the multi-node interaction module. In a specific embodiment, this improved reinforcement learning model focuses on the training scenario of low-voltage substation assessment meters, integrating multi-field cutting-edge technologies to achieve intelligent training. The spatio-temporal feature fusion module takes the Transformer spatio-temporal attention mechanism and the adversarial learning framework as the core, breaks through the limitations of traditional convolutional networks, and strengthens the extraction and differentiation of meter data features in complex electromagnetic environments; the data encoding module integrates knowledge graphs, federated learning, and variational autoencoders to build a dynamic semantic encoding system, endowing data with deep semantics and ensuring data privacy and model generalization ability; the spatio-temporal Bayesian network module introduces meta-learning, causal discovery algorithms, and quantum computing to achieve network dynamic evolution and efficient causal reasoning; the dynamic channel switching module combines deep reinforcement learning, digital twins, spectrum sensing, and blockchain to optimize channel resource management; the multi-node interaction module is based on edge intelligence, holographic projection, emotion computing, and game theory to create an immersive and personalized training interaction system. Each module operates in coordination to form a complete intelligent training solution. First, the spatio-temporal feature fusion module uses the Transformer-based spatio-temporal attention mechanism to capture the long-range dependence relationship of data, enhances the feature discrimination degree through adversarial learning, and outputs fused features; then, the data encoding module uses the knowledge graph to construct a semantic network, optimizes the encoding model in combination with federated learning, and generates virtual samples through VAE to complete data semantic encoding; subsequently, the spatio-temporal Bayesian network module adjusts the network structure with the help of meta-learning, mines the causal relationship of data through causal discovery, and uses quantum computing to accelerate reasoning, and outputs prediction results; then, the dynamic channel switching module learns the optimal switching strategy based on the deep Q network and digital twins, and realizes dynamic spectrum sharing in combination with spectrum sensing and blockchain; finally, the multi-node interaction module deploys an edge computing unit at the meter node, visualizes the interaction relationship through holographic projection, adjusts the training difficulty according to emotion computing, and simulates multi-agent interaction using game theory to complete the entire training process. This model significantly improves the intelligent level and training effect of low-voltage substation assessment meter training. In terms of feature processing, it solves the problem of data feature confusion in complex environments; during the encoding process, it takes into account semantic information and data privacy; in the reasoning link, it greatly improves the accuracy and efficiency of fault prediction; channel management effectively improves the utilization rate of frequency band resources; at the interaction level, it realizes a personalized and immersive training experience. Overall, the model provides a more realistic and efficient training environment for trainees, enhances the fit between training content and real scenarios, and is of great significance for improving trainees' skill levels and substation management capabilities.

