Vacuum circuit breaker service life prediction method

Abstract

The invention relates to the technical field of state monitoring and life evaluation of power equipment, in particular to a life prediction method of a vacuum circuit breaker. The objective of the invention is to solve the problem of key degradation signal loss caused by poor dynamic weight adaptability, modeling failure in a few-sample scene and multi-source feature fusion coarseness of a vacuum circuit breaker life prediction technology in the prior art. Through multi-source fine-grained feature extraction, working condition sensing dynamic weight fusion and a few-sample element learning strategy, the problems that a traditional method is poor in dynamic weight adaptability, few-sample modeling fails and feature fusion is coarse in granularity are solved, and the method has the advantages that prediction precision is improved, cross-model and cross-working-condition application cost is reduced, and key degradation signal loss is reduced.

Description

Technical Field

[0001] This application relates to the field of power equipment condition monitoring and life assessment technology, and more specifically, to a method for predicting the life of a vacuum circuit breaker. Background Technology

[0002] Vacuum circuit breakers, as core switching equipment in power systems, bear the dual functions of load dispatching and fault interruption. The degradation of their mechanical and insulation properties directly affects power grid safety. As power systems develop towards intelligence and high reliability, the traditional "periodic replacement" operation and maintenance model can no longer meet actual needs due to serious resource waste and delayed fault warnings. Accurate life prediction technology is urgently needed.

[0003] Existing vacuum circuit breaker life prediction technologies have three major drawbacks: First, dynamic weights rely on preset rules, resulting in poor adaptability. Existing technologies rely entirely on preset event rules for weight adjustment, which cannot handle complex working conditions without preset rules, leading to inaccurate weights. Secondly, modeling fails in scenarios with few samples, and costs are high across operating conditions and models. Existing technologies all rely on a large number of historical samples to train models or preset rules, and the models fail due to insufficient samples when new circuit breaker models or new operating conditions are introduced. Third, the coarse-grained fusion of multi-source features leads to the loss of key degradation signals. Some techniques use only a single vibration signal, or although they collect multi-source data, they do not make full use of fine-grained features, resulting in the loss of coupled degradation signals.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] (a) Technical problems to be solved To address the aforementioned issues, this invention proposes a method for predicting the lifespan of vacuum circuit breakers. This method aims to solve the problems faced by existing vacuum circuit breaker lifespan prediction technologies, such as poor dynamic weight adaptability, modeling failure in scenarios with few samples, and loss of key degradation signals due to coarse-grained fusion of multi-source features.

[0006] (II) Technical Solution The present invention provides a method for predicting the lifespan of a vacuum circuit breaker, the technical solution of which is as follows: acquiring multi-source monitoring data of the vacuum circuit breaker, the multi-source monitoring data including at least current data, mechanical waveform data, partial discharge data, environmental data, and contact temperature data; preprocessing the multi-source monitoring data to extract multi-source fine-grained features, the fine-grained features including at least the preset frequency band energy spectrum features extracted from the current data; inputting the multi-source fine-grained features and real-time operating parameters into an operating condition sensing weight network, calculating the dynamic weights corresponding to each feature, and performing weighted fusion of the multi-source fine-grained features based on the dynamic weights. The process involves: merging features; constructing and training a life prediction model using a few-shot meta-learning strategy, which includes meta-pre-training based on a meta-task set containing various circuit breaker models and operating conditions to obtain general initialization parameters; calibrating the model using no more than thirty samples for the target circuit breaker to obtain the target prediction model; inputting the fused features into the target prediction model to calculate the comprehensive health index and the aging coefficient after operating condition correction for the vacuum circuit breaker; calculating the remaining service life of the vacuum circuit breaker based on the comprehensive health index and the aging coefficient; and evaluating the confidence level of the calculation results.

[0007] Furthermore, this application proposes that the operating condition sensing weight network is a feedforward neural network, the input of the operating condition sensing weight network is the real-time ambient temperature and the main circuit current fluctuation coefficient, and the output layer adopts the Softmax activation function to output normalized dynamic weights.

[0008] Furthermore, this application proposes that the specific steps of meta-pre-training include: constructing a meta-task set containing at least five circuit breaker models and at least three typical operating conditions; in each meta-task, training an improved CNN-BiGRU backbone network using fused features and updating temporary parameters; calculating the average loss of all meta-tasks and updating the general initialization parameters based on this.

[0009] Furthermore, this application proposes an improved CNN-BiGRU backbone network in which the CNN part employs depthwise separable convolutions and the BiGRU part adds a layer normalization layer in the hidden layers.

[0010] Furthermore, this application proposes that the comprehensive health index is calculated as follows: based on dynamic weights, a weighted average is taken of the mechanical aging health index, the contact wear health index, and the insulation aging health index to obtain the comprehensive health index, the calculation formula of which is:

[0011] in, To assess overall health, , , These are the dynamic weights corresponding to the health index. , , These are the mechanical aging health index, the contact wear health index, and the insulation aging health index.

[0012] Furthermore, this application also proposes the following formula for calculating the aging factor after operating condition correction:

[0013] in, The aging factor is... This is the operating condition correction factor determined based on the main circuit current fluctuation coefficient. Based on the aging rate factor, Environmental sensitivity coefficient, This refers to the environmental corrosion health index.

[0014] Furthermore, this application also proposes that after calculating the remaining service life of the vacuum circuit breaker, a step of assessing the confidence level of the calculation results is included: using the Monte Carlo Dropout method to perform multiple forward inferences on the target prediction model to obtain multiple remaining service life prediction values; calculating the standard deviation of the multiple prediction values, and if the standard deviation is lower than a preset threshold of 5%, the confidence level is determined to be up to standard and the final service life prediction result is output.

