A high-voltage circuit breaker residual life prediction method based on PIO-XGBoost

Abstract

The application discloses a high-voltage circuit breaker residual life prediction method based on PIO-XGBoost, comprising the following steps: collecting the multi-dimensional mechanical characteristic parameters of the current node of the high-voltage circuit breaker, and splicing the multi-dimensional mechanical characteristic parameters to form mechanical characteristic data after preprocessing; constructing an XGBoost life stage classification submodel and an XGBoost residual life regression submodel, and optimizing the hyperparameter combinations of the two models through a pigeon optimization algorithm; inputting the mechanical characteristic data into the optimized XGBoost life stage classification submodel and the XGBoost residual life regression submodel to obtain the high-voltage circuit breaker current life stage category probability distribution and the residual operable number; calculating the prediction uncertainty of each model, combining the prediction uncertainty to determine the consistency of the high-voltage circuit breaker current life stage category probability and the residual operable number, and if the determination fails, selecting the corresponding model result output according to the prediction uncertainty and the decision path similarity.

Description

Technical Field

[0001] This invention relates to the field of high-voltage circuit breaker remaining life prediction technology, and in particular to a high-voltage circuit breaker remaining life prediction method based on PIO-XGBoost. Background Technology

[0002] As the core equipment in the power system that performs control and protection functions, the mechanical reliability of high-voltage circuit breakers under rated operating conditions has a significant impact on power grid safety. Because the components of the operating mechanism undergo wear, fatigue, and aging during service (e.g., spring fatigue, lubricant degradation), the mechanical characteristic parameters of the circuit breaker (such as opening time, closing time, opening and closing speed, asynchrony of opening and closing, coil current, and energy storage time) exhibit a regular degradation trend with the increase in the number of operations.

[0003] Regarding the aforementioned degradation process, if the current stage of the circuit breaker's lifespan degradation can be determined based on the multidimensional mechanical characteristic parameters of the circuit breaker at different operation cycles, and the remaining mechanically operable lifespan can be quantitatively predicted, then a quantitative basis can be provided for the formulation of preventive maintenance and repair plans, thereby reducing the risk of unplanned outages and maintenance costs.

[0004] Currently, circuit breaker mechanical life assessment and remaining life prediction fall into two main categories: First, physical model-based methods. These methods derive life decay laws by establishing motion equations for the operating mechanism, spring fatigue models, or related mechanical mechanisms. Their implementation typically relies on structural, material, and dynamic parameters to calculate or estimate the life decay process, and then determine the remaining life or life status based on the decay results. Second, data-driven methods. These methods use vibration and current signals collected during circuit breaker testing, combined with algorithms such as Support Vector Regression (SVR) and neural networks, to predict life. Typically, this involves extracting signal features at a single moment or node as input, and then outputting life estimation results or life status determination results through the model.

[0005] Existing data-driven methods often use single-time parameters or fixed thresholds for lifetime assessment, such as using whether the parameters at the end of the test exceed the limit as a pass / fail criterion. These methods fail to dynamically characterize the degradation evolution of mechanical property parameters with the number of operations, making it difficult to identify lifetime degradation risks before the end of the lifespan.

[0006] Data-driven methods often use only a single parameter feature (such as only using opening or closing time) for modeling, or fail to fully utilize the synergistic changes of multi-dimensional parameters such as closing speed, coil current, and energy storage time during the degradation process, thus limiting the model's ability to characterize the lifetime degradation mechanism and affecting prediction accuracy.

[0007] Lifespan prediction tasks are typically characterized by a limited number of samples and complex coupling of multiple parameters. Under these conditions, the performance of ensemble learning models (such as XGBoost) or other machine learning models is sensitive to hyperparameter configuration. Manual parameter tuning or grid search is often inefficient and prone to getting trapped in local optima, making it difficult to guarantee prediction accuracy and stability.