[0022] Furthermore: S1. Training scenario construction Based on the standardized training project template constructed by the knowledge graph, with the help of the 15 types of fault scenarios and more than 300 knowledge nodes output by the knowledge reasoning unit, and combined with the 8 typical substation area topologies of radial and loop network types in the topology structure library of the project design unit, the training scenario parameters are automatically generated; the system can dynamically configure 5 mainstream meter models of single-phase and three-phase watt-hour meters, and set the initial operating state of 15 monitoring parameters covering voltage and current according to the historical operation data containing more than 3,000 cases; among them, the voltage range is 0 - 400V, and the accuracy reaches ±0.5%FS; the current range is 0 - 100A, and the accuracy is ±0.2%FS; the data generation module loads and generates specific training tasks containing equipment topology relationships, operating parameters and potential fault hazards at a scenario switching speed of no more than 5 seconds, and at the same time, through the 10.1-inch high-definition touch screen with a resolution of 1920×1280 equipped in the intelligent interaction module, the detailed task information is displayed; S2. Fault injection After receiving the training task instruction, the fault simulation module can quickly respond within 1 second and inject faults through two mechanisms of hardware control and signal interference; for hardware faults such as open circuit and short circuit, the solid-state relay is used to control the on-off of the circuit; for signal faults such as data anomaly and leakage current, the interference waveform with an accuracy of ±2% is output by the analog signal generator; the system supports setting 8 types of fault types including wiring error and equipment insulation aging, and the fault recurrence accuracy is not less than 95%; while injecting faults, the intelligent interaction module shows the fault phenomena such as sudden change of voltage curve and flashing of alarm indicator light with a visualization animation with a frame rate of not less than 25fps, and then combined with the abnormal sound of the equipment generated by the environmental audio encoder, even in an environment with a signal-to-noise ratio of -5dB, the recognition rate of abnormal equipment sound can reach more than 90%, so as to create a highly realistic fault environment; S3. Trainee operation and data collection When the trainee is operating, the data acquisition unit captures three-dimensional space coordinates in real time at a sampling frequency of 100 Hz, with a position accuracy of ±2 mm, and combines a bidirectional gated recurrent unit and an attention mechanism to model the operation time series data; multi-source data is collected through a touch screen with a resolution of 1920×1080, an action capture camera with a frame rate of 60 fps, and a force feedback device with a sampling rate of 200 Hz; after these data are fused by tensor decomposition and Kalman filtering, the confidence of each modality within the range of 0 to 1 is calculated; when the confidence of a certain modality is lower than 0.6, the data completion algorithm is automatically started, and interpolation processing is performed using historical data and device status knowledge; the convolutional neural network module processes the data with an inference delay of 15 ms, extracts the operation trajectory features, and finally generates a multi-modal training dataset containing 256-dimensional operation vectors, completely recording the trainee's operation steps, changes in device parameters, and the whole process of fault handling; S4. Evaluation and Feedback The evaluation and feedback unit uses the fuzzy comprehensive evaluation method to score the trainee from four dimensions of operation standardization and fault diagnosis accuracy; the operation standardization is evaluated by checking 8 sub-indicators, and the fault diagnosis accuracy is determined based on 6 sub-indicators; through the analytic hierarchy process, the expert group constructs a judgment matrix, calculates the maximum eigenvalue and the corresponding eigenvector of the matrix, obtains the initial weight vector, and conducts a consistency test to ensure that the consistency ratio CR does not exceed 0.1, so as to determine the final weight vector W and divide the trainee's performance into five evaluation levels; at the same time, the improved Bayesian network reversely deduces the skill shortboards in the trainee's operation based on more than 1000 groups of expert-annotated data and historical fault cases, and generates a diagnostic report with a fault location accuracy of 92%; finally, the system generates an evaluation report containing quantitative scores, fault analysis, and improvement suggestions within 3 minutes, and feeds it back to the trainee through the intelligent interaction module, and at the same time synchronously pushes it to the teacher-end management system; S5. Training Optimization The intelligent optimization unit is based on the reinforcement learning algorithm, and dynamically adjusts the difficulty of the training according to the evaluation results of the trainees; the difficulty adjustment step is not less than 5 levels, and is achieved by increasing the complexity of the fault type, upgrading from a single short-circuit fault to a compound fault, and shortening the fault response time requirement; combined with the conditional probability table constructed by the domain adaptation module, the distribution of fault types in subsequent training projects is adjusted to increase the probability of adjacent equipment related faults by 30%; the entire optimization cycle is controlled within 10 minutes to ensure that the training content is dynamically matched with the trainees' skill level, and through continuous iterative training, the trainees' fault diagnosis and processing capabilities in the process of low-voltage area meter operation and maintenance are continuously improved. Furthermore, the steps of the convolutional neural network modeling are: 1) Data preprocessing and format conversion: The collected trainee operation trajectory data, including operation timestamps, operation equipment numbers, operation instruction codes, and equipment status parameter vectors, are formed into a multivariate sequence set; for missing values, an interpolation method based on Gaussian process regression is adopted, and the covariance matrix K is constructed using the kernel function, and the linear equation group is solved. Get the interpolation coefficients , to achieve high-precision filling; the outlier detection introduces the local outlier factor algorithm, which identifies and corrects abnormal data points by calculating the ratio of the local reachable density of sample x to the average local reachable density of neighborhood samples; when normalizing the operation timestamp, Min-Max normalization combined with time series smoothing technology is used; the generator G and the discriminator D are introduced to form an adversarial network, and the generator G generates a normalized time series The discriminator D determines the true normalization result The difference with the pseudo result is as follows: In formula (1), To adjust the parameters, used to control the range of the normalized timestamp, the non-interval sequence is converted into a time interval sequence through cubic spline interpolation; To balance the hyperparameters of adversarial loss; through this adversarial learning method, the normalization process not only meets the basic numerical transformation requirements, but also learns a more discriminative and robust normalization mode in the adversarial game, effectively dealing with noise and outlier interference in the data; Construct a dynamic sliding event window, the window size is determined by the adaptive parameter Control, convert the operation trajectory sequence into a three-dimensional tensor format , where N is the number of samples, T is the time step, M is the operation feature dimension, and D is the number of data channels; mathematical expression: , in the formula, is the feature conversion function for the Dth data channel, Denote the operation data of the nth sample at time step t and feature dimension m; 2) Spatiotemporal feature extraction: Input the processed three-dimensional tensor X into the spatiotemporal feature extraction module, adopting a multi-scale parallel convolution architecture; For each time step, use convolution kernels of different sizes , where l represents the convolution layer number, k is the size of the convolution kernel in the time dimension, D is the number of input channels, is the number of output channels, and perform convolution operations; Based on the meta-learning framework, enable the parameters of the convolution layer to be quickly and adaptively adjusted according to different task scenarios; During the training process, optimize the parameters of the meta-network through the meta-learning algorithm, so that the generated convolution layer parameters are more adaptable to different types of training data. During the training process, optimize the parameters of the meta-network through the meta-learning algorithm, so that the generated convolution layer parameters are more adaptable to different types of training data. The improved calculation process is as follows: Introduce a gating mechanism, and the output feature map of the lth convolution layer is calculated as: In formula (2), is the Sigmoid activation function, is the gating signal tensor represents element-wise multiplication, is the convolution kernel weight, is the bias term, is the pooling weight; Expand the receptive field through the dilated convolution layer, and combine the residual connection mechanism to solve the problem of gradient disappearance; are the parameters of the meta-network. In this way, the network can quickly adjust the convolution calculation method under different training scenarios of low-voltage substation meter reading, and improve the generalization ability of the model; In the pooling stage, adopt the adaptive weighted pooling method to dynamically adjust the pooling weight according to the local method of the feature map; 3) Application of the attention mechanism: Input the extracted spatiotemporal feature map Y into the attention mechanism module, adopting the multi-head self-attention mechanism; Introduce the spatiotemporal dynamic weight matrix , which is composed of the temporal dependence coefficient in the time dimension and the feature correlation coefficient in the space dimension. By capturing the continuity of the trainee's operation trajectory in the time series and the relevance of the operation device spatial layout, dynamically adjust the calculation of the attention weight. For each head h, calculate the attention weight using the formula: In formula (3), are the query vector and the key vector respectively, obtained through linear transformation, is the key vector dimension; can be obtained by calculating the reciprocal of the operation time interval and the autocorrelation of the time series, Determined based on the physical connection relationship and functional relevance among operating devices; this approach enables more precise focus on key operation links when processing training data of low-voltage substation assessment meters. For example, in the fault handling process, higher weights are assigned to operation steps with sequential dependencies. Concatenate the multi-head attention outputs and perform a linear transformation to obtain a feature vector with attention; 4) Model training and parameter optimization: Input the feature vector Z with attention into the fully connected layer, adopt a hybrid loss function L, combine the cross-entropy loss , mean squared error loss and regularization term . Combine the attention mechanism, introduce dynamic attention weights for each loss sub-term, and construct an attention weight matrix, . Dynamically generate weights by calculating the importance of different tasks in the current training stage; the formula is: In formula (4), is the weight parameter, which can be calculated through the self-attention mechanism based on the change rate of task loss and the uncertainty factor of prediction results, realizing collaborative optimization among multiple tasks and improving the comprehensive performance of the model in the training task of low-voltage substation assessment meters; 5) Prediction and evaluation: Use the trained model to predict new trainee operation trajectory data; for classification tasks, output the probability distribution of operation behavior categories; for regression tasks, output the quantitative score of operation skills; introduce an ensemble learning evaluation method, and combine the weighted sum of multiple evaluation indicators as the final evaluation score , where is the index weight, is the i-th evaluation index; the SHAP value analysis method is used to perform interpretability analysis on the model prediction results to identify the key operation features that affect the prediction results; virtual operation trajectory data is generated through a generative adversarial network to expand the test data set, and further verify the generalization ability and robustness of the model. In a specific embodiment, the technology constructs a high-precision and interpretable training data analysis model for the operation trajectory data of trainees through data preprocessing, spatio-temporal feature extraction, application of attention mechanism, model training optimization and prediction evaluation. Its core lies in integrating the technologies of generative adversarial network, meta-learning, and multi-head self-attention to dynamically optimize data processing and model parameters, solve the problems of data noise, complex features, and model generalization, and achieve accurate classification of trainee operation behaviors and quantitative evaluation of skills. Data preprocessing and format conversion: Integrate the operation trajectory data to form a multi-variable sequence, use Gaussian process regression to fill in missing values, and use local outlier factor to detect outliers; improve Min-Max normalization through a generative adversarial network, and combine the interval sequence generated by cubic spline interpolation; construct a dynamic sliding window to convert the data into a three-dimensional tensor. Spatio-temporal feature extraction: Input the three-dimensional tensor into a multi-scale parallel convolution architecture, dynamically adjust the convolution parameters based on meta-learning, and combine the gated mechanism, dilated convolution, and residual connection to extract features; use adaptive weighted pooling to optimize the pooling process. Application of attention mechanism: Through the multi-head self-attention mechanism, fuse the spatio-temporal dynamic weight matrix, capture the operation continuity and device correlation, dynamically adjust the attention weight, and focus on the key operation links. Model training and parameter optimization: Use the feature vector with attention as the input, adopt a mixed loss function with dynamic weights, calculate the task weights through the self-attention mechanism, and achieve multi-task collaborative optimization. Prediction and evaluation: Use the trained model to complete classification or regression tasks; combine ensemble learning and SHAP value analysis to achieve comprehensive evaluation and result interpretation; verify the model generalization through a generative adversarial network to expand the data set. In the data preprocessing stage, Gaussian process regression realizes interpolation based on kernel function to measure data similarity, and a generative adversarial network optimizes the normalization strategy through game learning; spatio-temporal feature extraction uses meta-learning to quickly adapt to different data distributions, and the gated mechanism and residual connection enhance the network stability; the multi-head self-attention mechanism strengthens feature correlation through the spatio-temporal weight matrix; the mixed loss function balances the training objectives according to the task dynamic weights; the SHAP value quantifies the feature contributions from the perspective of game theory to ensure the interpretability of the model.

[0023] By comparing the traditional image recognition method and the recognition method based on the convolutional neural network model of deep learning, 100 power equipment fault images are classified and recognized, and compared with the actual fault types to determine the recognition effect of the convolutional neural network model on the power equipment fault images. Table 1 shows the result comparison between the traditional recognition method and the convolutional neural network model recognition method: As can be seen from the data in Table 1, the error rate of traditional image recognition methods in the recognition of 100 power equipment fault images is as high as 32 - 70%, while the recognition method based on the convolutional neural network model of deep learning achieves 100% correct recognition with an error rate of 0. It can be seen that the convolutional neural network model has significantly better recognition effect on power equipment fault images than traditional image recognition methods, and can more accurately identify the types of power equipment faults, providing reliable technical support for the operation and maintenance of power equipment.

