Switch cabinet performance prediction method, system, equipment and medium

Abstract

The invention discloses a switch cabinet performance prediction method, system and device and a medium, and the method comprises the steps: collecting the real-time operation data of a switch cabinet, carrying out the preprocessing of the real-time operation data, and obtaining a key performance index; obtaining historical switch cabinet fault data, and training the historical switch cabinet fault data to obtain a multi-scale prediction model; combining the multi-scale prediction model with the key performance indexes to predict the prediction state of the switch cabinet; and taking corresponding measures according to the prediction state. According to the method, the key performance indexes are obtained by collecting the real-time operation data of the switch cabinet and preprocessing the real-time operation data, the prediction state of the switch cabinet can be accurately predicted in combination with the multi-scale prediction model obtained by training the historical switch cabinet fault data, and advanced prediction and fault early warning of the performance of the switch cabinet are achieved.

Description

Technical Field

[0001] This invention relates to the field of switchgear technology, and in particular to a method, system, device and medium for predicting the performance of switchgear. Background Technology

[0002] In power supply systems, especially with the rapid development of modern power systems towards high voltage, large generating units, and large capacity, the requirements for power supply reliability are becoming increasingly stringent. Switchgear, as a crucial device for opening, closing, controlling, and protecting electrical equipment, plays a vital role in the power generation, transmission, distribution, and energy conversion processes of the power system. The performance of switchgear directly affects the safety and stability of the entire power supply system; therefore, effective performance prediction of switchgear is of great significance.

[0003] In recent years, switchgear accidents have occurred frequently, causing economic losses, personal injury, and other adverse social impacts. The potential hazards and inherent defects mainly focus on wiring methods, internal arc release capabilities, internal insulation, heat generation, and anti-misoperation interlocking. Existing switchgear performance prediction methods suffer from the following problems: first, the prediction models are too simplistic to comprehensively assess the performance status of multiple subsystems within the switchgear; second, there is a lack of simulation analysis of fault risks under extreme operating conditions, resulting in low prediction accuracy. Therefore, a method capable of accurately predicting the performance status of switchgear in advance is urgently needed to reduce the occurrence of accidents. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the technical problem to be solved by the present invention is to provide a method, system, device and medium for predicting the performance of switchgear, so as to make accurate performance prediction of switchgear in advance and provide timely warning when the prediction result indicates that a failure may occur, thereby reducing the loss of switchgear failure.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting the performance of a switchgear, comprising: Real-time operating data of the switchgear is collected, and the real-time operating data is preprocessed to obtain key performance indicators; historical switchgear fault data is obtained, and a multi-scale prediction model is trained using the historical switchgear fault data; the multi-scale prediction model and the key performance indicators are combined to predict the predicted state of the switchgear; and corresponding measures are taken based on the predicted state.

[0007] As a preferred embodiment of the switchgear performance prediction method of the present invention, the step of preprocessing the real-time operating data includes: performing noise filtering on the real-time operating data to obtain preliminary processed data; performing data alignment on the preliminary processed data to obtain secondary processed data; and performing feature extraction on the secondary processed data to obtain the key performance indicators.

[0008] As a preferred embodiment of the switchgear performance prediction method of the present invention, the step of training a multi-scale prediction model using the historical switchgear fault data includes: dividing the historical switchgear fault data into a training set and a test set; constructing an initial prediction model; training the initial prediction model based on the training set to obtain a trained prediction model; testing the trained prediction model based on the test set; and determining the trained prediction model as the multi-scale prediction model when the loss of the trained prediction model reaches a preset threshold.

[0009] As a preferred embodiment of the switchgear performance prediction method of the present invention, the step of predicting the predicted state of the switchgear includes: inputting the key performance indicators into the multi-scale prediction model; calculating the predicted state of the switchgear through the multi-scale prediction model; wherein the predicted state of the switchgear includes the operating state of the switchgear and the predicted fault information of the components in the future time period, and the fault information includes the faulty components and the causes of the faults.

[0010] As a preferred embodiment of the switchgear performance prediction method of the present invention, the step of taking corresponding measures according to the predicted state includes: determining whether the predicted state of the switchgear indicates that a component will fail in the future time period; when the determination result is yes, issuing an early warning according to the cause of the failure, and replacing the corresponding component according to the cause of the failure.

[0011] As a preferred embodiment of the switchgear performance prediction method of the present invention, the multi-scale prediction model includes a contact welding risk prediction sub-model; the application steps of the contact welding risk prediction sub-model include: collecting core feature data; obtaining historical melting cases, and training the contact welding risk prediction sub-model through the historical melting cases; inputting the core feature data into the contact welding risk prediction sub-model to predict the contact welding risk of the switchgear.

[0012] The beneficial effects of this preferred technical solution are as follows: by conducting specialized analysis of core feature data through the contact welding risk prediction sub-model, the risk of contact welding can be predicted in a targeted manner, which improves the accuracy and pertinence of fault prediction for key components of switchgear and effectively reduces the probability of contact welding accidents.

[0013] As a preferred embodiment of the switchgear performance prediction method of the present invention, the step of obtaining the contact welding risk prediction sub-model through training the historical fusing cases includes: dividing the historical fusing cases into a fusing training set and a fusing test set; constructing an initial contact welding risk prediction sub-model; training the initial contact welding risk prediction sub-model according to the fusing training set to obtain a trained contact welding risk prediction sub-model; testing the trained contact welding risk prediction sub-model based on the fusing test set; and determining the trained contact welding risk prediction sub-model as the contact welding risk prediction sub-model when the loss degree of the trained contact welding risk prediction sub-model reaches a preset loss threshold.

[0014] Secondly, the present invention provides a switchgear performance prediction system, comprising: The data acquisition module is used to collect real-time operating data of the switchgear; The data preprocessing module is used to preprocess the real-time running data to obtain key performance indicators; The data acquisition module is used to acquire historical switchgear fault data; The model training module is used to train a multi-scale prediction model using the historical switchgear fault data. The prediction module is used to combine the multi-scale prediction model and the key performance indicators to predict the predicted state of the switchgear. The decision execution module is used to take corresponding measures based on the predicted state.

[0015] Thirdly, the present invention provides an electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the switch cabinet performance prediction method.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the switchgear performance prediction method.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention collects and preprocesses real-time operating data of switchgear to obtain key performance indicators. Combined with a multi-scale prediction model trained using historical switchgear fault data, it accurately predicts the switchgear's state, enabling early prediction of switchgear performance and fault warning. Compared to existing technologies, this invention's multi-scale prediction model comprehensively evaluates the performance status of multiple subsystems of the switchgear, overcoming the shortcomings of traditional methods such as single prediction models and incomplete evaluations. This improves the accuracy and reliability of predictions, effectively reducing the probability of switchgear failures and minimizing economic losses and safety incidents.

[0018] This invention also constructs a sub-model for predicting contact welding risks, performing specific risk predictions for key components of the switchgear. By injecting short-circuit current based on fault mode and effects analysis, it simulates contact welding risks under extreme conditions, predicting arc energy and contact erosion levels. This virtual simulation testing method can cover extreme conditions that are difficult to achieve in actual testing, not only reducing testing costs but also improving the accuracy of fault risk prediction under extreme conditions. Simultaneously, this invention combines historical load data and equipment aging models to predict the remaining lifespan of the switchgear, providing a scientific basis for equipment maintenance and replacement, realizing intelligent management throughout the equipment's lifecycle, and improving the operational reliability of the switchgear and the overall safety of the power supply system. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the overall process of the switchgear performance prediction method according to an embodiment of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for predicting the performance of a switchgear is provided, comprising the following steps S100~S400: S100. Collect real-time operating data of the switchgear, preprocess the real-time operating data, and obtain key performance indicators. S200. Obtain historical switchgear fault data, and train a multi-scale prediction model using the historical switchgear fault data. S300. Combine the multi-scale prediction model with the key performance indicators to predict the predicted state of the switchgear. S400. Take corresponding measures based on the predicted state.