[0008] Existing data-driven methods for circuit breaker life assessment or remaining life prediction typically employ a single model or a single output. While these methods offer some improvement in average prediction accuracy, they still face reliability issues: when input mechanical characteristic parameters are subject to noise, operating condition deviations, or changes in feature distribution, the model may produce contradictory results in life stage classification and remaining life regression. For example, the range of operational times implied by the stage classification result may differ from the remaining operational times obtained through regression, thus reducing the credibility of the prediction conclusions. Relying solely on the output value of a single prediction makes it difficult to determine whether the model's current prediction falls within a "credible region." In scenarios with multiple models or multi-task structures, without methods to quantify the confidence level of model predictions, the prediction results cannot effectively implement reliable risk classification and mitigation strategies.

[0009] To address the problems of the existing technology, a method for predicting the remaining life of high-voltage circuit breakers based on PIO-XGBoost is designed. Summary of the Invention

[0010] This invention provides a method for predicting the remaining life of high-voltage circuit breakers based on PIO-XGBoost, comprising: S1 collects the multi-dimensional mechanical characteristic parameters of the current node of the high-voltage circuit breaker, preprocesses them, and then splices the multi-dimensional mechanical characteristic parameters to form mechanical feature data. S2 generates life stage labels according to the preset number of operations in the training samples. Based on the sequential stacking of several decision trees, an XGBoost life stage classification sub-model is constructed with mechanical feature data as input and life stage category probability distribution as output. Based on the sequential stacking of several decision trees, an XGBoost remaining life regression sub-model is constructed with mechanical feature data as input and the remaining number of operable operations of the circuit breaker to the rated mechanical life as output. The hyperparameter combination of the two models is optimized by pigeon flock optimization algorithm. S3 inputs the mechanical feature data into the optimized XGBoost life stage classification sub-model and XGBoost remaining life regression sub-model to obtain the current life stage category probability distribution and remaining operable times of the high-voltage circuit breaker. S4 calculates the prediction uncertainty of each model, and combines the prediction uncertainty to determine the consistency of the current life stage category probability and remaining operable times of the high-voltage circuit breaker. If the determination fails, the corresponding model result is output based on the prediction uncertainty and decision path similarity.

[0011] The beneficial effects of this invention are: First, by comprehensively and multidimensionally collecting the mechanical characteristic parameters of high-voltage circuit breakers and performing refined preprocessing and feature engineering, the subtle performance degradation of the equipment under complex operating conditions is effectively captured, enhancing the ability of features to represent the health status of the equipment. This provides high-quality and structured input data for the accurate prediction of subsequent models, thereby significantly improving the accuracy of predictions and supporting the efficient training of models.

[0012] Secondly, the Pigeon Optimization (PIO) algorithm is used to efficiently optimize the hyperparameters of the XGBoost classification and regression sub-models, significantly improving the model's predictive performance and generalization ability, effectively avoiding overfitting, and enhancing the model's robustness under complex real-world conditions. This automated hyperparameter tuning process not only improves model development efficiency but also ensures that the model can consistently and stably output high-quality prediction results.

[0013] Third, by synergistically utilizing optimized and complementary XGBoost classification and regression sub-models, the life status of high-voltage circuit breakers can be comprehensively and accurately assessed from both qualitative (life stage category probability) and quantitative (remaining operable times) dimensions. This multi-dimensional and complementary prediction system not only overcomes the limitations of single-model prediction information but also provides a rich and necessary data foundation for subsequent key consistency verification steps, thereby making the overall prediction results more informative and valuable for decision guidance.

[0014] Fourth, by introducing consistency verification of prediction results, model uncertainty assessment, and decision path similarity analysis, the reliability and intelligence level of the prediction system are significantly enhanced. This step can intelligently diagnose the nature of contradictions in multi-model predictions, and achieve automated decision-making in some complex situations (such as when decision paths are similar and one model has low uncertainty), avoiding blind acceptance or unnecessary human intervention. At the same time, for the most complex or dangerous contradiction situations, this step can trigger timely and evidence-based human review and provide in-depth diagnostic information, thereby optimizing the efficiency and accuracy of human intervention and ensuring that the prediction system can make safe and reliable decisions under any circumstances. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings required 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.

[0016] Figure 1 This is a flowchart of the method in this embodiment. Detailed Implementation

[0017] To facilitate understanding by those skilled in the art, the structure of the present invention will now be described in further detail with reference to the accompanying drawings. It should be understood that, unless otherwise specified, the order of the steps mentioned in this embodiment can be adjusted according to actual needs, and they can even be executed simultaneously or partially simultaneously.