[0024] Furthermore, the working process of the reinforcement learning algorithm is as follows: 1) Construct For the training data sequence of low-voltage substation assessment meters Data features with a time dimension of D, calculate the attention weights through a multi-attention mechanism, and fuse a new type of multi-head attention mechanism that combines adaptive weight adjustment and multi-scale feature fusion. The formula for the multi-attention mechanism is: In formulas (5), (6), and (7), where Q, K, and V are the query, key, and value matrices respectively, which are obtained by linear transformation of the input data X; h is the number of heads; is the dimension of the key matrix; is the linear transformation parameter matrix for each head, is the output transformation matrix; this mechanism captures the long-range dependence relationships between data at different time scales and different spatial positions in parallel, realizing deep feature fusion across regions and time periods; A is the global adaptive weight matrix; G is the local adjustment matrix; is the Hadamard product, which realizes element-wise multiplication and weighting of matrices; is the multi-scale feature fusion operation, which concatenates the features extracted by different convolutional kernels; is a learnable scalar that dynamically adjusts the dimension scaling of the key matrix; is the scientific system vector, and the non-linear change value matrix enhances the feature expression; 2) Introduction of the adversarial learning framework Introduce a discriminator D and a feature extractor G for adversarial training, and improve the feature discrimination degree through a min-max game; the adversarial loss function is: In formula (8), is the true data distribution, is the noise distribution, z is the noise vector; the discriminator D is used to distinguish the true data features from the generated features, forcing the feature extractor G to learn more discriminative representations, and solving the problem that the meter data features are easily confused in a complex electromagnetic environment; is the multi-scale discriminator, which consists of m sub-discriminators; is the gradient penalty coefficient; 3) Construction of the dynamic semantic encoding system Construct a semantic association network using a knowledge graph, map the data sample x to the semantic vector space to obtain the semantic encoding s(x); s(x) can be obtained by aggregating the relevant entity and relationship vectors; 4) Collaborative optimization of the federated learning framework Multiple training sites each have local data , and each site trains a local encoding model , and the server updates the global model parameters through an aggregation algorithm , and the formula is: where n is the number of training sites, is the data volume of the i-th site, enabling the encoding model to adapt to different training scenarios; is the aggregated global model parameter, is the local model parameter of the i-th site; is the data scale exponential factor; is the AGC performance index of the i-th site model; is the performance weight index; is the semantic consistency score between the j-th site and the knowledge graph; Knowledge graph constraint loss, measuring the consistency between model parameters and prior knowledge 5) Generative encoding strategy based on the variational autoencoder VAE VAE generates virtual data samples by maximizing the evidence lower bound ELBO, and the formula is: In formula (10), is the encoder, is the decoder, is the KL divergence, is the prior distribution of the latent variable z, improving the generalization ability of the model; is the prior distribution of the latent variable z; represents the true data distribution; is the KL divergence, measuring the difference between the posterior distribution and the prior distribution; is the Jensen - Shannon divergence, used to measure the distance between the generated data distribution and the true data distribution; is a hyperparameter that balances each item, where controls the matching degree between the posterior and the prior, adjusts the similarity between the generated data and the true data, and dynamically adjusts to optimize the model performance; 6) Implementation of the spatio-temporal Bayesian network module Introduce a meta-learning mechanism, and quickly adjust the network structure and prior distribution of parameters according to different training tasks and fault simulation scenarios S ; Mining the causal relationships between data using causal discovery algorithms to construct a causal graph G=(V,E); introducing quantum superposition states into Bayesian inference for accelerating the calculation of multi-node probability distributions for acceleration calculation; 7) Implementation of the dynamic channel switching module Policy learning based on the deep Q-network DQN: Define the state space S, action space A, and reward function R(s,a). Learn the optimal channel switching policy through the deep Q-network. The Q-network minimizes the loss function: In formula (11), is the target Q value, is the discount factor, are the main network parameters, are the target network parameters, is the dynamic discount factor, is the entropy regularization coefficient, Entropy is the action distribution entropy; Construct a digital twin to generate virtual training samples to feed back to the reinforcement learning model; Use spectrum sensing to monitor the channel state C(t), and combine with blockchain to achieve dynamic spectrum sharing; 8) Implementation of the multi-node interaction module Deploy edge computing units at the meter nodes to process the data and make an interaction decision d; Use holographic projection to display the node interaction relationship R = GNN(H,A) calculated by the graph neural network; Adjust the training difficulty according to the trainee feedback information f through the emotion calculation model; Design a multi-agent interaction mechanism to simulate strategic games by solving the Nash equilibrium of the game. In a specific embodiment, this reinforcement learning algorithm constructs an intelligent training interaction system, performs one-hot encoding on the meter operation parameters, fault types, and trainee historical operation sequences to construct a multi-dimensional state space S, and defines a discrete action space A including operations such as line inspection and component replacement; Design a reward function, comprehensively considering the effectiveness of fault handling and operation standardization, giving a positive reward +r for successfully eliminating faults, giving a negative penalty -r for violating operations that trigger safety risks, and providing a small maintenance reward r for regular operations; Update the policy network based on the Q-learning algorithm, and introduce an adversarial training mechanism composed of a generator and a discriminator to optimize the Q-value iteration formula and improve the robustness of the model. Table 2 data shows that the reinforcement learning algorithm significantly improves the training effect compared with traditional training methods: the average fault handling duration is shortened by 42.4%, the operation standard compliance rate is increased by 31.3%, the trainee knowledge mastery efficiency is increased by 33.0%, the strategy generalization success rate and the correct rate of dealing with abnormal situations are increased by 42.7% and 50.4% respectively, effectively realizing the dynamic balance between training difficulty and trainee ability, and improving the training efficiency and accuracy. It can be seen from the data in Table 2 that the reinforcement learning algorithm is significantly superior to the traditional training method: the average fault handling time is shortened by 42.4%, the compliance rate of operation specifications is increased by 31.3%, the knowledge mastery efficiency of trainees is improved by 33.0%, and the success rate of policy generalization and the correct rate of coping with abnormal situations are increased by 42.7% and 50.4% respectively. The Q-network optimized by adversarial training effectively enhances the policy robustness, realizes the dynamic balance between the training difficulty and the trainees' abilities, provides more efficient and accurate training guidance for trainees, and greatly improves the training effect and learning efficiency.