[0023] It should be noted that switchgear, as a key piece of equipment in the power system, plays a crucial role in power generation, transmission, and distribution. During its operation, the electrical parameters, temperature, and mechanical condition of the switchgear fluctuate significantly, causing dynamic changes in its operating status and fault risk. Component performance also deteriorates due to load fluctuations and environmental factors. Furthermore, the high-voltage, high-current environment of the operating switchgear makes real-time monitoring of its internal condition difficult, often resulting in delayed fault warnings. Simultaneously, frequent contact operation and arcing cause continuous wear and erosion of the contact surfaces, harming equipment health. The complex operating environment also significantly impacts fault prediction and can lead to equipment damage due to aging. Therefore, health monitoring and performance prediction of switchgear are extremely important.

[0024] Therefore, to address the aforementioned issues in operation monitoring and performance prediction, steps S100-S400 are used to collect real-time operational data and extract key performance indicators. These are then combined with historical fault data to train a multi-scale prediction model, enabling accurate assessment of the switchgear's operational status. The multi-scale prediction model analyzes key performance indicators to achieve real-time prediction of fault risks for each component of the switchgear, dynamically monitor equipment performance trends, and provide early warnings of abnormal conditions. Simultaneously, targeted measures are taken based on the predicted status to achieve intelligent operation and maintenance management and fault prevention for the switchgear, effectively reducing the probability of faults and improving the reliability and safety of the power supply system.

[0025] Example 2, refer to Figure 1 As an embodiment of the present invention, a method for predicting the performance of switchgear is provided based on the above embodiment.

[0026] In this embodiment of the application, the step of preprocessing the real-time running data in step S100 includes A1~A3: A1. Noise filtering is performed on the real-time running data to obtain preliminary processed data.

[0027] Specifically, real-time operating data of the switchgear is collected through sensors. This real-time operating data includes, but is not limited to, electrical parameters, temperature data, mechanical status data, and environmental data. Electrical parameters are collected using CT / PT sensors, including current, voltage, and power factor. Temperature data is collected using infrared thermal imagers and fiber optic temperature measurement systems, focusing on monitoring the temperature distribution of busbar joints and circuit breaker contacts. Mechanical status data is collected using vibration sensors and partial discharge sensors to detect loosening of mechanisms and insulation degradation. Environmental data includes temperature, relative humidity, and dust concentration values ​​at any location on the front, back, left, right, and top of the switchgear.

[0028] Wavelet transform is used to filter noise in the real-time operating data, removing high-frequency interference signals such as electromagnetic interference (EMI) noise. The wavelet transform uses the Daubechies wavelet basis function, with a decomposition level of 3-5 layers. A thresholding method is used to remove noise components while retaining valid signal components. A data quality assessment mechanism is established, including data integrity checks, outlier identification, and signal-to-noise ratio calculation, to ensure the validity of the filtered data. Preliminary processed data is obtained after filtering. The data sampling frequency is set according to different parameter types: electrical parameters are sampled at a frequency of no less than 1 kHz, temperature data at 1 Hz, and mechanical status data at 10 kHz.

[0029] A2. Align the preliminary processed data to obtain secondary processed data.

[0030] Since real-time operational data originates from multiple sensor systems, and each sensor differs in sampling frequency, data format, and time reference, data alignment is required to achieve time synchronization of multi-source data. Timestamp synchronization technology is employed to calibrate the data from each sensor using a unified time reference, controlling the time synchronization error to within 1ms. A data buffer mechanism is established to cache and resample data with different sampling frequencies, achieving data frequency unification through linear interpolation or spline interpolation methods. To address data transmission latency, a latency compensation model is established to estimate and compensate for latency based on the network transmission path and device response time of each sensor. A data time-series alignment algorithm is used to identify time offsets between different data sequences through cross-correlation analysis, achieving precise time alignment. A multi-source data fusion framework is established, organizing and storing aligned electrical parameters, temperature data, mechanical status data, and environmental data according to a unified time series. The data storage format adopts a time-series database structure, supporting efficient time-series data querying and analysis. After data alignment, secondary processed data is obtained, ensuring the consistency and comparability of multi-source data across the time dimension.

[0031] A3. Perform feature extraction on the secondary processing data to obtain the key performance indicators.

[0032] Based on secondary processed data, key performance indicators reflecting the operating status and performance changes of the switchgear are extracted using feature engineering methods. Feature extraction includes three categories: time-domain features, frequency-domain features, and statistical features. Time-domain features include temperature rise rate, current change rate, and vibration amplitude, calculated using numerical differentiation methods. Frequency-domain features include current harmonic distortion rate (THD) and vibration spectrum characteristics, extracted using Fast Fourier Transform (FFT). Statistical features include data mean, standard deviation, peak value, and peak-to-peak value, calculated using statistical analysis methods. For temperature data, the temperature rise rate is calculated using the formula ΔT / Δt, where ΔT is the temperature change and Δt is the time interval. For electrical parameters, the current harmonic distortion rate (THD) is calculated using the formula... ,in, The effective value of the nth harmonic current. The fundamental current is the effective value. A feature library is established to store historical data and statistical characteristics of various key performance indicators, including normal operating range, abnormal thresholds, and trends. Feature selection algorithms are used, employing correlation analysis and principal component analysis (PCA) to screen features that contribute significantly to fault prediction, reducing feature dimensionality and improving the efficiency of subsequent model training. The final key performance indicators include, but are not limited to: temperature rise rate, current harmonic distortion rate, contact resistance, vibration intensity, partial discharge intensity, and ambient temperature and humidity. These indicators serve as input parameters for the multi-scale prediction model.

[0033] In one alternative implementation, data preprocessing can also employ deep learning feature extraction techniques. Specifically, a convolutional neural network (CNN) is used to extract deep features from infrared thermal images, automatically learning temperature distribution patterns through multi-layer convolution and pooling operations; a recurrent neural network (RNN) or long short-term memory network (LSTM) is used to learn features from time-series data, capturing the time dependencies of the switchgear's operating status; an autoencoder is introduced for data dimensionality reduction and feature compression, extracting the essential feature representations of the data; an end-to-end feature learning framework is established, directly inputting the raw data into the neural network model, which automatically learns the optimal feature representations; and transfer learning techniques are employed, utilizing models pre-trained in similar equipment or scenarios for feature extraction, improving the accuracy and generalization ability of feature extraction.

[0034] In another alternative implementation, data preprocessing can also combine expert knowledge and physical models. A feature engineering method based on physical mechanisms is established, designing feature extraction rules according to the electrical, thermodynamic, and mechanical characteristics of the switchgear; expert experience is introduced, combining the field experience of maintenance personnel to define fault characteristic indicators; fuzzy logic is used to handle the uncertainty of features, converting continuous variables into fuzzy membership values; a multi-scale feature extraction framework is established to extract features at different time scales (seconds, minutes, hours) and spatial scales (single point, local, global) to comprehensively reflect the operating status of the switchgear; an ensemble learning method is used to fuse multiple feature extraction results, obtaining the final key performance indicators through weighted averaging or voting mechanisms, thereby improving the robustness and reliability of feature extraction.

[0035] In this embodiment of the application, step S200, which involves training a multi-scale prediction model using the historical switchgear fault data, includes steps B1 to B4: B1. Divide the historical switchgear fault data into a training set and a test set.

[0036] Specifically, historical switchgear fault data is first obtained from the switchgear operation and maintenance database. This historical fault data includes historical key performance indicators and corresponding fault information. The historical key performance indicators are consistent with the key performance indicators extracted in step S100, including parameters such as temperature rise rate, current harmonic distortion rate, contact resistance value, vibration intensity, and partial discharge intensity. The fault information includes detailed records such as fault occurrence time, faulty component, fault type, fault cause, and fault severity. A data cleaning mechanism is established to perform quality checks on the historical fault data, removing samples with severe data loss, incorrect labeling, or outliers to ensure the effectiveness and reliability of the training data. To address the data imbalance problem, where normal operation data far exceeds fault data, data balancing techniques are used. These include oversampling methods (such as SMOTE synthetic minority class oversampling technology) to increase the number of fault samples, or undersampling methods to reduce the number of normal samples, ensuring that the ratio of normal samples to fault samples in the training data reaches a reasonable range. The historical switchgear fault data is divided into a training set and a test set according to chronological order or random sampling. The training set is used for model parameter learning, and the test set is used for model performance verification. The split ratio is typically set to 70%-80% for the training set and 20%-30% for the test set, ensuring a sufficient number of test set samples for reliable performance evaluation. A data standardization process is established to normalize or standardize the training and test sets respectively, mapping feature data of different dimensions to a unified numerical range, thereby improving the convergence speed and prediction accuracy of model training.