[0018] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost, including: S1 collects the multi-dimensional mechanical characteristic parameters of the current node of the high-voltage circuit breaker, preprocesses them, and then splices the multi-dimensional mechanical characteristic parameters to form mechanical feature data. S101 collects the mechanical characteristic parameters of the current high-voltage circuit breaker under both opening and closing operations. S1011 collects the opening time, average opening speed, asynchrony of opening, and maximum current of the opening coil of the high-voltage circuit breaker. S1012 collects data on the closing time, average closing speed, closing asynchrony, maximum current of the closing coil, energy storage time, and starting current of the energy storage motor of the high-voltage circuit breaker.

[0019] In this step, the mechanical life test of the high-voltage circuit breaker is used as the data source. After the circuit breaker completes 0, 2000, 4000, 6000, 8000, and 10000 operating cycles, mechanical characteristic tests are performed at three control voltages: rated voltage (DC220V), maximum voltage (DC242V), and minimum voltage (DC143V). The following multi-dimensional parameters are collected: Tripping characteristic parameters: tripping time (ms), average tripping speed (m / s), tripping asynchrony (ms), and maximum tripping coil current (A). Closing characteristic parameters: closing time (ms), average closing speed (m / s), closing asynchrony (ms), maximum current of closing coil (A), energy storage time (s), energy storage motor starting current (A).

[0020] The aforementioned parameters were repeatedly collected at each operation node, with the parameters under the rated control voltage serving as the primary modeling feature. Simultaneously, the average values ​​of the parameters at different voltage levels were extracted as a comprehensive feature. The parameter data collected at each node constituted the parameter evolution sequence of the circuit breaker throughout its entire lifespan. In actual lifespan prediction, the corresponding data were also collected as input for subsequent models.

[0021] S102 For each dimension of mechanical characteristic parameter, the minimum and maximum values ​​of that dimension parameter are determined based on the training set samples, and the dimension parameter is mapped to the [0,1] interval according to the minimum and maximum normalization method; In this step, due to the significant differences in the dimensions and numerical ranges of the various mechanical characteristic parameters, it is necessary to normalize all parameters. This invention employs the following method to uniformly map each parameter to the [0,1] interval:

[0022] in For the first The first sample The original value of the dimension parameter, and The training set The minimum and maximum values ​​of the dimension parameter. These are the normalized values. The normalization process used... and All values ​​are determined by fitting the training set data and are uniformly applied to the validation and test sets.

[0023] S103 concatenates the mechanical characteristic parameters under the current operation in parameter dimension order to form mechanical characteristic data.

[0024] In this step, the measured values ​​of all mechanical characteristic parameters at the current operation node are directly concatenated to form the feature vector of the sample. :

[0025] in, The total dimension of the parameters. For the first The normalized values ​​of each mechanical characteristic parameter at the current measurement node. This feature vector directly represents the mechanical state of the circuit breaker at this operation count node using the current measured values ​​of each parameter. Since the parameter values ​​corresponding to different operation count nodes inherently imply differences in the degree of degradation, the XGBoost model can automatically learn the mapping relationship between each parameter value and the life stage through nonlinear combination of these raw features. These features are directly used as input feature vectors for the XGBoost life stage classification sub-model and the XGBoost remaining life regression sub-model for model training and predictive inference.

[0026] S2 generates life stage labels according to the preset number of operations in the training samples. Based on the sequential stacking of several decision trees, an XGBoost life stage classification sub-model is constructed with mechanical feature data as input and life stage category probability distribution as output. Based on the sequential stacking of several decision trees, an XGBoost remaining life regression sub-model is constructed with mechanical feature data as input and the remaining number of operable operations of the circuit breaker to the rated mechanical life as output. The hyperparameter combination of the two models is optimized by pigeon flock optimization algorithm. In this step, XGBoost is an optimized version of Gradient Boosting Decision Tree (GBDT) and an iterative ensemble learning algorithm. An ensemble model consisting of decision trees, which is effective for the first decision tree. The predicted value for each sample can be expressed as:

[0027] in, Representing the A decision tree, For the first The input feature vector of each sample is the constructed feature vector of mechanical characteristic parameters. XGBoost completes training by minimizing the objective function consisting of the loss function and model complexity.