[0015] Furthermore, this application also proposes preset frequency band energy spectrum features extracted from current data, including 50-200Hz frequency band energy values ​​to reflect mechanical vibration and 1-2kHz frequency band energy values ​​to reflect mechanical jamming.

[0016] Furthermore, this application also proposes a computing device comprising: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the above-described vacuum circuit breaker life prediction method.

[0017] Furthermore, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described vacuum circuit breaker life prediction method.

[0018] (III) Beneficial Effects Compared with the prior art, the beneficial effects of the present invention are as follows: This invention solves the problems of poor dynamic weight adaptability, failure of modeling with few samples, and coarse-grained feature fusion in traditional methods by using multi-source fine-grained feature extraction, dynamic weight fusion based on working condition perception, and a few-sample meta-learning strategy. It has the advantages of improving prediction accuracy, reducing the cost of cross-model and cross-working condition applications, and reducing the loss of key degradation signals.

[0019] This invention can dynamically adapt to the weight allocation requirements under unknown operating conditions, improving the sensitivity of early defect detection. Fine-grained feature extraction effectively captures coupling degradation signals in multi-source data, improving prediction accuracy. The meta-learning strategy reduces the sample requirements for cross-scenario modeling, reducing operation and maintenance costs. The confidence assessment mechanism ensures the reliability of prediction results, providing a basis for preventive maintenance decisions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0021] Figure 1 A logic diagram of a method for predicting the lifespan of vacuum circuit breakers. Detailed Implementation

[0022] The following will refer to the appendix to this application. Figure 1 The technical solutions in this application are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] In existing technologies, vacuum circuit breakers, as core switching equipment in power systems, face challenges in lifespan prediction, including poor dynamic weight adaptation, modeling failure in scenarios with few samples, and loss of key degradation signals due to coarse-grained multi-source feature fusion. Traditional methods rely on preset rules to adjust weights, which cannot handle complex operating conditions without pre-set rules. Furthermore, modeling across different models or new operating conditions requires a large number of samples, leading to high maintenance costs. Existing feature extraction methods are coarse-grained, making it difficult to capture early degradation signals and affecting prediction accuracy.

[0024] To address the aforementioned issues, the inventors discovered that existing weight adjustment mechanisms lack dynamic adaptability and cannot handle feature correlations under unknown operating conditions. By analyzing the coupling relationships between multi-source data, they proposed using neural networks to achieve dynamic weight allocation, replacing manually preset rules. To address the problem of scarce samples, a meta-learning strategy was introduced, enabling the model to quickly adapt to new scenarios with limited samples. Further investigation revealed that coarse-grained features lead to signal loss, necessitating the extraction of specific frequency band energy spectra from current data to capture degradation modes such as mechanical vibration and jamming.

[0025] Therefore, this application proposes a method for predicting the lifespan of a vacuum circuit breaker, comprising the following steps: S100: Acquire multi-source monitoring data of the vacuum circuit breaker; S200. Preprocess the data to extract multi-source fine-grained features; S300: Input the features and real-time operating condition parameters into the operating condition perception weight network to calculate dynamic weights and perform weighted fusion. S400: Construct a lifetime prediction model using a few-sample meta-learning strategy and calibrate it. S500 calculates the overall health index and the aging factor after operating condition correction, and assesses the remaining service life and confidence level.

[0026] The multi-source monitoring data includes current data, mechanical waveform data, partial discharge data, environmental data, and contact temperature data, covering key degradation factors such as mechanical vibration and contact wear through multi-dimensional data acquisition. Fine-grained feature extraction is achieved through preset frequency band energy spectrum analysis, such as using Fast Fourier Transform to extract energy values ​​in the 50-200Hz frequency band to capture abnormal mechanical vibrations. The operating condition perception weight network adopts a feedforward neural network structure, taking real-time ambient temperature and main circuit current fluctuation coefficient as input, and outputting normalized weights through the Softmax function to achieve dynamic adjustment of feature contribution. The few-sample meta-learning strategy is pre-trained by constructing meta-task sets of multiple models and operating conditions to obtain universal initialization parameters, enabling the model to have cross-scenario generalization ability. The comprehensive health index calculation adopts a dynamic weighted average method, combining mechanical aging, contact wear, and insulation aging health indices to quantify the overall health status of the equipment.

[0027] Specifically, this method first collects multi-source monitoring data through devices such as vibration sensors and current transformers. After filtering and standardization, fine-grained features reflecting different degradation modes are extracted. These features, along with real-time operating parameters, are input into a trained operating condition sensing weight network to generate dynamic weights that match ambient temperature and current fluctuations, thus completing feature fusion.

[0028] Based on a pre-trained meta-learning model, a prediction model adapted to specific scenarios is constructed by calibrating parameters using a small number of samples from target circuit breakers. After fusing features into the model, a comprehensive health index and an aging coefficient corrected for operating conditions are calculated, and the remaining lifespan is predicted by combining this with historical degradation curves. Monte Carlo Dropout is used for multiple inferences, and the confidence level of the results is evaluated through the standard deviation of the predicted values ​​to ensure output reliability.

[0029] Traditional methods rely on manually preset rules to adjust weights, which cannot handle undefined complex operating conditions. This solution, however, uses neural networks to dynamically generate weights, improving adaptability to different operating conditions. Existing feature extraction methods use only a single data source or coarse-grained features; this solution captures early degradation signals through frequency band energy spectrum analysis. Traditional cross-scene modeling requires a large number of samples; this solution employs a meta-learning strategy to significantly reduce data requirements and solve the problem of scarce samples for new equipment models.

[0030] This application can dynamically adapt to the weight allocation requirements under unknown operating conditions, improving the sensitivity of early defect detection. Fine-grained feature extraction effectively captures coupling degradation signals in multi-source data, improving prediction accuracy. The meta-learning strategy reduces the sample requirements for cross-scenario modeling, reducing operation and maintenance costs. The confidence assessment mechanism ensures the reliability of prediction results, providing a basis for preventive maintenance decisions.