[0037] B2. Construct an initial prediction model. Train the initial prediction model based on the training set to obtain the trained prediction model.

[0038] A network architecture for a multi-scale prediction model is constructed, employing a multi-layer neural network structure capable of simultaneously processing feature information at different time and spatial scales. The model architecture includes an input layer, multiple hidden layers, and an output layer. The input layer receives key performance indicator data, the hidden layers learn complex mapping relationships in the data through nonlinear activation functions, and the output layer outputs prediction results including the operating status and fault risk of each component. The initial prediction model can employ a deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN), or a combination thereof. Considering the characteristics of time-series data, a Long Short-Term Memory (LSTM) network or gated recurrent unit (GRU) structure is preferred to effectively capture the time dependence and long-term trends of the switchgear's operating status. Model parameter initialization uses Xavier initialization or He initialization methods to ensure gradient stability in the early stages of training.

[0039] Configure model hyperparameters, including learning rate, batch size, number of hidden layer nodes, and dropout ratio. The initial learning rate is set to 0.001-0.01, and a learning rate decay strategy is used to gradually reduce the learning rate during training. The batch size is set to 32-128. The number of hidden layer nodes is determined based on the input feature dimension and task complexity, typically set to 2-10 times the input dimension. The dropout ratio is set to 0.2-0.5 to prevent overfitting. The model is trained using backpropagation and gradient descent optimization methods, preferably employing the Adam optimizer or RMSprop optimizer, which have adaptive learning rate adjustment capabilities. Define a loss function: cross-entropy loss for classification tasks, mean squared error loss for regression tasks, or a combined loss function to optimize multiple prediction targets simultaneously. Set the training epochs to 100-500, and use early stopping to monitor validation set performance. Training is terminated early when the validation loss no longer decreases for several consecutive epochs to avoid overfitting. During training, model checkpoints are saved periodically, and training metrics such as training loss curves, validation loss curves, and accuracy curves are recorded for model performance analysis and tuning. After training, the trained prediction model is obtained, with model parameters optimized using the training set data.

[0040] B3. Test the trained prediction model based on the test set.

[0041] The test set data is input into the trained prediction model to obtain the model's prediction output. The prediction output includes information such as the operating status classification of each component (normal, warning, fault), fault probability value, and fault type identification. A model performance evaluation system is established, using multiple evaluation metrics to comprehensively evaluate the model's prediction performance. For classification tasks, metrics such as accuracy, precision, recall, and F1 score are calculated. Accuracy reflects the overall proportion of correct predictions, precision reflects the proportion of actual faults among samples predicted as faults, recall reflects the proportion of correctly identified actual fault samples, and the F1 score comprehensively considers precision and recall. For multi-class problems, a confusion matrix is ​​plotted to analyze the prediction situation of each category and identify the model's weaknesses. For regression tasks, metrics such as root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are calculated to evaluate the degree of deviation between predicted and true values. Plot the ROC (Receiver Operating Characteristic Curve) and calculate the AUC (Area Under Curve) to evaluate the model's classification performance at different thresholds. An AUC value closer to 1 indicates better model performance. Perform error analysis to identify samples with large prediction errors and analyze the causes of these errors, such as inappropriate feature selection, incorrect sample labeling, or an unreasonable model structure.

[0042] B4. When the loss of the trained prediction model reaches a preset threshold, the trained prediction model is determined as the multi-scale prediction model.

[0043] A preset threshold is set for the performance of the multi-scale prediction model. This preset threshold is determined based on actual application requirements and industry standards. For switchgear fault prediction tasks, the required accuracy is no less than 90%, recall no less than 85%, F1 score no less than 0.88, and AUC no less than 0.92. The loss threshold is set to a loss function value on the test set below 0.1 or a loss function value on the validation set that does not decrease for 10 consecutive rounds. The trained prediction model is then judged to meet the preset threshold conditions. If all performance indicators on the test set reach or exceed the preset threshold, and the loss is lower than the set value, the multi-scale prediction model is considered successfully trained. The trained prediction model is then designated as a multi-scale prediction model, and its parameters and structural information are saved for subsequent switchgear performance prediction. If the performance metrics do not reach the preset threshold, multi-scale prediction model optimization is required. Optimization methods include: adjusting model hyperparameters, such as increasing the number of hidden layers or nodes, adjusting the learning rate, and modifying the batch size; improving feature engineering, adding more effective features or combining features; expanding training data, collecting more fault samples or using data augmentation techniques; and changing the model architecture, trying different neural network structures or ensemble learning methods.

[0044] Repeat steps B2 and B3 to iteratively train and test the multi-scale prediction model until its performance reaches a preset threshold. Establish a version management mechanism for the multi-scale prediction model, recording model parameters, performance metrics, training configurations, and other information for each training iteration, supporting backtracking and comparative analysis. The final multi-scale prediction model possesses good generalization ability and prediction accuracy, and can accurately predict the operating status and fault risks of switchgear.

[0045] In one alternative implementation, multi-scale prediction model training can also employ ensemble learning methods to improve prediction performance. Specifically, multiple base prediction models are constructed, including different types of models such as decision trees, random forests, support vector machines, and neural networks, each learning data features from different perspectives; a Bagging method is used to generate multiple training subsets through bootstrapping, training multiple models separately, and then fusing the prediction results through voting or averaging; a Boosting method, such as AdaBoost or XGBoost, is used to train multiple weak learners sequentially, with each learner focusing on samples that the previous learner predicted incorrectly, gradually improving the overall prediction capability; a Stacking method is used to input the prediction results of multiple base models as features into a meta-learner, which learns the optimal fusion strategy; and a model selection mechanism is established to automatically select the most suitable model or model combination based on different fault types and operating scenarios, improving the flexibility and adaptability of prediction.

[0046] In another alternative implementation, model training can also combine transfer learning and online learning techniques. Transfer learning utilizes a pre-trained model on other types of switchgear or similar equipment as the initial model, and fine-tunes it to adapt to the characteristics of the current switchgear, reducing training data requirements and training time. A domain adaptation mechanism is established to reduce the distribution differences between the source and target domains through adversarial training or feature alignment methods, improving the model's transfer performance. Online learning enables the model to continuously learn from new data, dynamically updating model parameters to adapt to equipment aging and changes in the operating environment. An incremental learning framework is established so that when new fault samples are collected, the entire model does not need to be retrained; only local updates are required. An active learning strategy intelligently selects the most valuable samples for labeling and training, improving data utilization efficiency. A model monitoring and automatic retraining mechanism is established to trigger the model retraining process when a decline in model performance or a change in data distribution is detected, ensuring the model always maintains good predictive performance.

[0047] In this embodiment of the application, the step of predicting the predicted state of the switchgear in step S300 includes C1~C2: C1. Input the key performance indicators into the multi-scale prediction model.

[0048] The key performance indicator data extracted in step S100 undergoes format conversion and preprocessing to meet the input requirements of the multi-scale prediction model. Key performance indicator data includes real-time acquired multi-dimensional feature parameters such as temperature rise rate, current harmonic distortion rate, contact resistance, vibration intensity, partial discharge intensity, and ambient temperature and humidity. A data interface layer is established to achieve seamless integration between the real-time data stream and the prediction model. Data transmission uses a standardized protocol to ensure data integrity and consistency. The key performance indicators are normalized using the same normalization parameters as in the training phase, mapping each dimension of feature data to a unified numerical range, typically [value missing]. or Intervals eliminate the influence of dimensional differences between different characteristics.

[0049] To address the characteristics of time-series data, a time window sequence is constructed, organizing historical data over a period into a time-series input format. The time window length is set based on the model structure and prediction task, typically ranging from 10 to 100 time steps. A data caching mechanism is established to maintain a sliding time window, updating the latest key performance indicator data in real time and removing outdated historical data to ensure the timeliness of the input data. Integrity checks are performed on the input data to identify missing and outlier values. Missing data is handled using interpolation or forward imputation methods. Criteria or box plot methods are used to identify and process outlier data. The processed key performance indicator (KPI) data is organized according to the input tensor format defined by the model, including batch, time, and feature dimensions, forming a standardized input data structure. The KPI data is then fed into the multi-scale prediction model through the model input interface, triggering the model's forward inference computation process.