[0028] in, Measure the error between the predicted lifespan stage and the actual lifespan stage label; This is a regularization term to suppress model overfitting. For the lifespan prediction task, this invention simultaneously constructs two types of sub-models, with the specific steps as follows: S201 uses multi-class cross-entropy as the loss function for the XGBoost lifetime stage classification sub-model, and iteratively updates the model parameters based on this loss function. The calculation formula is as follows: , in, This represents the cross-entropy loss function for multi-class classification. Indicates the category number, Indicates the total number of categories. Indicates an indicator function, This indicates the true lifespan stage category of the training samples. This represents the feature vector of the training sample. Indicates input Below, the prediction category is The conditional probability, This indicates taking the logarithm of the probability; In this step, lifetime prediction is treated as an N-class task, with each class corresponding to a lifetime degradation stage from 0 to N operations, labeled 0-N respectively. The model input is the normalized feature vector. Each component corresponds to a normalized measured value of a mechanical characteristic parameter at the current measurement node.

[0029] The XGBoost classification sub-model is composed of The decision trees are formed by stacking them sequentially. The structure function of a tree is denoted as The model is based on the input samples. The Categories The original prediction score for ) is:

[0030] in, Indicates the first The tree is for the first The score contribution of each category is considered. The raw scores are then transformed into probability distributions for each category using the Softmax function.

[0031] Ultimately, the category with the highest probability will be used as the prediction result for lifespan stage:

[0032] The model is trained using multi-class cross-entropy as the loss function. For indicator functions, when the real label equals The value is 1 if the condition is met, and 0 otherwise. The model, with its hyperparameters optimized using PIO on the training set, aims to minimize the cross-entropy loss on the training set, performing gradient boosting iteratively tree by tree until the optimal number of iterations is reached. Finally, the classification model that achieves the best generalization performance on the test set is output.

[0033] Using high-voltage circuit breaker mechanical life tests as the data source, and dividing the circuit breaker lifespan into stages based on the number of operations, this invention serves as a classification and prediction target. The mechanical life test data of this invention can be divided into the following stages: Phase 0 (Initial Phase): The number of operations is [0, 1000); Phase 1 (Early break-in phase): Approximately 1000 to 3000 operations; Phase 2 (Normal Operation Phase): Approximately 3000 to 5000 operations. Phase 3 (Intermediate Degradation Phase): Approximately 5000, 7000 operations; Phase 4 (Accelerated Degradation Phase): Approximately 7000, 9000 operations; Phase 5 (End of Life Stage): Approximately [9000, 10000] operations; For samples located between two test nodes, the lifetime stage label is determined based on the range of operations completed. For example, if 3000 operations have been completed, it belongs to stage 2; if 5800 operations have been completed, it belongs to stage 3.

[0034]

[0035] The parameter measurement records corresponding to each stage are used as labeled samples for training and validating the PIO-XGBoost life prediction model. For each circuit breaker, the mechanical characteristic parameter measurements collected at each operation count node are organized sequentially in ascending order of operation count to form a numerical sequence for each parameter. Taking the opening time of the mechanical life test data of this invention as an example, its numerical sequence is as follows:

[0036] in Indicates completion The opening time was measured after the operation. The same process was applied to other mechanical characteristic parameters (closing time, average opening speed, average closing speed, opening asynchrony, closing asynchrony, maximum current in the opening coil, maximum current in the closing coil, energy storage time, starting current of the energy storage motor, etc.), generating a numerical sequence of length N for each parameter. This sequence serves as the numerical sample for subsequent model training.

[0037] Given the limited number of samples at each operation count node in the circuit breaker mechanical life test data, directly using it for model training may lead to overfitting. This invention employs a Gaussian noise-based data augmentation strategy, adding a normally distributed random perturbation to each original sample to generate several augmented samples:

[0038] in, For the enhanced sample values, For a mean of 0 and a standard deviation of Random noise sampled from a normal distribution, Set as the first in the training set The standard deviation of the augmented sample is set to 5% to ensure that the augmented sample fluctuates within a reasonable neighborhood of the original sample, without introducing data distortion beyond the actual measurement error range. The above process is repeated for each original sample. The sample size for each lifespan stage is then increased to a sufficient quantity to support model training. Data augmentation is performed only on the training set; the validation and test sets remain unchanged to ensure the objectivity of the evaluation results.