[0025] Although the specific implementation manners of the present invention have been described above, those skilled in the art should understand that these specific implementation manners are only illustrative. Without departing from the principle and essence of the present invention, those skilled in the art can make various omissions, substitutions and changes to the details of the above methods and systems. For example, combining the above method steps, so as to perform substantially the same function according to a substantially same method to achieve substantially the same result, then it belongs to the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. A training device for low-voltage substation metering instruments, comprising a device body, characterized in that: The device body is a rectangular main frame (1) made of aluminum alloy. Adjustable support feet (2) and universal wheels (3) are arranged at the bottom of the main frame (1); a ventilation and heat dissipation structure (5) is arranged on the front of the main frame (1); a meter simulation module (6) and an intelligent interaction module (7) are arranged on the multi-layer detachable mounting plate on the front of the main frame (1); a data generation module, a fault simulation module and a training management module are arranged inside the main frame (1): The meter simulation module (6) is compatible with the metering instruments for single-phase watt-hour meters, three-phase watt-hour meters and intelligent water meters; The intelligent interaction module (7) is equipped with a 10.1-inch high-definition touch screen. The data generation module generates electrical parameter data of voltage, current and power for testing based on the load characteristics of the actual substation area, supports ≥10 customized test scenarios, and the scenario switching time ≤5 seconds; The fault simulation module sets fault types such as open circuit, short circuit, wiring error, data anomaly and leakage current. The fault setting response time ≤1 second, which is used to train the students' fault troubleshooting ability, and the fault recurrence accuracy rate ≥95%; The training management module is responsible for the creation, distribution and progress tracking of training projects. The training management module includes students' operation records, performance evaluation and skill assessment; the training management module includes a project design unit, an operation guidance unit, a data collection unit, an evaluation feedback unit and an intelligent optimization unit; the project design unit constructs a standardized training project template based on the typical substation area topology and fault case library by using a knowledge graph, and the generation time of a single template ≤10 minutes; the operation guidance unit provides a standardized operation process and safety specification guidance, and the guidance response delay ≤500ms; the data collection unit models the operation trajectory data of students collected by the data collection unit through a deep learning algorithm, the trajectory sampling frequency ≥10Hz, and the modeling accuracy rate ≥90%; the evaluation feedback unit uses the fuzzy comprehensive evaluation method to score the students' operations in multiple dimensions, combines the improved Bayesian network driven by cognition to analyze the skill short board, and generates a personalized improvement plan, and the plan generation time ≤3 minutes; The intelligent optimization unit dynamically adjusts the training difficulty and fault types based on the reinforcement learning algorithm, and the difficulty adjustment step size ≥5 levels, realizing the adaptive optimization of training projects, and the optimization period ≤10 minutes; the output end of the project design unit is connected to the input end of the operation guidance unit, the output end of the data collection unit is connected to the input end of the evaluation feedback unit, and the project design unit is connected to the intelligent optimization unit; the output end of the evaluation feedback unit is connected to the input end of the operation guidance unit; the meter simulation module is connected to the training management module; the data generation module is connected to the training management module; the fault simulation module is connected to the training management module; the intelligent interaction module is connected to the training management module.

2. The training device for low-voltage substation metering instruments according to claim 1, characterized in that: The meter simulation module (6) has a simulated measurement accuracy of 0.5S level, a data acquisition frequency of ≥100Hz, and supports the communication protocol transmission functions of RS485 and LoRa; The intelligent interaction module (7) has a resolution of 1920×1200, displays training data and operation results through a graphical interface, supports a virtual simulation operation delay of ≤200ms, a remote collaboration video transmission frame rate of ≥25fps, and realizes a data interaction rate of ≥100Mbps between the internal modules of the device, and an instruction transmission delay of ≤100ms; The aluminum alloy material of the rectangular main frame (1) has a tensile strength of ≥200MPa and a yield strength of ≥150MPa; the height adjustment range of the adjustable support feet (2) is 0 - 50mm, and the universal wheels (3) can carry a weight of ≥100kg; a lighting module (4) is provided at the top of the main frame (1), the brightness of the lighting module (4) is ≥500 lumens, and the irradiation angle can be adjusted by ±45°; the ventilation and heat dissipation structure (5) uses a silent fan with a diameter of 120mm, the air volume is ≥50CFM, the noise is ≤35dB, and a dust-proof net is provided at the air inlet of the heat dissipation structure, and the filtration efficiency is ≥90%; The mounting plate of the main frame (1) is made of 6061-T6 aluminum alloy with a thickness of 3mm, and the detachable structure supports quick replacement within 5 seconds; The knowledge graph includes a knowledge extraction unit, a knowledge fusion unit, and a knowledge reasoning unit. Each unit collaborates through data streams and algorithms to construct an intelligent knowledge system in the field of power training; among them: The knowledge extraction unit uses the BERT-Transformer pre-training model combined with the CRF sequence labeling algorithm to extract entities such as transformers, lines, and meter devices from typical substation topology structure documents and fault case reports with an accuracy of ≥92%. At the same time, it identifies 12 types of relationships including "connected to" and "fault caused", and can process up to 200KB of text per second to quickly form initial knowledge triples; for professional terms in the power field, more than 2000 professional dictionaries are built in to ensure the accuracy of entity extraction; The knowledge fusion unit uses an entity alignment algorithm based on cosine similarity and a relationship matching strategy based on rules to eliminate the ambiguity of multiple expressions of the same entity by setting a similarity threshold of 0.85, and at the same time solve the knowledge conflict problem between different data sources; it uses block indexing to improve the fusion efficiency of millions of knowledge triples by 40%, and constructs a unified knowledge graph ontology architecture containing 10 major categories and more than 500 attributes, supporting OWL semantic description and SPARQL query; The knowledge inference unit adopts a hybrid inference mode that combines the rule engine Drools and the graph neural network GNN. It performs rapid logical deduction through more than 500 custom rules in the power field. At the same time, it uses the GNN model to perform deep learning on device operation parameters and more than 5,000 historical fault data to mine potential knowledge relationships with an accuracy rate of ≥88%. The inferred potential faults can generate a visual analysis report within 30 seconds, providing extended knowledge support for the construction of training project templates, including 15 types of fault scenarios and more than 300 knowledge nodes, effectively improving the richness and practicality of training content. The output end of the knowledge extraction unit is connected to the input end of the knowledge fusion unit, and the output end of the knowledge fusion unit is connected to the input end of the knowledge inference unit.

3. A low-voltage substation assessment meter training device according to claim 1, characterized in that: The meter simulation module (6) simulates a measurement accuracy of 0.5S level, the data acquisition frequency is ≥100Hz, and it supports the communication protocol transmission functions of RS485 and LoRa. The intelligent interaction module (7) has a resolution of 1920×1200, displays training data and operation results through a graphical interface, supports a virtual simulation operation delay of ≤200ms, a remote collaboration video transmission frame rate of ≥25fps, and realizes a data interaction rate of ≥100Mbps between the internal modules of the device and an instruction transmission delay of ≤100ms. The aluminum alloy material of the rectangular main frame (1) has a tensile strength of ≥200MPa and a yield strength of ≥150MPa; the height adjustment range of the adjustable support feet (2) is 0-50mm, and the universal wheels (3) can bear a weight of ≥100kg; a lighting module (4) is provided at the top of the main frame (1), the brightness of the lighting module (4) is ≥500 lumens, and the irradiation angle can be adjusted by ±45°; the ventilation and heat dissipation structure (5) uses a silent fan with a diameter of 120mm, the air volume is ≥50CFM, the noise is ≤35dB, and a dust-proof net is provided at the air inlet of the heat dissipation structure, and the filtration efficiency is ≥90%. The mounting plate of the main frame (1) is made of 6061-T6 aluminum alloy with a thickness of 3mm, and the detachable structure supports rapid replacement within 5 seconds. The knowledge graph includes a knowledge extraction unit, a knowledge fusion unit, and a knowledge inference unit. Each unit collaborates through data flow and algorithms to build an intelligent knowledge system in the field of power training; among them: The knowledge extraction unit uses the BERT-Transformer pre-training model combined with the CRF sequence labeling algorithm to extract entities of transformers, lines, and meter devices from typical substation topology structure documents and fault case reports with an accuracy rate of ≥92%. At the same time, it identifies 12 types of relationships including "connected to" and "fault caused", and can process a text volume of up to 200KB per second to quickly form initial knowledge triples; for professional terms in the power field, more than 2,000 professional dictionaries are built in to ensure the accuracy of entity extraction. The knowledge fusion unit uses an entity alignment algorithm based on cosine similarity and a rule-based relationship matching strategy. By setting a similarity threshold of 0.85, it eliminates the ambiguity of multiple representations of the same entity and resolves knowledge conflicts between different data sources. It adopts block indexing to improve the fusion efficiency of millions of knowledge triples by 40% and constructs a unified knowledge graph ontology architecture containing 10 major categories and more than 500 attributes, supporting OWL semantic description and SPARQL query. The knowledge reasoning unit adopts a hybrid reasoning mode that combines the rule engine Drools and the graph neural network GNN. It performs rapid logical deduction through more than 500 custom rules in the power domain, and at the same time uses the GNN model to perform deep learning on device operation parameters and more than 5000 historical fault data to mine potential knowledge relationships with an accuracy of ≥88%. The inferred fault hazards can generate a visual analysis report within 30 seconds, providing extended knowledge support for the construction of training project templates, including 15 types of fault scenarios and more than 300 knowledge nodes, effectively improving the richness and practicality of training content. The output end of the knowledge extraction unit is connected to the input end of the knowledge fusion unit, and the output end of the knowledge fusion unit is connected to the input end of the knowledge reasoning unit.