[0050] C2. The predicted state of the switchgear is calculated using the multi-scale prediction model.

[0051] The multi-scale prediction model receives key performance indicator inputs and performs forward propagation calculations according to the trained network structure and parameters. The model's multi-scale characteristic is reflected in its ability to simultaneously handle prediction tasks at short-term, medium-term, and long-term time scales. Short-term predictions focus on state changes within the next 1-24 hours, medium-term predictions focus on trend evolution within the next 1-7 days, and long-term predictions focus on performance degradation within the next 1-3 months. The model extracts deep feature representations of the input data through a multi-layer neural network structure. The hidden layers use nonlinear activation functions (such as ReLU, Tanh, and Sigmoid) to achieve complex nonlinear mappings. The attention mechanism layer automatically identifies key features that significantly impact the prediction results, and the time series modeling layer (such as LSTM or GRU units) captures long-term dependencies and dynamic evolution patterns in the time series. The model output layer adopts a multi-task learning architecture, simultaneously outputting multiple prediction targets, including the overall operating status classification of the switchgear, the health scores of each key component, and the probability values ​​of fault risk. The operating status classification typically includes four levels: normal state, warning state, abnormal state, and fault state. The probability distribution of each state is output through the Softmax activation function, and the category with the highest probability is selected as the predicted state. Key components include circuit breaker contacts, busbar joints, insulation supports, operating mechanisms, and secondary circuits. The model outputs a health score (0-100) for each component; a lower score indicates more severe performance degradation and a higher risk of failure. The failure risk probability value is output using a Sigmoid activation function, ranging from 0 to 1, representing the likelihood of a component failing within the predicted time period. A risk threshold is set (typically 0.7-0.8); when the probability value exceeds the threshold, it is considered a high-risk state.

[0052] The predicted status of the switchgear includes the operating status of the switchgear and the predicted fault information of the components in the future time period. The fault information includes the faulty component and the cause of the fault.

[0053] In one alternative implementation, the predicted state calculation can also incorporate uncertainty quantification techniques. A Bayesian deep learning approach is employed to introduce a probability distribution into the model. The posterior distribution of model parameters is estimated using variational inference or Markov Chain Monte Carlo (MCMC) methods, thereby quantifying the uncertainty of the prediction results. An ensemble prediction method is used, running multiple trained model instances to predict the same input, and assessing the reliability of the predictions through the variance or standard deviation of the prediction results. A confidence interval estimation mechanism is established to provide a confidence interval for each predicted output, such as predicting the failure probability of a component as 0.75 ± 0.08 (95% confidence interval), helping decision-makers understand the credibility of the prediction results. Monte Carlo dropout technology is used to maintain the activation of the dropout layer during the inference phase, obtaining the prediction distribution through multiple forward propagations, and evaluating the cognitive and data uncertainties of the model. A prediction quality assessment index is established, issuing a warning when the prediction uncertainty is too high, indicating the need for manual review or the collection of more data.

[0054] In another alternative implementation, the predicted state calculation can also integrate multimodal data and external knowledge. This involves integrating multimodal information such as image data (infrared thermal images, visible light images), audio data (equipment operation sounds), and text data (historical maintenance records), and extracting cross-modal features through a multimodal fusion neural network to improve the comprehensiveness of the prediction. It also involves introducing external knowledge graphs to construct knowledge representations of the switchgear equipment itself, fault modes, and causal relationships, combining deep learning models with knowledge reasoning to achieve a hybrid prediction driven by both data and knowledge. Furthermore, it employs graph neural networks (GNNs) to model the topological relationships and mutual influences between internal components of the switchgear, capturing the propagation patterns of faults among components. A context-aware mechanism is established, considering contextual information such as the equipment's operating history, maintenance records, and load curves to improve the specificity of the prediction. Finally, causal inference methods are used to identify true causal relationships rather than simple correlations, improving the accuracy and interpretability of fault cause analysis.

[0055] In this embodiment of the application, step S400, which involves taking corresponding measures based on the predicted state, includes steps D1 to D2: D1. Determine whether the predicted status of the switchgear indicates that a component will fail in the future.

[0056] Specifically, a fault determination mechanism is established to comprehensively analyze and judge the predicted status of the switchgear obtained in step S300. The determination mechanism is based on a multi-dimensional threshold system, including multiple judgment criteria such as fault probability threshold, health score threshold, and key performance indicator threshold. First, the fault probability values ​​of each key component in the predicted status are extracted and compared with preset fault probability thresholds. The fault probability thresholds are set in stages according to the importance of the component and the severity of the fault consequences. Lower thresholds (e.g., 0.6-0.7) are set for core components such as circuit breaker contacts, and higher thresholds (e.g., 0.8-0.9) are set for auxiliary components. When the fault probability of a component exceeds its corresponding threshold, it is preliminarily determined that the component has a fault risk. Next, the health score of each component is analyzed. The health score uses a percentage system. Under normal conditions, the health score should be above 80 points. When the health score is below 60 points, it is judged as an abnormal state; below 40 points, it is judged as a high-risk state. Recheck key performance indicators to ensure they do not exceed normal operating ranges. Establish normal value ranges and warning thresholds for key performance indicators, including warning values ​​for temperature rise rate (e.g., exceeding 5℃ / min), current harmonic distortion rate (e.g., THD > 10%), contact resistance (e.g., exceeding 200% of the initial value), vibration intensity, and partial discharge intensity. When key performance indicators exceed warning thresholds, make a comprehensive judgment based on failure probability and health score.

[0057] A multi-level judgment logic is established, employing decision trees or rule engines for automated judgment. Level 1 judgment: Check the overall operational status; if the predicted status is a fault or abnormal state, directly determine that there is a fault risk. Level 2 judgment: Check key components one by one, identifying components whose fault probability exceeds a threshold or whose health score is below the warning line. Level 3 judgment: Analyze the changing trends of key performance indicators, identifying abnormal patterns of rapid deterioration or sudden changes in indicators. Level 4 judgment: Combine historical data and statistical patterns to identify abnormal behaviors deviating from the normal operating trajectory. Fuzzy logic is used to handle boundary cases; when the fault probability approaches the threshold or multiple indicators are simultaneously in a critical state, a comprehensive judgment result is obtained through fuzzy reasoning. A time window verification mechanism is established to avoid misjudgments caused by instantaneous fluctuations; a fault judgment is only triggered when the abnormal state persists for a certain period (e.g., 5-10 minutes) or recurs at multiple consecutive time points. An expert rule base is introduced to store the experience knowledge of maintenance personnel and typical fault patterns; the credibility of the judgment is enhanced when the prediction result matches the expert rules. A tiered risk assessment system is established, classifying fault risks into four levels: low risk, medium risk, high risk, and emergency risk. Low risk indicates a slight performance degradation of the component but no immediate risk of failure; medium risk indicates potential issues requiring monitoring; high risk indicates a possible failure within a future period (e.g., 7-30 days) requiring immediate attention; and emergency risk indicates an imminent failure (e.g., within 24 hours) requiring immediate action. The assessment process and basis are recorded, including key indicators triggering the assessment, threshold comparison results, risk level, and assessment time. A assessment log is established for auditing and analysis. The final assessment result is output, clearly indicating whether there is a risk of component failure within a future period. If the assessment result is yes, proceed to step D2 to implement corresponding measures; if the assessment result is no, continue normal monitoring and return to step S100 for the next round of data collection and prediction.

[0058] D2. When the judgment result is yes, issue an early warning based on the cause of the fault, and replace the corresponding component based on the cause of the fault.

[0059] When the determination result of step D1 indicates that a component poses a risk of failure in the future, the emergency response mechanism is immediately activated, and graded and categorized response measures are taken according to the cause of the failure and the risk level. First, the early warning function is executed, establishing a multi-level early warning system, including local early warning, remote early warning, and linked early warning. Local early warning issues audible and visual alarm signals through the switchgear monitoring terminal; indicator lights installed at the equipment site change color (green for normal, yellow for early warning, and red for failure) and emit a buzzer, alerting on-site inspection personnel. Remote early warning pushes early warning information to the monitoring center and operation and maintenance management platform via the communication network. The monitoring center's large-screen display system displays the location of the early warning equipment, the faulty component, and the risk level in real time. Simultaneously, relevant responsible personnel, including equipment managers, operation and maintenance team leaders, and technical experts, are notified through various means such as SMS, telephone, and mobile APP. The early warning information includes: early warning time, equipment number, equipment location, name of the faulty component, description of the cause of the failure, risk level, expected failure time, and suggested handling measures, among other detailed information. The linkage early warning system integrates with other systems, triggering the work order system to automatically generate maintenance work orders. The work orders include fault diagnosis information, processing priority, suggested processing solutions, and a list of required spare parts, which are then assigned to the corresponding maintenance personnel for execution.