[0039] S202 uses the mean squared error as the loss function for the XGBoost remaining lifetime regression sub-model, and iteratively updates the model parameters based on this loss function. The calculation formula is as follows: , in, This represents the mean squared error loss function. This indicates the number of samples involved in the mean squared error calculation. Indicates the sample index. This represents the actual remaining number of operations for the i-th sample. This indicates that the i-th sample is in the input At that time, the regression model predicts the remaining number of operations. This represents the feature vector of the i-th sample; In this step, the remaining lifespan regression sub-model uses the same normalized feature vector. As input, the number of remaining operable cycles of the circuit breaker before its rated mechanical life. As a continuous prediction target:

[0040] in Next is the rated mechanical life. This represents the number of operations completed so far. Taking the mechanical life test data of this invention as an example, each operation count node corresponds to... The actual values ​​were 10000, 8000, 6000, 4000, 2000, and 0, respectively. The predicted output of the XGBoost regression sub-model was:

[0041] During training, the mean squared error (MSE) is used as the loss function, where The number of samples in the training set. and The first The model calculates the true remaining lifetime and model predictions for each sample. The model uses PIO optimization to determine the optimal hyperparameter configuration, and then performs gradient boosting iterations on the training set with the objective of minimizing the MSE, ultimately outputting a continuous estimate of the remaining operable cycles of the circuit breaker under test. .

[0042] The classification and regression sub-models operate independently during the training phase, with each model determining its optimal hyperparameters through its respective PIO optimization process. During the prediction phase, both sub-models simultaneously infer the same feature vector. The output of the classification sub-model provides interpretable discrete lifetime stage judgments, while the output of the regression sub-model provides a continuous quantitative estimate of remaining lifetime. The combination of these two sub-models constitutes a complete lifetime prediction result.

[0043] S203 constructs the solution vector of hyperparameters to be optimized for the pigeon flock optimization algorithm by using the model learning rate, maximum depth, minimum leaf node weight, subsample ratio, and number of iterations. S204 uses the pigeon flocking optimization algorithm to perform a phased global search and a landmark operator phased directional convergence optimization in the map-compass phased update mechanism of the pigeon flocking optimization algorithm on the hyperparameter solution vector to be optimized, and obtains the optimal hyperparameter solution vector corresponding to the model.

[0044] S2041 initializes the pigeon flock algorithm, maps the position of each pigeon to a candidate hyperparameter solution vector, and uses the candidate hyperparameter solution vectors to train the XGBoost life stage classification sub-model and the XGBoost remaining life regression sub-model respectively, to obtain the classifier and regressor corresponding to each candidate solution vector; In the current iteration, S2042 uses five-fold cross-validation to obtain the average accuracy of any candidate hyperparameter solution vector in the pigeon flock as its fitness value, and updates the historical best position corresponding to the candidate hyperparameter solution vector and the global best position of the group based on the fitness value. This process is repeated until the fitness evaluation of the current set of candidate hyperparameter solution vectors is completed. In the current iteration, S2043 uses five-fold cross-validation to obtain the reciprocal of the five-fold average RMSE of the XGBoost residual lifetime regression submodel corresponding to any candidate hyperparameter solution vector in the pigeon flock, and updates the historical best position corresponding to the candidate hyperparameter solution vector and the global best position of the group based on the fitness value. This process is repeated until the fitness evaluation of the current set of candidate hyperparameter solution vectors is completed. After completing the fitness evaluation of the current iteration round and updating the historical and global optimal positions, S2044 executes the map-compass stage update mechanism of the pigeon flock optimization algorithm. It updates the speed and position of each pigeon based on its historical optimal position and the global optimal position of the group, generates a new set of candidate hyperparameter solution vectors, and enters the next iteration round until the termination condition is met. S2045 When the termination condition is met, the candidate hyperparameter solution vector with the highest fitness value is selected as the optimal hyperparameter solution vector. The classification optimal hyperparameter solution vector is used to construct the XGBoost lifetime stage classification sub-model, and the regression optimal hyperparameter solution vector is used to construct the XGBoost remaining lifetime regression sub-model.