[0031] This application further proposes a technical solution for constructing a working condition sensing weight network using a feedforward neural network. The input parameters of this network include real-time ambient temperature and main circuit current fluctuation coefficient, and the output layer generates normalized dynamic weights through the Softmax activation function.

[0032] Among them, the feedforward neural network refers to a multilayer perceptron with unidirectional propagation characteristics. Specifically, it can be implemented using a hierarchical structure of input layer, hidden layer, and output layer. The number of neurons in the hidden layer can be configured according to the complexity of the actual operating conditions. The network parameters are optimized through a backpropagation algorithm to achieve a nonlinear mapping from input to output. Real-time ambient temperature refers to the temperature measurement of the environment in which the circuit breaker is located. Specifically, it can be achieved by collecting and filtering data from a temperature sensor, used to quantify the impact of environmental factors on the aging of mechanical components and insulation materials. The main circuit current fluctuation coefficient is a quantitative indicator of the degree of fluctuation in the main circuit current of the circuit breaker. Specifically, it can be calculated as the ratio of the standard deviation to the mean of the current's effective value, used to reflect the accelerating effect of grid load fluctuations on contact erosion. The Softmax activation function is a multi-classification function that transforms the network output into a probability distribution. Specifically, it can be implemented by combining exponential operations and normalization processing to ensure that the sum of the weights of each feature is constant at 1, avoiding weight allocation conflicts under different operating conditions.

[0033] Specifically, the feedforward neural network is trained using historical multi-condition data to learn the mapping relationship between ambient temperature and current fluctuation coefficient on feature weight allocation. During training, network parameters are optimized to capture the impact of temperature changes on the thermal expansion of mechanical components and the accelerating effect of current harmonic anomalies on contact wear. When the real-time ambient temperature rises, the network automatically increases the weights of environment-related features to reflect the accelerated effect of high temperature on insulation aging; when the main circuit current fluctuation coefficient increases, the network increases the weights of current-related features to address the risk of contact ablation caused by harmonic anomalies. The Softmax function of the output layer converts the calculation results of the hidden layer into normalized weights, making the weight allocation smoothly transition with continuous changes in operating parameters, avoiding the abrupt weight changes in traditional rule-based methods under critical operating conditions.

[0034] Traditional solutions rely on manually preset event rules to adjust weights, such as adjusting weights based on the number of closing events or short-circuit tripping events. These rules cannot cover complex operating conditions that are not pre-defined, such as the simultaneous presence of high temperatures and current fluctuations. In contrast, the feedforward neural network in this solution, through a self-learning mechanism, can handle arbitrary combinations of operating parameter inputs without the need for pre-defined event rules. For example, in intermittent high-temperature scenarios, the network can dynamically adjust weights based on temperature fluctuation trends, without relying on fixed threshold trigger rules.

[0035] This application achieves adaptive adjustment of dynamic weight allocation, resolving the poor adaptability issue caused by preset rules. Under complex operating conditions, the accuracy of weight allocation is significantly improved, the early defect missed rate is reduced, and the workload of manually maintaining weight rules is decreased. This solution can effectively cope with unpredictable scenarios such as power grid load fluctuations and changes in ambient temperature and humidity, improving the reliability of vacuum circuit breaker life prediction.

[0036] This application further proposes specific steps for meta-pre-training, including constructing a meta-task set containing at least five circuit breaker models and at least three typical operating conditions; in each meta-task, training an improved CNN-BiGRU backbone network using fused features and updating temporary parameters; calculating the average loss of all meta-tasks and updating the general initialization parameters based on this.

[0037] The meta-task set refers to a collection of tasks composed of different circuit breaker models and operating conditions. Specifically, it can be constructed using circuit breaker models with different structural types such as spring-operated, permanent magnet-operated, and hydraulically operated circuit breakers, as well as typical operating conditions such as urban substations, industrial parks, and remote power stations. By covering diverse equipment types and operating environments, the model can extract common degradation patterns across equipment. The improved CNN-BiGRU backbone network refers to a network structure combining depthwise separable convolutions and bidirectional gated recurrent units. Specifically, depthwise separable convolutions can be used to reduce the number of parameters, and layer normalization operations can be added to the BiGRU hidden layers. By reducing computational complexity and stabilizing parameter updates during multi-task training, the model's ability to process multi-source time-series data is improved. The average loss refers to the comprehensive error index after training all meta-tasks. Specifically, it can be calculated by weighted averaging the root mean square errors of each meta-task. By balancing the impact of different tasks on parameter updates, it ensures that the general initialization parameters retain common knowledge while being compatible with scenario-specific characteristics.

[0038] Specifically, when constructing the meta-task set, it is necessary to select circuit breaker models with different structural types and differentiated operating conditions. For example, the ZN63 type spring operating mechanism and the VS1 type permanent magnet operating mechanism should be included in the model range, while also covering high load fluctuations and harsh temperature and humidity environments. The sample size in each meta-task should simulate a low-sample scenario, such as including 40-60 normal samples and 8-12 fault samples. In the improved CNN-BiGRU backbone network, depthwise separable convolutional layers extract the spatial distribution features of the current energy spectrum, and layer-normalized BiGRU layers capture the temporal degradation trend caused by the increase in the number of operations. During training, each meta-task independently updates temporary parameters to learn the specific task rules, and finally updates the general parameters through average loss calculation, enabling the model to identify common features across models, such as the correlation between the frequency band energy decrease of the tripping current and mechanical aging.