[0060] Based on the cause of the fault, develop targeted handling plans and establish a knowledge base mapping fault causes to handling measures. For overheating faults, handling measures include: checking the load condition and, if necessary, transferring the load to reduce the operating current; checking the heat dissipation system and cleaning dust from vents and cooling devices; checking contact surfaces and grinding or replacing oxidized or ablated contacts; installing temperature monitoring equipment and increasing the sampling frequency to once per minute. For electrical faults, handling measures include: using a partial discharge detector for precise detection to locate insulation defects; using an infrared thermal imager for comprehensive inspection to identify poor electrical contacts; measuring insulation resistance to assess the aging degree of insulation materials; and, if necessary, replacing insulation components or performing insulation reinforcement. For mechanical faults, handling measures include: checking the operating mechanism and tightening loose parts; checking springs and transmission mechanisms and replacing fatigued parts; conducting mechanical characteristic tests to test parameters such as opening and closing time, stroke, and speed; and lubricating and maintaining moving parts to restore mechanical performance. For environmental faults, the following measures can be taken: improve ventilation in the switchgear room and reduce ambient temperature and humidity; install dehumidifiers or heaters to prevent condensation; clean dust and pollutants inside and outside the switchgear; and install environmental monitoring sensors to continuously monitor changes in environmental parameters.

[0061] The decision to replace a component is based on the cause of the failure and the level of risk. Component replacement decision rules are established as follows: When a component's health score is below 30 or its failure probability exceeds 0.9, the component is considered severely degraded and must be replaced immediately; when a component has irreparable defects (such as severe contact erosion, insulation breakdown, or mechanical breakage), it must be replaced; when a component is repairable but the repair cost is close to or exceeds the replacement cost, direct replacement is recommended; when a component is nearing or has exceeded its design life and its performance has significantly deteriorated, preventative replacement is recommended. Develop a component replacement plan, including: determining the replacement time window and arranging emergency replacement (within 24 hours), planned replacement (within 7 days), or periodic replacement (within 30 days) according to the risk level; preparing spare parts and allocating the corresponding models and specifications from the spare parts warehouse, and making emergency purchases if the inventory is insufficient; developing a power outage plan, coordinating load scheduling, arranging equipment power outage times, and notifying affected users; organizing a construction team and allocating construction personnel with appropriate qualifications and experience according to the complexity of the work; preparing construction tools and safety protection equipment to ensure construction safety; and developing a construction plan and safety measures, clarifying construction steps, quality control points, and emergency plans.

[0062] The component replacement process includes: implementing safety measures such as power outage, voltage testing, and grounding of the switchgear according to safety operating procedures; removing the faulty component and recording any abnormalities found during removal; conducting failure analysis on the removed faulty component, photographing and recording the damage status to verify the accuracy of the fault cause assessment; installing the new component, ensuring correct installation location, reliable fastening, and good contact; conducting electrical and mechanical performance tests, including insulation resistance testing, withstand voltage testing, contact resistance measurement, and mechanical characteristic testing, to ensure the new component meets performance standards; restoring equipment operation by resuming switchgear operation according to power-on operating procedures; conducting trial operation observation, continuously monitoring key performance indicators, and confirming normal equipment operation. After replacement, updating the equipment ledger and maintenance records, recording information such as replacement time, replaced component, fault cause, and handling process; updating the equipment parameters of the multi-scale prediction model, incorporating the performance baseline of the new component into the model; summarizing lessons learned, analyzing the underlying causes of the fault, proposing improvement suggestions, and refining preventive measures. A closed-loop management mechanism is established to continuously track the operating status of the replaced equipment, verify the effectiveness of the handling measures, and initiate in-depth investigation and expert consultation if the problem is not resolved or recurs.

[0063] In one alternative implementation, emergency response can also employ an intelligent decision support system. This includes establishing an AI-based fault diagnosis expert system that integrates a fault diagnosis knowledge base, case library, and inference engine to automatically generate detailed fault analysis reports and handling suggestions; using augmented reality (AR) technology to assist on-site maintenance, allowing maintenance personnel to wear AR glasses to view 3D models of equipment, fault location markings, and operation guidance information, improving maintenance efficiency and accuracy; establishing a remote expert support system, enabling on-site personnel to receive real-time remote guidance from experts through video conferencing and data sharing; employing digital twin technology to simulate the effects of maintenance plans in a virtual environment and verify the feasibility of handling measures; establishing a fault contingency plan library, pre-developing standardized handling procedures for common fault types to achieve rapid response; and using mobile work terminals, allowing maintenance personnel to receive work orders, view equipment information, record handling processes, and upload on-site photos via handheld devices or tablets, achieving full-process information management of operations.

[0064] In another alternative implementation, emergency response can be combined with preventative maintenance strategy optimization. This involves establishing a condition-based maintenance (CBM) system, scheduling maintenance activities based on the actual operating status of equipment rather than fixed cycles, improving the targeting and economy of maintenance; employing predictive maintenance (PdM) strategies to predict the remaining service life of components in advance, performing maintenance or replacement at the optimal time before failure occurs, balancing maintenance costs and reliability; establishing a maintenance optimization model to comprehensively consider multiple factors such as failure risk, maintenance costs, power outage losses, and spare parts inventory to formulate the optimal maintenance plan; adopting a group maintenance strategy to coordinate the maintenance sequence and resource allocation when multiple switchgears issue warnings simultaneously, improving overall maintenance efficiency; establishing a spare parts optimization management system to dynamically adjust spare parts inventory strategies based on failure prediction results, ensuring the availability of critical spare parts while reducing inventory costs; and establishing a maintenance knowledge management platform to systematically accumulate maintenance experience and failure cases, continuously optimizing maintenance processes and technical standards, forming a virtuous cycle of maintenance knowledge and a continuous improvement mechanism.

[0065] It is also important to know that the multi-scale prediction model includes a sub-model for predicting the risk of contact welding. It should be noted that switchgear involves numerous subsystems and related parameters. Predicting faults across all subsystems using a single model would require processing excessive data, resulting in a massive mesh (tens of millions of cells) and unacceptable computation time. Therefore, a dedicated sub-model for predicting contact welding risk was constructed specifically for the critical contact system of the switchgear, enabling accurate prediction of this serious fault. Contact welding is one of the most dangerous faults in switchgear. When a short-circuit current passes through the contacts, the immense Joule heating causes localized overheating and even melting of the contact surface. During opening, the liquid metal adheres, preventing contact separation and potentially leading to equipment explosions, personal injury, and other serious consequences. Traditional periodic maintenance struggles to detect potential contact welding hazards early. However, the dedicated contact welding risk prediction sub-model provides early warnings based on real-time monitoring data, reducing the probability of contact welding accidents.

[0066] The application steps of the contact welding risk prediction sub-model include E1~E3: E1. Collect core feature data.