[0045] The predictive performance of the XGBoost model is highly dependent on hyperparameter configuration. This invention introduces Pigeon-inspired Optimization (PIO) to automatically optimize key hyperparameters of XGBoost, adapting to the parameter optimization needs in small-sample scenarios of lifetime prediction. PIO uses a phased search with map-compass and landmark operators. Each pigeon represents a candidate solution in the hyperparameter space, and its position vector is composed of the XGBoost hyperparameters to be optimized. The map-compass phase is used for global search, and the pigeon's speed and position are updated according to the following formula:

[0046]

[0047] In the formula, and The first Only one pigeon in the 1st The speed and position of the next iteration; For map-compass factors; This is the globally optimal position within the current population. for Random numbers are generated between the given numbers. This stage gradually shrinks the search area by learning from the global optimum. When the number of iterations exceeds a set threshold, the algorithm enters the landmark operator stage, simulating the behavior of a flock of pigeons using landmarks for navigation as they approach their destination. In this stage, the population size is halved in each iteration, retaining individuals with better fitness, and the remaining pigeons calculate the weighted center. Subsequently, all individuals move towards the center, causing the algorithm to quickly converge to the globally optimal region.

[0048]

[0049] For the lifespan prediction task, five parameters that significantly affect model performance are selected to form the solution vector: learning rate. Maximum depth of the tree Minimum leaf node weights Subsample ratio and number of iterations .

[0050] For classification tasks, the accuracy is used, and for regression tasks, the inverse of the root mean square error (RMSE) is used. The average prediction accuracy of XGBoost's five-fold cross-validation on the training set is used as the fitness function. PIO continuously updates the hyperparameter combination during the iteration process and finally outputs the optimal parameters that maximize the fitness, thereby constructing the PIO-XGBoost lifetime prediction model.

[0051] S3 inputs the mechanical feature data into the optimized XGBoost life stage classification sub-model and XGBoost remaining life regression sub-model to obtain the current life stage category probability distribution and remaining operable times of the high-voltage circuit breaker. S301 inputs mechanical feature data into the optimized XGBoost life stage classification sub-model. The model calculates the score of the leaf node of each decision tree in turn, and outputs the predicted probability distribution of each category after being transformed by the Softmax function. The category with the highest probability is taken as the life stage prediction result. S302 inputs mechanical characteristic data into the optimized XGBoost remaining life regression sub-model and outputs a continuous estimate of the remaining number of operable operations. In this step, Input a trained PIO-XGBoost lifespan stage classification sub-model. The model calculates the score for each leaf node in sequence, and outputs the predicted probability distribution of N lifespan stages after transformation by the Softmax function. The category with the highest probability is the life stage prediction result. In addition to the category with the highest probability, the second highest probability category and its probability value are also recorded for subsequent uncertainty assessment.

[0052] S303 trims the continuous estimate of the remaining number of operations, constraining it to the interval [0, ... Within this range, the final number of remaining operations is obtained, calculated using the following formula: , in, This indicates the final number of operations remaining after trimming. This represents the function that takes the minimum value. This represents the function that takes the maximum value. This indicates the maximum rated mechanical life of the circuit breaker.

[0053] In this step, the same Input the trained PIO-XGBoost remaining lifetime regression sub-model and output a continuous estimate of the remaining number of operations. (Unit: times). To ensure the reasonableness of the output results, boundary constraints are applied to the regression output, that is, the predicted values ​​are clipped to [0, ... Within the specified range, avoid unreasonable outputs that result in negative values ​​or exceed the rated lifespan due to model extrapolation.

[0054] S4 calculates the prediction uncertainty of each model, and combines the prediction uncertainty to determine the consistency of the current life stage category probability and remaining operable times of the high-voltage circuit breaker. If the determination fails, the corresponding model result is output based on the prediction uncertainty and decision path similarity.