4. The low-voltage substation assessment meter training device according to claim 3, wherein: The convolutional neural network module includes a data preprocessing module, a spatio-temporal feature extraction module, an attention mechanism module, a model training module, and a prediction evaluation module. Each module realizes end-to-end intelligent analysis of power data through a standardized data flow interface and a feature enhancement algorithm chain. The data preprocessing module designs a multi-modal data normalization pipeline. For image data with a resolution of not less than 1024×1024, zero-mean normalization is adopted, setting the data mean to 0 and the standard deviation to 1, and combining the CLAHE contrast-limited adaptive histogram equalization technology to expand the pixel dynamic range to the interval of 0 to 255. For time-series data with a sampling frequency of 50Hz or above, normalization is performed through a sliding window with a length of 100 points and a step size of 10 points, screening and retaining the valid values within the 95th percentile of the data to effectively suppress abnormal pulse interference. For audio data with a sampling rate of 44.1kHz, it is converted into an 80-dimensional Mel spectrogram using a Mel filter bank, and a random gain jitter of -6dB to +6dB is added to enhance the robustness of the model to different audio signals. The data preprocessing module supports parallel processing of 32 channels of data, and the single-batch processing delay is controlled within 50ms. The spatio-temporal feature extraction module constructs a four-branch dilated convolutional network architecture, and the dilation rates of the four branches are set to 1, 3, 5, and 7 respectively, corresponding to realizing the receptive field coverage in the pixel range from 1×1 to 15×15. Each branch contains a triple structure: first is a double residual block, designed with a bottleneck structure, a channel compression ratio of 4:1, and embedded cross-layer connection and weight normalization technology to improve the stability of the training process while reducing the computational complexity. Followed by the multi-scale pooling layer, which performs three pooling operations of 1×1, 3×3, and 5×5 in parallel, dynamically generates weights through a gating mechanism, and realizes the adaptive fusion extraction of multi-granularity features; finally, there is the feature calibration layer, which based on the learned affine transformation parameters γ and β, uniformly adjusts the brightness and contrast of cross-branch features to ensure the consistency of the feature space; when this module runs on the NVIDIA A100 GPU, the floating-point operation volume reaches 12.8G, and the feature extraction frame rate can be as high as 200FPS; The attention mechanism module designs a three-dimensional attention fusion framework, covering attention mechanisms in three dimensions: spatial, channel, and temporal: Among them: Spatial attention: Generates a spatial attention map through 7×7 convolution operations, precisely focuses on the key areas of power equipment, and uses a soft mask mechanism to suppress background noise interference. The generated attention map has the same resolution as the input feature map; Channel attention: Builds the dependence relationship between channels by means of global average pooling operation and combines two fully connected networks. For channels reflecting electrical parameters, a weight gain of 1.2 times or more is given; Temporal attention: For long-sequence data with a time span of 30 minutes or more, a Transformer encoder structure is introduced, and an 8-head multi-head self-attention mechanism is used to capture the dependence relationship across time points, and the maximum can support the processing of a temporal window of 5000 time steps; The overall parameter ratio of this attention module does not exceed 15%, and the feature discrimination ability is improved by more than 20%; The model training module adopts a mixed-precision training strategy, realizes the dynamic switching of FP16 and FP32 data precisions based on the PyTorch framework, and reduces the video memory occupancy by 40% without loss of precision; During the training process, the AdamW optimizer and the SGD optimizer are combined. The initial learning rate of AdamW is set to 1e-4, the weight decay value is 0.01, the SGD momentum is set to 0.9, and the learning rate is adjusted in a cycle of 200 epochs through the cosine annealing learning rate decay strategy to improve the stability of model convergence; The data augmentation strategy is customized according to the data type: for image data, random rotation, Gaussian blur, and CutOut augmentation methods are used; for temporal data, random time offset, amplitude scaling, and injection of noise with a signal-to-noise ratio not lower than 20dB are performed; for audio data, time stretching, pitch transformation, and adding background noise from the substation environment noise library are used for augmentation; This module supports distributed training with a maximum of 8 cards synchronized, and the single-card training efficiency is increased to more than 75%; The prediction and evaluation module constructs a multi-task output head, supporting three core task types: classification, regression, and anomaly detection. Classification task: The cross-entropy loss function is adopted, and the number of neurons in the output layer is set according to the number of fault types, not less than 20 types, and the classification accuracy rate on the standard test set reaches 98% or above. Regression task: The smooth L1 loss function is used to output the state parameters of power equipment, such as temperature and load rate, and the prediction error is controlled within ±3% of the true value. Anomaly detection: A boundary model is constructed based on One-Class SVM, and the degree of anomaly in the feature space is measured by the Mahalanobis distance. The F1-score of anomaly detection is not less than 0.

95. In addition, the evaluation index real-time visualization module supports generating more than 10 types of charts such as confusion matrices, ROC curves, and error distribution histograms, and the result update delay is controlled within 1 second. The output end of the data preprocessing module is connected to the input end of the spatio-temporal feature extraction module; the output end of the spatio-temporal feature extraction module is connected to the attention mechanism module; the output end of the attention mechanism module is connected to the input end of the model training module; the output end of the model training module is connected to the input end of the prediction and evaluation block.

5. The low-voltage substation assessment meter training device according to claim 1, characterized in that: The fuzzy comprehensive evaluation method is designed based on a modular architecture, including an index system construction module, a weight calculation module, a fuzzy relationship generation module, and a comprehensive evaluation operation module. Each module works together to achieve a quantitative evaluation of the trainee's training performance. The index system construction module is responsible for building a scientific evaluation index system to measure the trainee's training performance from four core dimensions. In the dimension of operation standardization, the compliance of the trainee's operation is evaluated by checking 8 sub-indices of the completeness of operation steps and the compliance with safety specifications. In the dimension of fault diagnosis accuracy, the diagnostic accuracy is determined based on 6 sub-indices of the correct rate of fault type identification and the deviation of fault point location. In the dimension of fault handling efficiency, the processing speed is quantified based on 4 time parameters of fault response time and repair time. In the dimension of data interpretation ability, the analysis ability of the trainee on power data is evaluated by 5 sub-indices of the depth of data anomaly analysis and the accuracy of parameter correlation reasoning. Finally, a multi-level evaluation index system including 23 sub-indices is constructed to fully cover the training assessment requirements. The weight calculation module uses the Analytic Hierarchy Process (AHP) to determine the weights of each indicator. A panel consisting of 5 power domain experts and 3 education evaluation experts constructs a judgment matrix according to the 1-9 scale method, and scores the relative importance of each indicator through pairwise comparison. By calculating the maximum eigenvalue and the corresponding eigenvector of the matrix, the initial weight vector is obtained, and a consistency test is carried out to ensure the logical rationality of the judgment matrix. If the consistency ratio CR > 0.1, the scores are readjusted. When CR ≤ 0.1, the final weight vector W = [w1, w2, w3, w4] is determined, where w1, w2, w3, w4 correspond to the weights of operation standardization, fault diagnosis accuracy, fault handling efficiency, and data interpretation ability respectively, realizing differential weighting for different assessment dimensions; The fuzzy relation generation module divides the training performance of the trainees into five evaluation levels: excellent, good, medium, pass, and fail; scores are given according to the records of the trainees' operation process and the data of the fault handling results. For each evaluation indicator, the proportion of the number of trainees whose performance belongs to each level is counted to form membership degree vectors R1, R2, R3, R4. The membership degree vectors of the four indicators are arranged by row to construct a 4×5 dimensional fuzzy relation matrix R, visually presenting the correlation degree between the trainees' performance and each evaluation level; The comprehensive evaluation operation module calculates the comprehensive evaluation vector B through matrix multiplication operation B = W × R. This vector contains the comprehensive membership degrees of the trainees' performance belonging to the five evaluation levels; The vector B is normalized so that the sum of its elements is 1 to enhance the comparability of the results. According to the principle of the maximum membership degree, the level corresponding to the element with the largest value in B is selected as the comprehensive evaluation level of the trainees. At the same time, corresponding score intervals are assigned to excellent (90 - 100 points), good (80 - 89 points), medium (70 - 79 points), pass (60 - 69 points), and fail (0 - 59 points). Combining the membership degrees, the weighted comprehensive score is calculated to achieve accurate quantitative evaluation of the trainees' training performance and provide data support for skill improvement; The output end of the indicator system construction module is connected to the input end of the weight calculation module; the output end of the weight calculation module is connected to the input end of the fuzzy relation generation module; the output end of the fuzzy relation generation module is connected to the input end of the comprehensive evaluation operation module.