[0067] Specifically, a dedicated data acquisition system for predicting contact welding risks is established. The core characteristic data consists of high-frequency, high-precision monitoring parameters specifically designed for the contact welding mechanism and characteristics. These core characteristic data include multi-dimensional parameters such as dynamic resistance values, infrared thermographic data, transient current, opening speed, triaxial acceleration, and metal vapor concentration. Dynamic resistance values ​​are acquired in real-time using a contact resistance measuring device with a sampling frequency of 100Hz-1kHz. This reflects the electrical contact quality of the contact surface; under normal conditions, the contact resistance remains stable, but increases when the contact surface oxidizes, ablates, or experiences poor contact. Infrared thermographic data is acquired using a high-resolution infrared thermal imager with a sampling frequency of 1-10Hz. This focuses on monitoring the temperature distribution and hot spots on the contact surface. The thermographic data is stored in matrix form, containing the temperature value of each pixel in the contact area, enabling the identification of localized hot spots. Transient current is acquired using a high-speed current sensor with a sampling rate of no less than 10kHz. This captures current surge waveforms during transient processes such as short circuits and closing. The peak value and rate of change of the transient current are closely related to the contact welding risk. The tripping speed is measured in m / s using displacement or velocity sensors, reflecting the rapidity of the circuit breaker's tripping action. A slow tripping speed prolongs the arc burning time, increasing the risk of contact erosion and welding. Triaxial acceleration is acquired using a high-frequency vibration sensor with a sampling rate of 50kHz, monitoring the mechanical vibration characteristics during circuit breaker operation. The vibration signal reflects the impact characteristics of contact separation and the health status of the mechanical system. Metal vapor concentration is measured in ppm (parts per million) using a spectrometer or gas sensor. Severe arc erosion of the contacts generates metal vapor, and an increase in vapor concentration is a significant indicator of contact damage. A synchronous acquisition mechanism for core characteristic data is established, with all sensors using a unified time base for timestamping to ensure time consistency of data from different sources, with time synchronization errors controlled within 0.1ms. To address the differences between high-frequency data (e.g., triaxial acceleration at 50kHz, transient current at 10kHz) and low-frequency data (e.g., infrared thermography at 1Hz), a multi-sampling-rate data fusion mechanism is established, achieving a unified data organization format through time alignment and data interpolation. A trigger-based data acquisition mode is established to automatically initiate high-speed data acquisition and capture complete transient process data when critical events such as switching operations, short-circuit faults, and load surges occur in the switchgear. A data preprocessing workflow is established to filter, denoise, and remove outliers from the acquired core feature data; background temperature compensation and emissivity correction are performed on infrared thermographic data; and baseline drift correction and power frequency interference filtering are performed on transient current data. Feature engineering processing is implemented to extract higher-level feature parameters from the original core feature data, such as the rate of change of dynamic resistance, the highest temperature and temperature gradient of the infrared thermographic image, the peak value and waveform distortion rate of the transient current, the standard deviation of the switching speed, the energy spectrum of the vibration signal, and the cumulative value of the metal vapor concentration.A core feature database is established to store and manage the collected data in real time. The database adopts a time-series database architecture, which supports high-speed data writing and fast query analysis.

[0068] E2. Obtain historical melting failure cases, and train the contact welding risk prediction sub-model using the historical melting failure cases.

[0069] Historical contact failure cases were collected from multiple sources, including equipment operation and maintenance records, fault record databases, and test reports. These historical failure cases include actual cases of contact welding failures and cases simulated through testing. Actual failure cases include key characteristic data from a period prior to the failure (typically 1-30 days before the failure), data records at the time of the failure, post-failure equipment inspection reports, and failure analysis results. Simulated test cases are derived from controlled testing environments such as short-circuit tests, type tests, and reliability tests, and include preset test condition parameters (such as short-circuit current amplitude, duration, and contact material), key characteristic data collected during the test, and post-test assessments of contact damage (such as ablation depth, weld marks, and metallographic analysis results). Establish criteria for judging contact fusion welding, clarifying the conditions under which contact fusion welding is identified: Welding cases confirmed by metallographic analysis, where the melting and resolidification characteristics of the contact surface microstructure are observed under a microscope; an operating force exceeding 150% of the standard value and inability to open the circuit normally indicate contact adhesion; a sudden increase in contact resistance exceeding 300% accompanied by sustained high temperature indicates severe deterioration of the contact surface. Collect historical melting cases of different types of contacts, including cases of copper-tungsten alloy (Cu-W) contacts, silver-nickel alloy (Ag-Ni) contacts, copper-chromium alloy (Cu-Cr) contacts, etc., because the melting points, thermal conductivity, and arc resistance of contacts of different materials vary significantly, resulting in different welding mechanisms and characteristics. For each historical fuse failure case, detailed case attributes are labeled, including contact material type, rated voltage level, rated current, short-circuit current amplitude, number of operations, service life, environmental conditions, and other background information. The evolution trend of core characteristic data before the failure is labeled to identify key early warning indicators. The severity level of fusion welding is labeled, such as mild fusion welding (local adhesion on the contact surface but separable), moderate fusion welding (contact separation is difficult and requires significant operating force), and severe fusion welding (contacts are completely adhered and cannot be separated). A historical fuse failure case database is established to store complete case data and labeling information. The database includes positive samples (cases where fusion welding occurred) and negative samples (normal cases where fusion welding did not occur), with a positive-to-negative sample ratio designed to be 1:3 to 1:5 to ensure the model can learn the distinguishing features between normal and abnormal states. The total sample size of the database is no less than 500 cases, of which no less than 100 are fusion welding cases, covering diverse cases of different equipment types, different operating conditions, and different failure modes.

[0070] The steps for training the contact welding risk prediction sub-model using the historical melting failure cases include E2.1 to E2.4: E2.1 Divide the historical circuit breaker cases into a circuit breaker training set and a circuit breaker test set.

[0071] The collected historical circuit breaker cases were divided into datasets to ensure the representativeness and independence of the training and test sets. Stratified sampling was used to ensure a consistent ratio of positive and negative samples in both sets, avoiding model bias caused by uneven sample distribution. Stratification was based on key attributes such as contact material type, equipment voltage level, and fault severity, ensuring that each category of cases is representative in both sets. The partition ratio was set to 75% for the training set and 25% for the test set. For a total sample size of 500 cases, the training set would contain approximately 375 cases, and the test set approximately 125 cases. A time-series partitioning principle was adopted, assigning earlier cases to the training set and later cases to the test set. This simulates scenarios where historical data is used to predict future faults in real-world applications, avoiding overly optimistic assessments due to data leakage. For multiple cases involving the same equipment, all cases from the same equipment were grouped into the same dataset to avoid highly correlated samples between the training and test sets. A data augmentation mechanism was established to expand the training set to address the insufficient sample size of circuit breaker cases. Data augmentation methods include: time warping, which slightly stretches or compresses the time axis of time-series data; noise injection, which adds small-amplitude random noise to the original data to simulate measurement errors; sliding window sampling, which extracts multiple overlapping time window samples from long-term series data; and synthetic minority oversampling (SMOTE), which generates synthetic samples in the feature space. After data augmentation, the training set sample size can be expanded to 2-3 times the original size, but the degree of augmentation must be controlled to avoid introducing unrealistic data patterns. Data standardization is performed on both the training and test sets. The mean and standard deviation of each feature in the training set are calculated, and the training set data is Z-score standardized. The test set uses the same standardization parameters as the training set to undergo the same transformation, ensuring consistency in data preprocessing. A dataset management mechanism is established, assigning a unique identifier to each dataset and recording metadata such as the dataset partitioning method, sample size, class distribution, and creation time, supporting dataset version management and traceability. The partitioned circuit breaker training and test sets are stored in a standard format for easy model training and evaluation.

[0072] E2.2 Construct an initial contact welding risk prediction sub-model. Train the initial contact welding risk prediction sub-model based on the welding training set to obtain the trained contact welding risk prediction sub-model.