[0039] Traditional solutions rely on data from a single circuit breaker model or operating condition to train the model, leading to the need to collect a large number of samples when applying it across different scenarios. For example, CN120654571A only builds a model for a specific circuit breaker model and cannot adapt to new models or operating conditions. In contrast, this solution covers multiple models and operating conditions through a meta-task set, enabling the model to grasp common patterns during the pre-training stage and significantly reducing the sample requirements during subsequent calibration. Furthermore, existing technologies use ordinary convolutional neural networks to process data, which struggles to address the efficiency issues of multi-task training. This solution, through a collaborative design of depthwise separable convolution and layer normalization, improves training stability and computational efficiency.

[0040] This application addresses the issues of model failure in scenarios with few samples and the high cost of cross-model modeling. During the pre-training phase, the model learns common degradation patterns across different circuit breaker models and operating conditions, enabling subsequent calibration for new models or operating conditions to be completed with only a small number of samples, significantly reducing data acquisition costs. Simultaneously, the improved network structure enhances the ability to extract multi-source temporal features, avoids the loss of key degradation signals, and improves the accuracy of lifetime prediction.

[0041] This application further proposes an improved CNN-BiGRU backbone network, in which the CNN part adopts depthwise separable convolutions, and the BiGRU part adds a layer normalization in the hidden layers.

[0042] Depthwise separable convolution refers to decomposing standard convolution into two independent steps: spatial convolution and channel-wise convolution. Specifically, this can be achieved by first performing a 3×3 convolution kernel on the input feature map in the spatial dimension, and then performing a 1×1 convolution on each channel individually. By separating the feature extraction processes in the spatial and channel dimensions, the computational cost of parameters is reduced while preserving the spatial distribution details of fine-grained features. Layer normalization refers to adding a normalization module after the output of the BiGRU hidden layer. Specifically, this can be achieved by calculating the mean and variance normalization for each sample output from the hidden layer, and then restoring the feature representation through learnable scaling and translation parameters. This eliminates the impact of differences in data distribution across different tasks on temporal modeling, improving the network's stability in capturing fine-grained temporal features.

[0043] Specifically, when using depthwise separable convolution in the CNN part, the spatial convolution stage only filters the spatial dimension of the input feature map, preserving the spatial distribution features of high-frequency fine-grained signals in multi-source monitoring data, such as the peak shape of the energy spectrum in the 50-200Hz band of the tripping current signal. The channel-by-channel convolution stage performs convolution on each feature channel separately, enhancing the specific expression of different monitoring data channels, such as the independent feature extraction of the energy spectrum in the 1-2kHz band of the current channel and the energy spectrum in the 500Hz band of the vibration channel. After adding layer normalization in the BiGRU part, the hidden layer output is standardized to eliminate parameter fluctuations caused by differences in data distribution under different operating conditions. For example, the impact of the difference in current fluctuation coefficient between urban substations and industrial parks on time series modeling is effectively suppressed, so that the gradual trend of the tripping current energy spectrum changing with the number of operations can be stably captured.

[0044] Existing methods for predicting the lifespan of vacuum circuit breakers often employ standard convolution operations in their CNN components. This leads to the aliasing of fine-grained features from different channels during convolution, such as the averaging effect of spatial information from current and vibration features under standard convolution kernels. Furthermore, existing BiGRU structures lack normalization processing, resulting in insufficient stability of temporal modeling in cross-condition meta-learning scenarios. This proposed solution achieves decoupled feature extraction of spatial and channel dimensions through depthwise separable convolution, combined with layer normalization to ensure the stability of temporal features under multi-task data, forming a dual improvement mechanism.

[0045] This application can effectively improve the extraction accuracy of multi-source fine-grained features, avoid the loss of key degradation signals in the feature fusion process, and enhance the training efficiency and cross-condition adaptability of neural networks in low-sample scenarios, providing a high-precision feature expression basis for vacuum circuit breaker life prediction.

[0046] This application further proposes that the comprehensive health index is calculated by weighting the mechanical aging health index, contact wear health index, and insulation aging health index based on dynamic weights, and the comprehensive health index is obtained by the following formula:

[0047] in, To assess overall health, , , These are the dynamic weights corresponding to the health index. , , These are the mechanical aging health index, the contact wear health index, and the insulation aging health index.

[0048] Among them, dynamic weight refers to the weight coefficients obtained in real time through the working condition sensing weight network. Specifically, it can be implemented by using a feedforward neural network combined with real-time ambient temperature and main circuit current fluctuation coefficient as input, which is used to adjust the contribution of different health indices according to the current working conditions.

[0049] Weighted average refers to multiplying each health index by its corresponding weight, summing the results, and then dividing by the total weight. Specifically, it can be achieved by matrix operations combined with normalization processing, and is used to realize the dynamic amplification and fusion of degradation signals.

[0050] The comprehensive health index is a quantitative indicator reflecting the overall health status of a vacuum circuit breaker. Specifically, it can be achieved by integrating three types of health indices: mechanical, contact, and insulation, to cover the main degradation modes of the equipment. The normalization process involves performing a division operation on the weighted sum of the results, specifically by calculating the sum of weights in the denominator, to eliminate the influence of differences in the absolute values ​​of the weights on the index range.

[0051] Specifically, dynamic weights are generated in real time through a condition-sensing weight network, autonomously adjusting the weight ratios of three health indices—mechanical aging, contact wear, and insulation aging—based on operating parameters such as ambient temperature and current fluctuations. During the weighted average calculation, the numerator multiplies each health index by its weight, thus amplifying key degradation signals under specific operating conditions. For example, when main circuit current fluctuations intensify, the weight of the contact wear health index automatically increases, significantly increasing the proportion of its corresponding degradation signal in the overall health index. The denominator is normalized through the sum of weights, ensuring the overall health index remains within a standard value range, facilitating horizontal comparisons of health status under different equipment or operating conditions. This calculation method comprehensively covers the three core degradation modes: mechanical performance degradation, contact material wear, and insulation performance deterioration, avoiding signal loss issues caused by single-feature fusion.