6. The low-voltage substation area assessment meter training device according to claim 1, characterized in that: The improved Bayesian network adopts a modular architecture design, including a network structure optimization module, a parameter learning algorithm module, an inference mechanism enhancement module, a cognitive interpretation module, and a domain adaptation module. Each module cooperates with each other to achieve accurate power fault diagnosis and training evaluation; The network structure optimization module is responsible for constructing a hierarchical causal network architecture, which is divided into an equipment layer, a parameter layer, and a phenomenon layer; The equipment layer integrates 20 types of core power equipment nodes such as transformers and meters; the parameter layer gathers 15 real-time monitoring parameter nodes of voltage, current and frequency; the phenomenon layer integrates 8 types of fault characterization nodes of alarm information and abnormal sound; by analyzing more than 1,000 groups of expert annotation data and more than 3,000 historical fault cases, the dependencies between nodes at each layer are determined, and accurate causal links are constructed; in view of the dynamic evolution characteristics of power faults, a dynamic Bayesian network architecture with a time window of 15 minutes is introduced, and 30 nodes are set for each time slice, which can fully capture the temporal characteristics of faults from occurrence to development, and realize dynamic modeling of the fault evolution process; The parameter learning algorithm module adopts a hybrid parameter estimation strategy to improve the accuracy of the model; for common fault nodes with rich data samples, the maximum likelihood estimation method is used to calculate the conditional probability; for rare fault event nodes, Bayesian estimation is used and Jeffreys prior distribution is configured; in the learning process, an adaptive learning rate mechanism is introduced, and the weight of historical data decays exponentially, with a half-life of 30 days. The weight of newly collected data is 20% higher than that of historical data, ensuring that the model can quickly adapt to new fault modes and improve the learning ability of complex fault scenarios; The reasoning mechanism enhancement module improves the reasoning performance through evidence weighted fusion and reverse diagnosis; designs an evidence weighting mechanism to assign credibility weights to evidence from different sources, among which the infrared thermal imaging evidence is weighted as 0.85 because it can directly reflect the abnormal temperature of the equipment; uses the Dempster-Shafer theory to deal with evidence conflicts, and when there are contradictions in multi-source evidence, it is effectively integrated through evidence synthesis rules; supports reverse diagnosis reasoning mode, starting from the fault phenomenon of tripping at the phenomenon layer, reversely derives the fault posterior probability of each node at the equipment layer, and in complex fault scenarios, the fault source location accuracy can reach more than 92%, significantly improving the fault diagnosis efficiency; the cognitive interpretation module has a built-in knowledge base containing 120 typical power fault modes, and each fault mode is associated with standard operating procedures, safety specification requirements and repair plan documents; based on Bayesian reasoning, the C4.5 algorithm is used to generate a decision tree interpretation path; The domain adaptation module constructs a conditional probability table of typical radial and ring network structures based on the topological relationship of the substation equipment; fully considers the electrical distance and influence coefficient between the equipment, and uses probability intervals to represent the probability of faults in view of the uncertainty of power system operation. Compared with precise point estimation, it is more in line with the actual operation scenario of the power system and improves the applicability of the model in low-voltage substation training; The output end of the network structure optimization module is connected to the input end of the parameter learning algorithm module; the output end of the parameter learning module is connected to the input end of the reasoning mechanism enhancement module; the output end of the reasoning mechanism enhancement module is connected to the input end of the cognitive interpretation module; the input end of the cognitive interpretation module is connected to the input end of the domain adaptation module.

7. The low voltage area meter assessment training device according to claim 1 is characterized by: The improved reinforcement learning model includes a spatiotemporal feature fusion module, a data encoding module, a spatiotemporal Bayesian network module, a dynamic channel switching module and a multi-node interaction module; The spatiotemporal feature fusion module introduces a Transformer-based spatiotemporal attention mechanism, which captures the long-range dependencies between data of different time scales and different spatial locations in parallel through a multi-head attention mechanism, thus achieving deep fusion of features across regions and time periods. At the same time, an adversarial learning framework is embedded to train a discriminator to distinguish between real data features and generated features, forcing the feature extractor to learn more discriminative representations, effectively solving the problem of easy confusion of meter data features in complex electromagnetic environments. The data encoding module pioneered a dynamic semantic encoding system that integrates knowledge graphs and federated learning. For low-voltage area meter data, the knowledge graph is used to construct a semantic association network between meter parameters, fault types, and operating environments, giving data encoding deep semantic information. Combined with the federated learning framework, while protecting the data privacy of each training site, the encoding model is collaboratively optimized so that the encoding parameters can adapt to the data distribution differences in different training scenarios. The spatiotemporal Bayesian network module breaks through the limitations of traditional static modeling and constructs a dynamic evolutionary Bayesian network architecture. It introduces a meta-learning mechanism and combines it with a causal discovery algorithm to automatically mine the causal relationship between low-voltage meter data, rather than relying solely on statistical associations, so that the network can more accurately predict the fault propagation path. Integrate quantum computing principles into the Bayesian reasoning process, use quantum superposition to accelerate the parallel calculation of multi-node probability distribution, and greatly improve reasoning efficiency in complex scenarios; The dynamic channel switching module proposes a dual-drive strategy based on deep reinforcement learning and digital twins; through the deep Q network DQN combined with the experience replay mechanism, the optimal channel switching strategy in the low-voltage area communication environment is learned; a virtual area communication digital twin is constructed, the effects of different channel switching actions are rehearsed in the twin environment, and a virtual training sample is generated to feed back the reinforcement learning model; The multi-node interaction module creates a collaborative system based on edge intelligence and holographic interaction; a lightweight edge computing unit is deployed at each meter node to realize local preprocessing and interactive decision-making of data, and reduce data transmission delay; holographic projection technology is used to present the node interaction relationship obtained by the graph neural network calculation in a visual 3D dynamic map, to assist students in intuitively understanding the complex coupling relationship between the nodes in the substation; at the same time, an emotional computing model is introduced to dynamically adjust the node interaction difficulty and fault simulation strategy according to the feedback information of the students' expressions and voices during the practical training operation, so as to realize personalized practical training experience; in addition, a multi-agent interaction mechanism based on game theory is designed to simulate the strategic game of different stakeholders in substation management, and enhance the reality mapping and challenge of the practical training content; The output end of the spatio-temporal feature fusion module is connected to the input end of the data encoding module; the output end of the data encoding module is connected to the input end of the spatio-temporal Bayesian network module; the output end of the spatio-temporal Bayesian network module is connected to the input end of the dynamic channel switching module; the output end of the dynamic channel switching module is connected to the input end of the multi-node interaction module.