[0073] The network architecture of the contact welding risk prediction sub-model is designed. This sub-model is specifically optimized for contact welding prediction tasks and can effectively handle multi-source heterogeneous core feature data. The model architecture adopts a multi-channel input design, with dedicated processing channels for different types of core feature data. The time-series data channel processes time-series data such as dynamic resistance values, transient currents, and triaxial accelerations, using a one-dimensional convolutional neural network (1D-CNN) or LSTM network to extract time-series features. The image data channel processes infrared thermal imaging data, using a two-dimensional convolutional neural network (2D-CNN) to extract spatial temperature distribution features. The scalar data channel processes scalar parameters such as opening speed and metal vapor concentration, using fully connected layers for feature transformation. The features extracted from each channel are concatenated or weighted and fused in a fusion layer to form a comprehensive feature representation. The backbone network of the model adopts a deep neural network structure, containing multiple hidden layers for deep feature learning. To address the physical mechanism of contact welding prediction, an attention mechanism module is introduced to automatically identify the features and time periods with the greatest impact on welding risk, such as the peak of short-circuit current and the moment of temperature change. Residual connections are used to alleviate the vanishing gradient problem in deep networks and improve model training efficiency. The output layer adopts a binary classification structure, outputting the probability value of contact welding risk, with a probability value close to 1 indicating high risk and close to 0 indicating low risk. The model parameter size is designed according to data complexity, typically containing 100,000 to 500,000 trainable parameters. The model parameters are initialized using the He initialization method, suitable for networks with ReLU activation functions. Model hyperparameters are set, with an initial learning rate of 0.001, using a cosine annealing learning rate scheduling strategy; batch size is set to 16-32; dropout ratio is set to 0.3-0.5; and weight decay coefficient is set to 1e-4 for L2 regularization. A loss function is defined, using a weighted cross-entropy loss function to address the class imbalance problem in contact welding prediction, assigning higher weights (e.g., 3-5 times) to welding cases (minority class) to ensure the model prioritizes the learning of welding cases. The loss function formula is: ,in, For real labels, To predict probabilities, and Class weights are used. The Adam optimizer is employed for model training. The Adam optimizer combines the advantages of momentum and adaptive learning rate, resulting in fast and stable convergence. Mini-batch gradient descent is used during training. In each iteration, a batch of data is randomly sampled from the circuit breaker training set, the loss function and gradient are calculated, and the model parameters are updated. The training epochs are set to 200-500, with each epoch traversing the entire training set. During training, 10-20% of the samples in the training set are used as a validation set to monitor the training process and adjust hyperparameters. Model performance is evaluated on the validation set every 10-20 epochs, calculating the validation loss and validation accuracy, and plotting training curves to observe model convergence. An early stopping mechanism is used to prevent overfitting. The loss function value on the validation set is monitored; when the validation loss stops decreasing or begins to increase after 20-30 consecutive epochs, training is terminated early, and the model parameters at the point of lowest validation loss are restored. A learning rate decay strategy is employed. When the validation loss does not improve within 10 epochs, the learning rate is reduced to 50% of its original value to facilitate fine-tuning of the model in the later stages of training. During training, model checkpoints are saved periodically, including information such as model weights, optimizer status, and training epochs, to support recovery after training interruptions. Training logs are recorded, including metrics such as training loss, validation loss, training accuracy, validation accuracy, and learning rate for each epoch, used for analyzing the training process and diagnosing problems. After training, a trained sub-model for predicting contact welding risks is obtained. The model parameters have been fully optimized using a welding training set, possessing preliminary contact welding risk identification capabilities. To address the differences in contact materials, a material-specific sub-model training strategy is established. First, a general base model is trained using mixed data from all materials to learn the common characteristics of contact welding. Then, dedicated sub-models are trained for each material (e.g., copper-tungsten alloy, silver-nickel alloy). Using transfer learning, the general model is used as a pre-trained model, fine-tuned using the material-specific data. Only a few training epochs (50-100 epochs) are needed to obtain a high-performance material-specific model. During fine-tuning, the parameters of the first few layers are frozen and only the last few layers are trained, or a small learning rate (such as 0.0001) is used to fine-tune all layers. This method utilizes common knowledge across materials while capturing the specific characteristics of each material, achieving good predictive performance even with limited data.

[0074] E2.3 Test the trained contact welding risk prediction sub-model based on the aforementioned fusing test set.

[0075] The test set data for contact welding risk prediction is input into the trained sub-model, and model inference is performed to obtain the prediction results. During testing, the model is in evaluation mode, and random operations such as dropout during training are disabled to ensure the determinism and repeatability of the prediction results. For each case in the test set, the model outputs a probability value of contact welding risk, ranging from 0 to 1. A judgment threshold is set, typically 0.5. When the predicted probability is greater than 0.5, it is judged as welding risk; when it is less than or equal to 0.5, it is judged as normal. The judgment threshold can be adjusted according to actual application needs. If it is desired to improve recall and reduce false negatives, the threshold can be lowered to 0.3-0.4; if it is desired to improve precision and reduce false positives, the threshold can be increased to 0.6-0.7.

[0076] Establish a comprehensive performance evaluation system and calculate multiple evaluation metrics. Accuracy This reflects the proportion of correct overall predictions; accuracy rate Recall rate reflects the proportion of cases predicted as fusion welding that actually result in fusion welding; This reflects the proportion of correctly identified fusion welding cases in actual fusion welding scenarios; The result is the harmonic mean of precision and recall. TP represents a true positive (correctly predicted as weld), TN represents a true negative (correctly predicted as normal), FP represents a false positive (false alarm, actually normal but predicted as weld), and FN represents a false negative (missed detection, actually welded but predicted as normal). For contact weld prediction tasks, recall is particularly important because missed detections can lead to serious accidents. Typically, a recall rate of at least 90% is required, meaning that over 90% of weld cases can be identified in advance.

[0077] Plot a confusion matrix to display the number of true positives, true negatives, false positives, and false negatives in matrix form, visually reflecting the model's predictive performance across each category. Plot a ROC curve (Receiver Operating Characteristic curve), with the horizontal axis representing the false positive rate. The vertical axis represents the true positive rate. The area under the curve (AUC) reflects the model's overall classification ability; the closer the AUC value is to 1, the better the performance. For the contact welding prediction sub-model, an AUC of at least 0.90 is typically required. Plotting the Precision-Recall curve is particularly useful for scenarios with imbalanced classes. The horizontal axis represents recall, and the vertical axis represents precision; the closer the curve is to the upper right corner, the better the performance. Calculate the performance metrics under different decision thresholds to find the optimal balance between precision and recall.

[0078] Perform error analysis to identify cases of prediction errors and analyze the causes of these errors. For false positives (false alarms), analyze whether there are cases with similar characteristics that did not lead to welding, and check whether there are data labeling errors or whether the model is overly sensitive. For false negatives (missed alarms), analyze whether the characteristics before welding differ from typical cases, and check whether there is a lack of similar cases in the training set or whether the model is insufficient in recognizing certain abnormal patterns. Statistically analyze the prediction performance under different contact materials, voltage levels, and fault severity to identify the model's strengths and weaknesses. Perform time-dimensional analysis to evaluate how far in advance the model can predict welding risks, such as the proportion of cases that can accurately warn of welding 24 hours, 48 ​​hours, and 7 days before welding occurs. Build a test report, recording in detail various performance indicators, confusion matrix, ROC curve, error analysis results, typical case demonstrations, etc., to provide a basis for model optimization and practical application.

[0079] E2.4 When the loss of the trained contact welding risk prediction sub-model reaches the preset loss threshold, the trained contact welding risk prediction sub-model is determined as the contact welding risk prediction sub-model.

[0080] A preset loss threshold is set for the contact welding risk prediction sub-model. This preset loss threshold is determined based on the safety requirements and practical application needs of contact welding prediction. For contact welding, a serious fault, high performance standards are set, requiring a recall rate of no less than 90% (i.e., a false negative rate of no more than 10%), a precision rate of no less than 80% (i.e., a false positive rate of no more than 20%), an F1 score of no less than 0.85, and an AUC value of no less than 0.92. The loss threshold is set to a loss function value below 0.08 on the welding test set, or a loss function value that does not decrease for 30 consecutive rounds on the validation set. Simultaneously, the model is required to provide at least 24 hours' advance warning of welding risks, allowing sufficient processing time for maintenance personnel.

[0081] To determine whether the trained contact welding risk prediction sub-model meets the preset loss threshold, a comprehensive evaluation of various performance indicators on the welding test set is conducted. If all indicators reach or exceed the preset loss threshold, and the performance is balanced across sub-test sets of different contact materials and voltage levels, with no obvious weaknesses, the model is considered successfully trained, and the trained contact welding risk prediction sub-model is designated as the contact welding risk prediction sub-model. Complete model information is saved, including model structure definition, trained parameter weights, standardized parameters (mean and standard deviation), decision threshold, performance evaluation report, etc. Model metadata records are established, including training date, training dataset version, model version number, performance indicators, and applicable scope (e.g., applicable contact materials, voltage levels), to facilitate model management and traceability.

[0082] If the performance metrics do not reach the preset loss threshold, model optimization iterations are required. Optimization methods include: data-level optimization, collecting more circuit breaker case samples, especially those from scenarios where the model performs poorly, improving data annotation quality, and increasing the diversity of data augmentation; feature engineering optimization, introducing more physical mechanism-related features, such as arc energy integral, temperature spatiotemporal gradient, and current spectrum features, and performing feature selection to eliminate redundant features; model structure optimization, increasing network depth or width, introducing more advanced network modules such as Transformer and graph neural networks, and adjusting multi-channel fusion strategies; hyperparameter optimization, using grid search or Bayesian optimization methods to find the optimal hyperparameter combination; and training strategy optimization, adjusting class weight ratios, using loss functions more suitable for imbalanced data such as focal loss, extending training time, or adjusting the learning rate strategy. To address the issue of significant performance differences between contacts made of different materials, a dedicated sub-model is trained for each material, and the prediction results of multiple sub-models are fused using model ensemble methods. Steps E2.2 and E2.3 are repeated for iterative model training and testing, recording performance improvements in each iteration until the model performance reaches the preset loss threshold.