[0055] S401 calculates the entropy of the predicted probability distribution of the current sample as its prediction uncertainty through the XGBoost lifetime stage classification sub-model, and calculates the prediction standard deviation based on the integrated prediction results through the XGBoost remaining lifetime regression sub-model as its prediction uncertainty. For the XGBoost lifetime stage classification sub-model, S4011 uses the entropy of the predicted probability distribution as the prediction uncertainty, calculated as follows: , in, Entropy represents the predicted probability distribution. Indicates the category number, Indicates the total number of categories. This represents the feature vector of the test sample to be predicted after preprocessing. Indicates input Below, the prediction category is The conditional probability, This indicates taking the logarithm of the probability; In this step, when a sample is classified into a certain stage, if the probabilities of each stage are nearly uniform (high entropy), it indicates that the classification sub-model itself is uncertain; if the probability of a certain class is close to 1 (low entropy), it indicates that the classification decision is very confident. Therefore, entropy is the confidence level of the classification model, which does not require additional calibration and is directly derived from the model output.

[0056] S4012 performs predictions on several XGBoost remaining lifetime regression sub-models constructed from different sub-samplings during the inference phase, obtaining several estimates of remaining operable times. The standard deviation of these estimates is calculated as the prediction uncertainty, as shown in the following formula: , in, The standard deviation represents the number of remaining operands. This indicates the number of XGBoost remaining lifetime regression sub-models used for ensemble prediction. This represents the remaining number of operations predicted by the m-th XGBoost remaining lifetime regression sub-model.

[0057] In this step, the regression sub-model is an integration of multiple sub-models constructed from "different subsamples" during the inference phase (this is a key feature of the PIO-optimized XGBoost in this application). Each independent sub-model provides a residual count estimate, and the dispersion (standard deviation) of these estimates directly reflects the regression sub-model's confidence in the prediction. σ is small (high confidence) when the predictions among the subsamples are highly consistent, and large (low confidence) when there is significant divergence, thus transforming into a computable confidence signal.

[0058] S402 If the entropy of the predicted probability distribution is greater than the upper limit of entropy, the classification is judged to be low confidence and manual review is triggered; In this step, all subsequent consistency determinations are based on the lifetime stage given by the sub-model (and thus its remaining lifetime reference interval). If the sub-model itself is unsure about the current sample (high entropy), then the reference interval it provides is inherently problematic. Using this unreliable interval to verify the regression results may result in either false positives or false negatives.

[0059] Upper limit of entropy Using the validation set from the model training phase, all correctly classified samples are selected, and their prediction entropy is calculated. The 95th percentile value is then taken. This means that among the samples correctly classified by the model, 95% of the samples have a prediction entropy that does not exceed this value. If the prediction entropy of a new sample exceeds this upper limit, it indicates that its uncertainty has exceeded the model's performance within a normal and reliable range, and the classification result is deemed unreliable.

[0060] S403 If the entropy of the predicted probability distribution is less than or equal to the upper limit of entropy, obtain the remaining lifetime reference interval corresponding to the lifetime stage prediction result, calculate the absolute deviation of the remaining operable counts from the center of the remaining lifetime reference interval, and calculate its ratio to the prediction standard deviation. The calculation formula is as follows: , , in, This represents the ratio of absolute deviation to the predicted standard deviation. Indicates the remaining number of operations. Indicates the center of the remaining lifetime reference interval. Indicates the standard deviation of the forecast. and These represent the upper and lower bounds of the remaining lifetime reference interval, respectively. In this step, assuming the classification is reliable, the deviation is divided by the model's own prediction standard deviation. The same absolute deviation might be a small fluctuation for a model with low confidence, but a serious anomaly for a highly confident model, indicating that the deviation value does not fall within the model's normal fluctuation range. Simultaneously, selecting the center of the reference interval as a benchmark provides a central representation of all possible values ​​for this stage, offering a basis for subsequent stratification determination.

[0061] S404 determines the consistency between two models based on the remaining number of operands and the ratio of absolute deviation to the prediction standard deviation. If the consistency determination is passed, it outputs the remaining number of operands and its lifetime stage. S4041 If the remaining number of operable attempts is within the remaining lifetime reference range, the results of the two models are determined to be consistent, and the remaining number of operable attempts and their lifetime stage are output. S4042 If the remaining operable counts are outside the remaining lifetime reference range, and the ratio of the absolute deviation to the predicted standard deviation is greater than the ratio threshold, then the regression is determined to be low confidence and manual review is triggered. S4043 If the remaining operable counts are outside the remaining lifetime reference range, and the ratio of the absolute deviation to the prediction standard deviation is less than or equal to the ratio threshold, then the model results are considered inconsistent.