[0052] Traditional methods use fixed weights or adjust weights based on preset rules, failing to dynamically correlate real-time operating conditions with degradation mechanisms. For example, a certain existing technology sets a constant mechanical aging weight of 0.6. When equipment is in a high-temperature and high-humidity environment, insulation aging accelerates, but the weight does not increase accordingly, leading to an underestimation of the impact of insulation degradation in the overall health index. This solution, through a combination of dynamic weights and a weighted average formula, automatically increases the weight ratio of the insulation aging health index under high-temperature conditions, enabling the overall health index to accurately reflect the insulation degradation trend caused by environmental factors.

[0053] This application can dynamically adjust the fusion ratio of different health indices according to real-time operating conditions, so that the comprehensive health index accurately reflects the coupling effect of the three types of degradation signals: mechanical, contact and insulation. It solves the problem of health status assessment bias caused by the fixed weights in the existing technology, and at the same time, it ensures the comparability and interpretability of the assessment results through normalization processing.

[0054] This application further proposes the following formula for calculating the aging factor after operating condition correction:

[0055] in, The aging factor is... This is the operating condition correction factor determined based on the main circuit current fluctuation coefficient. Based on the aging rate factor, Environmental sensitivity coefficient, This refers to the environmental corrosion health index.

[0056] The main circuit current fluctuation coefficient refers to the degree of fluctuation of the main circuit current during the operation of the circuit breaker. Specifically, it can be quantified by the ratio of the current harmonic content to the fundamental current, which is used to reflect the dynamic impact of current conditions on equipment aging.

[0057] The baseline aging rate factor refers to the reference aging rate of the circuit breaker under standard operating conditions. It can be obtained by fitting historical data from accelerated aging tests in the laboratory and serves as the basic parameter for calculating the aging coefficient. The environmental sensitivity factor refers to the sensitivity of equipment materials to environmental corrosion factors. It can be differentiated according to the environmental type of the equipment installation area to quantify the differences in the contribution of different environments to the aging rate.

[0058] The environmental corrosion health index is a quantitative value of the health status of equipment affected by environmental factors. It can be calculated using environmental monitoring data and corrosion accumulation models, and is used to characterize the degree of progressive damage to equipment materials caused by environmental corrosion.

[0059] Specifically, this technical solution achieves dynamic correction of the aging coefficient by constructing a mathematical formula that couples multiple factors. The main circuit current fluctuation coefficient is calculated using real-time monitoring data, and the operating condition correction coefficient is automatically increased when current fluctuations intensify. The numerical value quantifies the accelerated aging effect of current conditions on mechanical components and electrical contacts. Environmental corrosion health index. The index is dynamically updated based on environmental monitoring data; it decreases as environmental corrosion worsens. This term reflects the degree to which environmental factors degrade material performance. The environmental sensitivity coefficient γ is preset according to the type of environment in which the equipment is located. For example, a higher coefficient value is used in coastal high-salt-fog environments, and a lower coefficient value is used in inland dry environments, achieving differentiated quantification of environmental impact. In the formula... and The product relationship design allows the coupling effect of current conditions and environmental corrosion to be nonlinearly amplified, accurately depicting the accelerated aging phenomenon under the synergistic effect of the two factors.

[0060] Traditional methods often employ fixed aging coefficients or single-factor linear correction models, failing to reflect the dynamic coupling effect between current fluctuations and environmental corrosion. For example, one existing technology adjusts the aging coefficient solely through temperature parameters, neglecting the accelerating effect of current fluctuations on material aging; another existing technology uses a simple superposition of environmental corrosion factors and the basic aging rate, failing to characterize the nonlinear interactions of multiple factors. This proposed solution utilizes a product-based composite correction term to model the synergistic effect of current conditions and environmental corrosion, overcoming the quantization bias of linear models in complex scenarios.

[0061] This application effectively addresses the technical deficiency of aging coefficients that do not consider the dynamic coupling effect of real-time operating conditions and environmental corrosion. The accelerated mechanical wear effect caused by current fluctuations and the material performance degradation effect caused by environmental corrosion are quantified simultaneously, and the actual aging rate under combined factors is accurately reflected through nonlinear coupling. This scheme significantly improves the calculation accuracy of aging coefficients under complex operating conditions, making the life prediction results more consistent with the actual operating state of equipment, and providing reliable technical support for the safe operation and maintenance of power grids.

[0062] This application further proposes a step of evaluating the confidence level of the calculation results after calculating the remaining service life of the vacuum circuit breaker, including using the Monte Carlo Dropout method to perform multiple forward inferences on the target prediction model to obtain multiple remaining service life prediction values, calculating the standard deviation of multiple prediction values ​​and comparing them with a preset threshold to determine whether the confidence level meets the standard.

[0063] The Monte Carlo Dropout method maintains a random deactivation mechanism during the neural network inference phase, capturing the uncertainty of model parameters through repeated inference. Specifically, this can be achieved by randomly masking some neuron connections during the forward propagation of the target prediction model. This feature effectively quantifies the model prediction volatility caused by limited sample calibration. Standard deviation calculation analyzes the dispersion of remaining lifetime predictions obtained from multiple inferences. Specifically, it uses the sample standard deviation formula in statistics to calculate the fluctuation range of the predicted value set. This feature objectively reflects the stability of the model prediction results. The preset threshold refers to a pre-set upper limit for the standard deviation, which can be based on a critical value validated by historical data. This feature transforms the confidence assessment of the prediction results into an actionable operational decision-making standard.

[0064] Specifically, after obtaining the initial predicted lifespan from the target prediction model, the Dropout layer in the model is kept active, and a set of predicted values ​​is generated by repeatedly performing forward inference a preset number of times. During each inference process, some neuron connections are randomly masked, fully exposing the uncertainty of the model parameters. Subsequently, a dispersion analysis is performed on the set of predicted values. When the standard deviation is lower than a preset threshold, it indicates that the model's prediction results are consistent under different parameter configurations, and the final lifespan prediction value is output at this time. If the standard deviation exceeds the threshold, an early warning mechanism is triggered, prompting the need to supplement sample data or re-validate the model parameters.