[0083] A model validation mechanism was established, evaluating the model not only on the circuit breaker test set but also performing cross-validation on independent validation datasets or real-world operational data to ensure the model's generalization ability. Time extrapolation validation was conducted, using newly collected cases to test model performance and verify its adaptability to new data. Adversarial testing was performed, constructing boundary cases and difficult samples to test the model's robustness. The final determined contact welding risk prediction sub-model possesses high accuracy, high recall, and strong generalization ability, enabling it to reliably predict contact welding risks in practical applications and providing strong support for the safe operation of switchgear.

[0084] E3. Input the core feature data into the contact welding risk prediction sub-model to predict the contact welding risk of the switchgear.

[0085] In practical applications, the core feature data collected in real time in step E1 is input into the pre-trained contact welding risk prediction sub-model for online prediction. Preprocessing is performed before data input; the core feature data is normalized using the same standardized parameters as in the training phase to ensure data scale consistency. For time-series data, a time window matching the model input requirements is constructed, typically a historical data sequence of the most recent 10-100 time steps. For infrared thermal imaging data, image preprocessing is performed, including size adjustment, background temperature compensation, and emissivity correction, to ensure consistency with the image format used during model training. For scalar data, numerical standardization is directly applied. The preprocessed core feature data is then organized according to the multi-channel input format defined by the model to form a standard input tensor.

[0086] After receiving core feature data, the contact welding risk prediction sub-model performs forward inference calculations. The model's time-series data channel processes dynamic resistance values, transient currents, triaxial acceleration, and other data, extracting time-dependent features and anomalous patterns through convolutional or LSTM layers. The image data channel processes infrared thermal imaging data, extracting spatial temperature distribution features and identifying local hotspot areas through a two-dimensional convolutional network. The scalar data channel processes parameters such as opening speed and metal vapor concentration, performing feature transformation through fully connected layers. The features extracted from each channel are integrated in a fusion layer to form a comprehensive feature representation.

[0087] In summary, this invention obtains key performance indicators by collecting and preprocessing real-time operating data of the switchgear. Combined with a multi-scale prediction model trained using historical switchgear fault data, it can accurately predict the switchgear's state, achieving early prediction of switchgear performance and fault warning. Compared to existing technologies, this invention's multi-scale prediction model comprehensively evaluates the performance status of multiple subsystems of the switchgear, overcoming the shortcomings of traditional methods such as single prediction models and incomplete evaluations. This improves the accuracy and reliability of predictions, effectively reducing the probability of switchgear failures and minimizing economic losses and safety incidents.

[0088] This invention also constructs a sub-model for predicting contact welding risks, performing specific risk predictions for key components of the switchgear. By injecting short-circuit current based on fault mode and effects analysis, it simulates contact welding risks under extreme conditions, predicting arc energy and contact erosion levels. This virtual simulation testing method can cover extreme conditions that are difficult to achieve in actual testing, not only reducing testing costs but also improving the accuracy of fault risk prediction under extreme conditions. Simultaneously, this invention combines historical load data and equipment aging models to predict the remaining lifespan of the switchgear, providing a scientific basis for equipment maintenance and replacement, realizing intelligent management throughout the equipment's lifecycle, and improving the operational reliability of the switchgear and the overall safety of the power supply system.

[0089] Example 3 illustrates a schematic scheme for a switchgear performance prediction method. It should be noted that the technical solution of this switchgear performance prediction system belongs to the same concept as the technical solution of the switchgear performance prediction method described above. Details not described in detail in this embodiment can be found in the description of the technical solution of the switchgear performance prediction method described above.

[0090] This embodiment also provides a switchgear performance prediction system, including: The data acquisition module is used to collect real-time operating data of the switchgear; The data preprocessing module is used to preprocess the real-time running data to obtain key performance indicators; The data acquisition module is used to acquire historical switchgear fault data; The model training module is used to train a multi-scale prediction model using the historical switchgear fault data. The prediction module is used to combine the multi-scale prediction model and the key performance indicators to predict the predicted state of the switchgear. The decision execution module is used to take corresponding measures based on the predicted state.

[0091] This embodiment also provides an electronic device suitable for switchgear performance prediction, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the switchgear performance prediction method proposed in the above embodiment.

[0092] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the switch cabinet performance prediction method proposed in the above embodiments.

[0093] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for predicting switch cabinet performance proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0094] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0095] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the performance of a switchgear, characterized in that, include: Collect real-time operating data of the switchgear, preprocess the real-time operating data, and obtain key performance indicators; Acquire historical switchgear fault data, and train a multi-scale prediction model using the historical switchgear fault data; The multi-scale prediction model and the key performance indicators are combined to predict the predicted state of the switchgear. Take appropriate measures based on the predicted state.

2. The switchgear performance prediction method as described in claim 1, characterized in that, The steps for preprocessing the real-time running data include: The real-time running data is subjected to noise filtering to obtain preliminary processed data; The preliminary processed data is aligned to obtain secondary processed data; Feature extraction is performed on the secondary processed data to obtain the key performance indicators.

3. The switchgear performance prediction method as described in claim 2, characterized in that, The steps for training a multi-scale prediction model using the historical switchgear fault data include: The historical switchgear fault data is divided into a training set and a test set; Construct an initial prediction model, and train the initial prediction model based on the training set to obtain a trained prediction model; The trained prediction model is tested based on the test set; When the loss of the trained prediction model reaches a preset threshold, the trained prediction model is determined as the multi-scale prediction model.

4. The switchgear performance prediction method as described in claim 3, characterized in that, The steps for predicting the predicted state of switchgear include: Input the key performance indicators into the multi-scale prediction model; The predicted state of the switchgear is calculated using the multi-scale prediction model. The predicted status of the switchgear includes the operating status of the switchgear and the predicted fault information of the components in the future time period. The fault information includes the faulty component and the cause of the fault.

5. The switchgear performance prediction method as described in claim 4, characterized in that, The steps for taking appropriate measures based on the predicted state include: Determine whether the predicted state of the switchgear indicates that a component will fail within a future time period; When the judgment result is yes, an early warning is issued based on the cause of the fault, and the corresponding component is replaced based on the cause of the fault.

6. The switchgear performance prediction method as described in claim 5, characterized in that, The multi-scale prediction model includes a sub-model for predicting contact welding risk. The application steps of the contact welding risk prediction sub-model include: Collect core feature data; Obtain historical cases of circuit failure, and train a sub-model for predicting the risk of contact welding using these historical cases; The core feature data is input into the contact welding risk prediction sub-model to predict the contact welding risk of the switchgear.

7. The switchgear performance prediction method as described in claim 6, characterized in that, The steps for obtaining the contact welding risk prediction sub-model through training the historical melting failure cases include: The historical circuit breaker cases are divided into a circuit breaker training set and a circuit breaker test set. Construct an initial contact welding risk prediction sub-model, and train the initial contact welding risk prediction sub-model based on the welding training set to obtain the trained contact welding risk prediction sub-model; The trained contact welding risk prediction sub-model was tested based on the aforementioned fracturing test set. When the loss of the trained contact welding risk prediction sub-model reaches the preset loss threshold, the trained contact welding risk prediction sub-model is determined as the contact welding risk prediction sub-model.

8. A switchgear performance prediction system, using the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect real-time operating data of the switchgear; The data preprocessing module is used to preprocess the real-time running data to obtain key performance indicators; The data acquisition module is used to acquire historical switchgear fault data; The model training module is used to train a multi-scale prediction model using the historical switchgear fault data. The prediction module is used to combine the multi-scale prediction model and the key performance indicators to predict the predicted state of the switchgear. The decision execution module is used to take corresponding measures based on the predicted state.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the switch cabinet performance prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the switchgear performance prediction method according to any one of claims 1 to 7.

Images