8. A fault simulation control method for a training device of low-voltage substation assessment meters, characterized in that, Controlling a low-voltage substation assessment meter training device according to any one of claims 1-7, characterized in that: It includes the following steps: S1. Training scenario construction Based on the standardized training project template constructed by the knowledge graph, with the help of the 15 types of fault scenarios and more than 300 knowledge nodes output by the knowledge reasoning unit, and combined with the 8 typical substation topologies of radial and loop network types in the topology structure library of the project design unit, the training scenario parameters are automatically generated; the system can dynamically configure 5 mainstream meter models of single-phase and three-phase watt-hour meters, and set the initial operating state of 15 monitoring parameters covering voltage and current according to the historical operation data containing more than 3,000 cases; among them, the voltage range is 0-400V, and the accuracy reaches ±0.5%FS; the current range is 0-100A, and the accuracy is ±0.2%FS; the data generation module loads and generates specific training tasks containing equipment topology relationships, operating parameters and potential fault hazards at a scene switching speed of no more than 5 seconds, and at the same time, through the 10.1-inch high-definition touch screen with a resolution of 1920×1280 equipped by the intelligent interaction module, displays the detailed task information; S2. Fault injection After receiving the training task instruction, the fault simulation module can respond quickly within 1 second and inject faults through two mechanisms of hardware control and signal interference; for hardware faults such as open circuit and short circuit, a solid-state relay is used to control the on-off of the circuit; for signal faults such as data anomaly and leakage current, an analog signal generator is used to output interference waveforms with an accuracy of ±2%; the system supports setting 8 types of fault types including wiring error and equipment insulation aging, and the fault recurrence accuracy is not less than 95%; while injecting faults, the intelligent interaction module displays the fault phenomena such as sudden change of voltage curve and flashing of alarm indicator lights with a visualization animation at a frame rate of not less than 25fps, and then cooperates with the abnormal sound of the equipment generated by the environmental audio encoder. Even in an environment with a signal-to-noise ratio of -5dB, the recognition rate of abnormal equipment sounds can reach more than 90%, so as to create a highly realistic fault environment; S3. Trainee operation and data collection When the trainee is operating, the data acquisition unit captures three-dimensional spatial coordinates in real time at a sampling frequency of 100 Hz, with a position accuracy of ±2 mm. Combining a bidirectional gated recurrent unit and an attention mechanism, it models the operation time series data; multi-source data is collected through a touch screen with a resolution of 1920×1080, an action capture camera with a frame rate of 60 fps, and a force feedback device with a sampling rate of 200 Hz; after the data is fused by tensor decomposition and Kalman filtering, the confidence of each modality within the range of 0 to 1 is calculated; when the confidence of a certain modality is lower than 0.6, the data completion algorithm is automatically started, and interpolation processing is performed using historical data and device status knowledge; the convolutional neural network module processes the data with an inference delay of 15 ms, extracts the operation trajectory features, and finally generates a multi-modal training dataset containing 256-dimensional operation vectors, completely recording the trainee's operation steps, changes in device parameters, and the entire process of fault handling; S4. Evaluation and Feedback The evaluation and feedback unit uses the fuzzy comprehensive evaluation method to score the trainee from four dimensions of operation standardization and fault diagnosis accuracy; the operation standardization is evaluated by checking 8 sub-indicators, and the fault diagnosis accuracy is determined based on 6 sub-indicators; through the analytic hierarchy process, the expert group constructs a judgment matrix, calculates the maximum eigenvalue and the corresponding eigenvector of the matrix to obtain the initial weight vector, and conducts a consistency test to ensure that the consistency ratio CR does not exceed 0.1, thereby determining the final weight vector W and dividing the trainee's performance into five evaluation levels; at the same time, the improved Bayesian network, based on more than 1000 groups of expert-annotated data and historical fault cases, reversely deduces the skill shortboards in the trainee's operation and generates a diagnostic report with a fault location accuracy rate of 92%; finally, the system generates an evaluation report containing quantitative scores, fault analysis, and improvement suggestions within 3 minutes, which is fed back to the trainee through the intelligent interaction module and simultaneously pushed to the teacher-end management system; S5. Training Optimization The intelligent optimization unit, based on the reinforcement learning algorithm, dynamically adjusts the training difficulty according to the trainee's evaluation results; the difficulty adjustment step size is not less than 5 levels, which is achieved by increasing the complexity of the fault types, upgrading from a single short-circuit fault to a composite fault, and shortening the fault response time requirement; combined with the conditional probability table constructed by the domain adaptation module, the fault type distribution in subsequent training projects is adjusted, and the probability of adjacent device associated faults is increased by 30%; the entire optimization cycle is controlled within 10 minutes to ensure that the training content and the trainee's skill level are dynamically matched, and through continuous iterative training, continuously improve the trainee's fault diagnosis and handling ability during the operation and maintenance of low-voltage substation metering devices.