[0065] Traditional methods for predicting the lifespan of vacuum circuit breakers rely solely on single inference results and lack a confidence assessment mechanism, failing to identify prediction biases caused by insufficient samples or sudden changes in operating conditions. This solution, however, constructs a reliable quantitative system for prediction results through a probabilistic inference mechanism. This enables dynamic monitoring of prediction stability without requiring additional computational resources, while deeply integrating confidence assessment with the operation and maintenance decision-making process, forming a closed-loop feedback mechanism.

[0066] This application can effectively identify unreliable prediction results caused by insufficient convergence of model parameters or abnormal operating conditions, avoiding misjudgments by maintenance personnel based on highly volatile prediction values. When the prediction standard deviation exceeds a threshold, the system automatically triggers a supplementary sample collection process to ensure that the model gradually improves prediction accuracy during iterative optimization, ultimately achieving verifiability and operability of the remaining service life assessment results for vacuum circuit breakers.

[0067] This application further proposes to extract preset frequency band energy spectrum features from current data, including 50-200Hz frequency band energy values ​​to reflect mechanical vibration and 1-2kHz frequency band energy values ​​to reflect mechanical jamming.

[0068] The energy value in the 50-200Hz frequency band refers to the energy distribution characteristics of the current signal within this band. Specifically, this can be achieved by performing spectral analysis of the opening and closing coil current using wavelet transform or short-time Fourier transform to calculate the integral energy value of this band. This band corresponds to the mechanical vibration frequency during the normal operation of the circuit breaker's operating mechanism, and its energy changes can characterize mechanical wear conditions such as spring fatigue and increased linkage clearance. The energy value in the 1-2kHz frequency band refers to the energy distribution characteristics of the current signal within this band. Specifically, this can be achieved by calculating the peak energy value of this band using bandpass filtering combined with power spectral density. This band corresponds to the high-frequency vibration signal generated by abnormal friction in the contact system, and its sudden energy increase can characterize mechanical jamming faults such as contact sticking and guide sleeve deformation.

[0069] Specifically, during the opening and closing operation, the current signal of the opening coil is coupled with the vibration characteristics of the mechanical system. By collecting current waveform data, noise reduction is first performed, followed by spectrum analysis to extract the energy values ​​of two preset frequency bands: 50-200Hz and 1-2kHz. A decrease in energy in the 50-200Hz band can reflect insufficient spring energy storage or abnormal linkage transmission; for example, when the spring is fatigued, the energy value in this band decreases compared to the normal state. A sudden increase in energy in the 1-2kHz band can reflect abnormal contact friction; for example, when the contact guide sleeve is deformed, the energy value in this band increases significantly compared to the normal state. The synergistic analysis of the characteristics of the two frequency bands can distinguish between simple mechanical vibration abnormalities and mechanical jamming faults. For example, when the energy decrease in the 50-200Hz band is accompanied by an increase in the energy in the 1-2kHz band, it can be determined as a coupled fault of mechanical vibration and jamming.

[0070] Existing methods typically use the average energy across the entire frequency band or the peak value in the time domain as current characteristics. For example, patent CN120654571A only calculates the average energy across the 20-2000Hz frequency band, resulting in the masking of subtle changes in the 50-200Hz and 1-2kHz frequency bands. This proposed solution, however, avoids the loss of key signals caused by wideband analysis by selectively extracting energy values ​​from two narrow frequency bands, and is able to capture early mechanical degradation characteristics.

[0071] This application solves the problem of mechanical vibration and jamming signal loss caused by coarse-grained current feature extraction, and realizes accurate monitoring of mechanical degradation states such as spring fatigue and contact jamming. At the same time, feature extraction can be completed through current signal without the need to deploy additional vibration sensors.

[0072] This application further proposes a computing device, including at least one processor and a memory. The memory stores instructions, which, when executed by the at least one processor, cause the at least one processor to perform a vacuum circuit breaker life prediction method. The method includes acquiring multi-source monitoring data of the vacuum circuit breaker, preprocessing the multi-source monitoring data to extract multi-source fine-grained features, inputting the multi-source fine-grained features and real-time operating parameters into an operating condition perception weight network for dynamic weight calculation and feature fusion, constructing and training a life prediction model using a few-sample meta-learning strategy, performing sample calibration on a target circuit breaker to obtain a target prediction model, calculating a comprehensive health index and aging coefficient based on the fused features, and outputting the remaining service life and confidence assessment results.

[0073] Among them, multi-source monitoring data refers to current, mechanical waveform, partial discharge, environmental and contact temperature data collected by sensors. Specifically, Hall sensors, vibration sensors and infrared temperature measurement modules can be used to comprehensively cover the monitoring dimensions related to mechanical performance degradation and insulation aging.

[0074] Multi-source fine-grained features refer to the detailed signal features extracted from the original data. Specifically, fast Fourier transform can be used to extract energy spectrum features in a preset frequency band. For example, the 50-200Hz band reflects mechanical vibration, and the 1-2kHz band reflects mechanical jamming, solving the problem of missed early fault detection caused by high-frequency signal loss in traditional methods. The condition-aware weighted network refers to a computational model that dynamically adjusts feature weights based on real-time operating conditions. Specifically, a feedforward neural network structure can be used. The input layer receives ambient temperature and current fluctuation coefficients, and the output layer generates normalized weights through a softmax function, achieving adaptive feature fusion under unpreset operating conditions.

[0075] The few-shot meta-learning strategy refers to a pre-training method based on a multi-model, multi-operating-condition meta-task set. Specifically, it can use an improved CNN-BiGRU backbone network for meta-training, reduce computational complexity through depthwise separable convolutions, and improve cross-operating-condition generalization ability through layer normalization, enabling the target device to complete model calibration with only a small number of samples. Confidence assessment refers to the quantitative evaluation of the reliability of the prediction results. Specifically, the Monte Carlo Dropout method can be used to perform multiple forward inferences, and the stability of the results can be judged by calculating the standard deviation of the predicted values, avoiding the limitation of traditional point estimation methods in assessing the error range.

[0076] Specifically, when the processor executes instructions in the memory, it first acquires the current waveform, mechanical vibration signal and environmental parameters of the vacuum circuit breaker during operation through the multi-channel data acquisition module, and then uses frequency domain analysis methods to extract the energy spectrum characteristics that reflect different fault modes.

[0077] Subsequently, real-time operating parameters are input into the feedforward neural network to dynamically generate fusion weights for each feature. For example, the weight coefficient of contact wear features is automatically increased in high-temperature environments. Model parameters pre-trained using meta-learning are used as initial values, and parameter calibration is performed using a small number of samples collected from the target equipment. For instance, model adaptation is completed using 20 sets of operating data from a ZN28-12 circuit breaker in a new energy power station. Finally, based on the weighted fusion of the comprehensive health index and the environmental sensitivity coefficient, the aging rate is corrected, and the predicted remaining service life and its confidence interval are output.

[0078] Existing computing devices rely on fixed weight rules and single-model training when performing lifetime prediction. For example, adjusting weights using threshold triggering mechanisms can lead to prediction failure under unpredictable operating conditions, or require retraining the entire model for new device models. This solution achieves adaptive fusion of operating conditions through a dynamic weight network. For instance, when grid load fluctuations cause abnormal current harmonics, the weight ratio of mechanical vibration characteristics is automatically adjusted. The general model parameters established through meta-learning allow the target device to complete model adaptation with only 30 or fewer samples, significantly reducing the deployment cost of new device models.

[0079] This application can solve the problem of adaptability of dynamic weights under unknown operating conditions, such as accurately increasing the weight ratio of environmental corrosion health index in intermittent high temperature and high humidity environments; reduce the number of samples required for modeling new vacuum circuit breakers, for example, only 20 sets of samples are needed to complete model calibration for the ZN28-12 equipment; and improve the fine-grained level of multi-source feature fusion, for example, by analyzing the correlation between the energy spectrum of the tripping current and the phase of partial discharge, effectively capturing the coupled degradation signal of spring fatigue and insulation aging, thereby improving the accuracy of remaining life prediction.

[0080] This application further proposes a computer-readable storage medium storing a computer program. When executed by a processor, the program implements a method for predicting the lifespan of a vacuum circuit breaker, comprising the following steps: acquiring multi-source monitoring data of the vacuum circuit breaker, the multi-source monitoring data including at least current data, mechanical waveform data, partial discharge data, environmental data, and contact temperature data; preprocessing the multi-source monitoring data to extract multi-source fine-grained features, the fine-grained features including at least a preset frequency band energy spectrum feature extracted from the current data; inputting the multi-source fine-grained features and real-time operating parameters into an operating condition sensing weight network, calculating the dynamic weights corresponding to each feature, and basing the calculation on... Multi-source fine-grained features are weighted and fused using dynamic weights to obtain fused features. A few-shot meta-learning strategy is used to construct and train a life prediction model. The few-shot meta-learning strategy includes meta-pre-training based on a meta-task set containing multiple circuit breaker models and operating conditions to obtain general initialization parameters. For the target circuit breaker, no more than thirty samples are collected to calibrate the model and obtain the target prediction model. The fused features are input into the target prediction model to calculate the comprehensive health index and the aging coefficient after operating condition correction of the vacuum circuit breaker. The remaining service life of the vacuum circuit breaker is calculated based on the comprehensive health index and the aging coefficient, and the confidence level of the calculation results is evaluated.

[0081] The multi-source monitoring data refers to the current, mechanical vibration waveforms, partial discharge signals, ambient temperature and humidity, and contact temperature data collected by sensors. Specifically, this can be achieved using Hall effect sensors, vibration accelerometers, high-frequency current transformers, temperature and humidity sensors, and infrared thermometers. This data comprehensively reflects the mechanical performance, electrical performance, and environmental influencing factors of the vacuum circuit breaker. The preset frequency band energy spectrum features refer to the energy values ​​in the 50-200Hz and 1-2kHz frequency bands extracted after wavelet packet decomposition of the current signal. This can be implemented using the Daubechies wavelet basis function to capture early signals of abnormal mechanical vibration and contact jamming. The operating condition perception weight network is a dynamic weight generation module built based on a feedforward neural network. Specifically, it can be implemented using a three-layer network structure: an input layer receiving real-time operating condition parameters, a hidden layer with ReLU activation, and an output layer using Softmax normalization. This structure automatically adjusts the feature weights based on ambient temperature and current fluctuation coefficients. Among them, the few-shot meta-learning strategy refers to the method of obtaining model initialization parameters through multi-task pre-training. Specifically, it can be implemented using a model-agnostic meta-learning algorithm, which enables the model to quickly adapt to new equipment through meta-task training across different models and operating conditions. Confidence assessment refers to calculating the stability of prediction results through multiple inferences. Specifically, it can be implemented using the Monte Carlo Dropout method, which generates multiple prediction values ​​by randomly discarding neurons during the model inference phase and then performs statistical analysis.