9. A fault simulation control method for a low-voltage substation assessment meter training device, characterized in that, Applying a low-voltage substation metering training device according to any one of claims 1-7, characterized in that: The steps of the convolutional neural network modeling are as follows: 1) Data preprocessing and format conversion: The collected operation trajectory data of trainees, including operation timestamps, operation device numbers, operation instruction codes, and device status parameter vectors, are used to form a multivariate sequence set; for missing values, an interpolation method based on Gaussian process regression is adopted, and the covariance matrix K is constructed using the kernel function. By solving the linear equations the interpolation coefficients are obtained , achieving high-precision filling; for outlier detection, the local outlier factor algorithm is introduced. By calculating the ratio of the local reachability density of sample x to the average local reachability density of neighborhood samples, outlier data points are identified and corrected; when normalizing the operation timestamps, Min-Max normalization combined with time series smoothing technology is adopted; an adversarial network is constructed by introducing a generator G and a discriminator D. The generator G generates normalized time series and the discriminator D judges the difference between the true normalization result and the pseudo result, and the formula is as follows: In formula (1), is a regulation parameter used to control the range of the timestamps after normalization. The non-uniform interval sequence is converted into a uniform time interval sequence through cubic spline interpolation; is a hyperparameter for balancing the adversarial loss; through this way of adversarial learning, the normalization process is promoted to not only meet the basic numerical transformation requirements but also learn a more discriminative and robust normalization pattern in the adversarial game, effectively coping with the noise and outlier interference in the data; Construct a dynamic sliding event window, and the window size is controlled by an adaptive parameter to convert the operation trajectory sequence into a three-dimensional tensor format , where N is the number of samples, T is the time step, M is the operation feature dimension, and D is the number of data channels; the mathematical expression: In the formula, is the feature conversion function for the D-th data channel, represents the operation data of the n-th sample at time step t and feature dimension m; 2) Spatiotemporal feature extraction: Input the processed three-dimensional tensor X into the spatiotemporal feature extraction module, and adopt a multi-scale parallel convolution architecture; for each time step, use convolution kernels of different sizes , where l represents the convolution layer number, k is the time dimension size of the convolution kernel, D is the number of input channels, is the number of output channels, and perform convolution operations; based on the meta-learning framework, enable the parameters of the convolution layer to be quickly and adaptively adjusted according to different task scenarios; during the training process, optimize the meta-network parameters through the meta-learning algorithm to make the generated convolution layer parameters more suitable for different types of training data. During the training process, optimize the meta-network parameters through the meta-learning algorithm to make the generated convolution layer parameters more suitable for different types of training data. The improved calculation process is as follows: Introduce a gating mechanism, and the output feature map of the l-th convolution layer is calculated by the formula: In formula (2), is the Sigmoid activation function, is the gating signal tensor represents element-wise multiplication, is the convolution kernel weight, is the bias term, is the pooling weight; Expand the receptive field through the dilated convolution layer and combine the residual connection mechanism to solve the problem of gradient disappearance; are the parameters of the meta-network. In this way, the network can quickly adjust the convolution calculation method under different low-voltage distribution network inspection meter training scenarios and improve the generalization ability of the model; in the pooling stage, adopt the adaptive weighted pooling method to dynamically adjust the pooling weight according to the local method of the feature map; 3) Application of the attention mechanism: Input the extracted spatiotemporal feature map Y into the attention mechanism module and adopt the multi-head self-attention mechanism; Introduce the spatiotemporal dynamic weight matrix , which is composed of the temporal dependence coefficient in the time dimension and the feature correlation coefficient in the space dimension. By capturing the continuity of the trainee's operation trajectory in the time series and the correlation of the operation device spatial layout, dynamically adjust the calculation of the attention weight. For each head h, the formula for calculating the attention weight is: In formula (3), are the query vector and the key vector respectively, obtained through linear transformation. is the dimension of the key vector; can be obtained by calculating the reciprocal of the operation time interval and the autocorrelation of the time series. Determined based on the physical connection relationship and functional relevance between operating devices; this method enables more precise focus on key operation links when processing the training data of low-voltage substation assessment meters. For example, in the fault handling process, higher weights are assigned to operation steps with sequential dependencies. Concatenate the multi-head attention outputs and perform a linear transformation to obtain a feature vector with attention; 4) Model training and parameter optimization: Input the feature vector Z with attention into the fully connected layer, adopt a hybrid loss function L, combine the cross-entropy loss , mean squared error loss and the regularization term . Combine the attention mechanism, introduce dynamic attention weights for each loss sub-term, and construct an attention weight matrix. , dynamically generate weights by calculating the importance of different tasks in the current training stage; the formula is: In formula (4), is the weight parameter. can be calculated through the self-attention mechanism based on the change rate of the task loss and the uncertainty factor of the prediction result, realizing the collaborative optimization between multiple tasks and improving the comprehensive performance of the model in the low-voltage substation assessment meter training task; 5) Prediction and evaluation: Use the trained model to predict new trainee operation trajectory data; for classification tasks, output the probability distribution of operation behavior categories; for regression tasks, output the quantitative score of operation skills; introduce an ensemble learning evaluation method, combine the weighted sum of multiple evaluation indicators as the final evaluation score , where is the index weight, is the i-th evaluation index; adopt the SHAP value analysis method to perform interpretability analysis on the model prediction results and identify the key operation features affecting the prediction results; generate virtual operation trajectory data through a generative adversarial network to expand the test data set and further verify the generalization ability and robustness of the model.

10. According to the method for simulating and controlling faults of a low-voltage substation metering training device according to claim 9, characterized in that: The working process of the reinforcement learning algorithm is as follows: 1) Constructed based on the spatio-temporal attention mechanism of Transformer For the training data sequence of the low-voltage substation assessment meter For the data features with the time dimension of D, calculate the attention weights through the multi-attention mechanism, and fuse the new multi-head attention mechanism of adaptive weight adjustment and multi-scale feature fusion. The formula of the multi-attention mechanism is as follows: In formulas (5), (6), and (7), Q, K, and V are the query, key, and value matrices respectively, which are obtained by linearly transforming the input data X; h is the number of heads; is the dimension of the key matrix; is the linear transformation parameter matrix of each head, is the output transformation matrix; this mechanism captures the long-range dependencies between data at different time scales and different spatial positions in parallel, and realizes the deep fusion of features across regions and time periods; A is the global adaptive weight matrix; G is the local adjustment matrix; is the Hadamard product, which realizes element-wise multiplication and weighting of the matrix; is the multi-scale feature fusion operation, which concatenates the features extracted by different convolution kernels; is a learnable scalar that dynamically adjusts the dimension scaling of the key matrix; is the scientific vector, and the non-linear change value matrix enhances the feature expression; 2) Introduction of the adversarial learning framework The discriminator D and the feature extractor G are introduced for adversarial training to improve the feature discrimination through the min-max game; the adversarial loss function is as follows: In Equation (8), is the true data distribution, is the noise distribution, and z is the noise vector; the discriminator D is used to distinguish the true data features from the generated features, forcing the feature extractor G to learn more discriminative representations and solving the problem of easy confusion of the meter data features in a complex electromagnetic environment; is the multi-scale discriminator, which consists of m sub-discriminators; is the gradient penalty coefficient; 3) Construction of the dynamic semantic encoding system Construct a semantic association network using a knowledge graph, map the data sample x to the semantic vector space to obtain the semantic encoding s(x); s(x) can be obtained by aggregating the relevant entity and relationship vectors; 4) Collaborative optimization of the federated learning framework Multiple training sites each have local data , and each site trains a local encoding model , and the server updates the global model parameters through an aggregation algorithm , and the formula is: where n is the number of training sites, is the data volume of the i-th site, enabling the encoding model to adapt to different training scenarios; is the aggregated global model parameter, is the local model parameter of the i-th site; is the data scale exponential factor; is the AGC performance index of the i-th site model; is the performance weight index; is the semantic consistency score between the j-th site and the knowledge graph; Knowledge graph constraint loss, which measures the consistency between model parameters and prior knowledge 5) Generative encoding strategy based on the variational autoencoder VAE The VAE generates virtual data samples by maximizing the evidence lower bound ELBO, and the formula is as follows: In formula (10), is the encoder, is the decoder, is the KL divergence, is the prior distribution of the latent variable z, which improves the generalization ability of the model; is the prior distribution of the latent variable z; represents the true data distribution; is the KL divergence, which measures the difference between the posterior distribution and the prior distribution; is the Jensen-Shannon divergence, which is used to measure the distance between the generated data distribution and the true data distribution; is a hyperparameter that balances each term, where controls the matching degree between the posterior and the prior, adjusts the similarity between the generated data and the true data, and optimizes the model performance by dynamically adjusting; 6) Implementation of the spatio-temporal Bayesian network module Introduce the meta-learning mechanism to quickly adjust the network structure and prior distribution of parameters according to different training tasks and fault simulation scenarios S ; Use the causal discovery algorithm to mine the causal relationships between data and construct the causal graph G=(V,E); Introduce the quantum superposition state into Bayesian inference for the multi-node probability distribution for accelerated calculation; 7) Implementation of the dynamic channel switching module Policy learning based on Deep Q-Network (DQN): Define the state space S, action space A, and reward function R(s,a). Learn the optimal channel switching policy through the deep Q-network. The Q-network minimizes the loss function: In Equation (11), is the target Q-value, is the discount factor, are the main network parameters, are the target network parameters, is the dynamic discount factor, is the entropy regularization coefficient, Entropy is the entropy of the action distribution; Build a digital twin to generate virtual training samples to feed back the reinforcement learning model; Use spectrum sensing to monitor the channel state C(t) and achieve dynamic spectrum sharing in combination with blockchain; 8) Implement the multi-node interaction module Deploy edge computing units at the meter nodes for data and make an interaction decision d; use holographic projection to display the node interaction relationship R = GNN(H, A) calculated by the graph neural network; adjust the training difficulty according to the trainee feedback information f through the emotion computing model; design a multi-agent interaction mechanism and simulate a strategy game by solving the Nash equilibrium of the game.

Images