[0082] Specifically, when the program instructions stored in the computer-readable storage medium are executed, they first collect comprehensive data on the operating status of the vacuum circuit breaker through multi-source sensors. In the preprocessing stage, wavelet packet transform is performed on the current signal to extract specific frequency band energy spectrum features reflecting mechanical vibration and contact jamming. This is combined with the opening speed features from the mechanical waveform data and the pulse phase distribution features from the partial discharge data to form a fine-grained feature set. Real-time operating parameters and fine-grained features are input together into the operating condition perception weight network. This network dynamically adjusts the contribution weight of each feature in the health index calculation by analyzing the current ambient temperature and current fluctuations. In the model building stage, a meta-learning strategy is used to pre-train the model on a meta-task set containing various circuit breaker models and operating conditions, enabling the model to generalize across devices. For specific target devices, only a small number of samples are needed to complete the model parameter calibration. In the final prediction stage, the fused features are input into the calibrated prediction model. The comprehensive health index and the operating condition-corrected aging coefficient are calculated by weighting, and the stability of the remaining life prediction results is verified using the Monte Carlo Dropout method.

[0083] Traditional methods adjust feature weights using fixed rules, which cannot adapt to complex combinations of unpredictable operating conditions. This solution, however, uses a condition-sensing network to dynamically allocate weights, automatically adjusting feature importance based on real-time environmental temperature and current fluctuations. Existing technologies require extensive new sample collection when applying new equipment, while this solution uses universal parameters obtained through meta-learning pre-training, allowing for model adaptation to new equipment with only a small number of samples. Traditional feature fusion methods use only single signals or coarse-grained features, while this solution effectively captures early-degrading signals by extracting fine-grained features such as energy spectra in preset frequency bands.

[0084] This application solves the problem of poor adaptability of dynamic weights and realizes the autonomous optimization allocation of feature weights under different operating conditions; overcomes the defect of modeling failure in scenarios with few samples and significantly reduces the sample collection cost for the application of new circuit breakers; improves the granularity of multi-source feature fusion and enhances the ability to identify early degradation features, thereby enhancing the engineering applicability of vacuum circuit breaker life prediction technology while ensuring prediction accuracy.

[0085] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the lifespan of a vacuum circuit breaker, characterized in that, Includes the following steps: Acquire multi-source monitoring data of the vacuum circuit breaker, wherein the multi-source monitoring data includes at least current data, mechanical waveform data, partial discharge data, environmental data, and contact temperature data; The multi-source monitoring data is preprocessed to extract multi-source fine-grained features, which include at least preset frequency band energy spectrum features extracted from the current data. The multi-source fine-grained features and real-time operating parameters are input into the operating condition perception weight network to calculate the dynamic weights corresponding to each feature. Based on the dynamic weights, the multi-source fine-grained features are weighted and fused to obtain fused features. A lifetime prediction model is constructed and trained using a few-shot meta-learning strategy, which includes meta-pre-training based on a meta-task set containing multiple circuit breaker models and operating conditions to obtain general initialization parameters. For the target circuit breaker, no more than thirty samples are collected to calibrate the model and obtain the target prediction model; The fused features are input into the target prediction model to calculate the comprehensive health index and the aging coefficient after operating condition correction of the vacuum circuit breaker. Based on the comprehensive health index and the aging coefficient, the remaining service life of the vacuum circuit breaker is calculated, and the confidence level of the calculation results is evaluated.

2. The method for predicting the lifespan of a vacuum circuit breaker according to claim 1, characterized in that, The operating condition sensing weight network is a feedforward neural network. The input of the operating condition sensing weight network is the real-time ambient temperature and the main circuit current fluctuation coefficient. The output layer uses the Softmax activation function to output normalized dynamic weights.

3. The method for predicting the lifespan of a vacuum circuit breaker according to claim 1, characterized in that, The specific steps of the meta-pre-training include: Construct a meta-task set containing at least five circuit breaker models and at least three typical operating conditions; In each meta-task, the fused features are used to train an improved CNN-BiGRU backbone network and update temporary parameters. Calculate the average loss for all meta-tasks and update the general initialization parameters accordingly.

4. The method for predicting the lifespan of a vacuum circuit breaker according to claim 3, characterized in that, In the improved CNN-BiGRU backbone network, the CNN part uses depthwise separable convolution, and the BiGRU part adds a layer normalization in the hidden layer.

5. The method for predicting the lifespan of a vacuum circuit breaker according to claim 1, characterized in that, The calculation of the comprehensive health index is specifically as follows: Based on the dynamic weights, the mechanical aging health index, contact wear health index, and insulation aging health index are weighted and averaged to obtain the comprehensive health index, which is calculated using the following formula: in, To assess overall health, , , These are the dynamic weights corresponding to the health index. , , These are the mechanical aging health index, the contact wear health index, and the insulation aging health index.

6. The method for predicting the lifespan of a vacuum circuit breaker according to claim 1, characterized in that, The formula used to calculate the aging coefficient after the operating condition correction is as follows: in, The aging factor is... This is the operating condition correction factor determined based on the main circuit current fluctuation coefficient. Based on the aging rate factor, is the environmental sensitivity coefficient, and is the environmental corrosion health index.

7. The method for predicting the lifespan of a vacuum circuit breaker according to claim 1, characterized in that, After calculating the remaining service life of the vacuum circuit breaker, the method further includes a step of assessing the confidence level of the calculation results: The Monte Carlo Dropout method is used to perform multiple forward inferences on the target prediction model to obtain multiple remaining useful life prediction values. Calculate the standard deviation of the multiple predicted values. If the standard deviation is lower than a preset threshold of 5%, the confidence level is determined to be up to standard and the final life prediction result is output.

8. The method for predicting the lifespan of a vacuum circuit breaker according to claim 1, characterized in that, The preset frequency band energy spectrum features extracted from the current data include energy values ​​in the 50-200Hz frequency band that reflect mechanical vibration, and energy values ​​in the 1-2kHz frequency band that reflect mechanical jamming.

9. A computing device, characterized in that, include: At least one processor; as well as A memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the vacuum circuit breaker lifetime prediction method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the vacuum circuit breaker life prediction method as described in any one of claims 1-8.

Images