[0062] In this step, based on the positional relationship between the remaining operable times falling inside and outside the reference interval and the normalized ratio, the consistency determination of the two models is divided into three mutually exclusive and clearly defined cases, replacing the simple and crude "consistent / contradictory" dichotomy, because the dichotomy cannot handle the problem of falling outside the interval but with a small deviation.

[0063] When the regression output falls within the classification reference interval, even if it is not in the center of the interval, it is still within the normal range of values ​​for that stage. The two models do not conflict with each other, and direct output is the fastest path. When the regression output goes outside the range and still deviates significantly relative to its own uncertainty, it indicates that the regression has been significantly distorted. This is a serious deviation even when there is no certainty, which is the strongest signal of regression failure. No arbitration is worth accepting, and it should be manually reviewed directly. When the regression output is outside the range but the normalization deviation is not significant, it indicates that the regression is not obviously distorted but is indeed outside the range of this stage. This is a genuine contradiction where both parties have certain reasons and it cannot be decided by simple rules. It is also the only scenario that requires further in-depth arbitration.

[0064] If the consistency determination fails in S405, the prediction uncertainty of the two models is obtained. If the prediction uncertainty of both models is higher or lower than their corresponding uncertainty threshold, manual review is triggered. In this step, if the uncertainty of both models is higher than their threshold, it means that the results of both models are inaccurate and must be manually reviewed; if the uncertainty of both models is lower than their threshold, it means that the two confident models have reached contradictory conclusions, and at least one model has a problem, which also requires manual review.

[0065] S406 If only one model's prediction uncertainty is higher than the corresponding uncertainty threshold, extract the actual decision paths from the root node to the leaf node of each decision tree in the life stage classification sub-model and the remaining life regression sub-model for the current mechanical feature data, calculate the similarity of the decision paths of the two models, and output the model result with lower prediction uncertainty if the decision paths are similar.

[0066] S4061 For the current mechanical feature data, for each decision tree in each sub-model, extract the internal nodes that the current mechanical feature data actually passes through from the root node to the leaf node, and obtain the actual decision path of the sample corresponding to the decision tree. S4062 records the split feature identifier, split operator, and split threshold to form a triplet for each internal node of the actual decision path; S4063 records the node layer number, starting with the root node having a layer number of 0, and increasing layer by layer along the direction from the root node to the leaf node. The depth weight of each triplet is calculated using the following formula: , in, Indicates depth weights, Indicates the number of node layers; S4064 identifies all common triples in two decision paths that have the same feature identifier, the same splitting operator, and a splitting threshold within a preset tolerance range. For each common triple, it obtains the node layer number and depth weight of the corresponding node in the two decision paths, and selects the smaller depth weight as the matching weight. S4065 calculates the similarity score of decision paths based on the sum of the depth weights of the triples in the two decision paths and the sum of the matching weights of the common triples. If the similarity score is greater than the similarity threshold, the decision paths are considered similar. The calculation formula is as follows: , in, This represents the sum of the matching weights of the common triples. This represents the sum of depth weights in the decision path of the XGBoost lifetime stage classification submodel. This represents the sum of depth weights in the decision path of the XGBoost remaining lifetime regression sub-model.

[0067] In this step, when predictions are contradictory but at least one model is relatively certain, we attempt to intelligently select the more reliable prediction result by analyzing the model's internal reasoning logic, thereby reducing unnecessary human intervention.

[0068] The key steps of making the complex decision-making logic within the XGBoost model (i.e., the splitting judgments of each decision tree on features) explicit, structured, and computable are revealed. By capturing which features, under what conditions, and at what threshold the model makes judgments at each internal node, the model's inference chain is obtained.

[0069] The splitting feature identifier, splitting operator, and splitting threshold together constitute the judgment conditions of a node. When there are common triples in two decision paths, it indicates that both the life stage classification sub-model and the remaining life regression sub-model are making judgments based on the same mechanical features and similar thresholds. Therefore, the more common triples there are, the more consistent the judgment criteria of the two models are.

[0070] Splitting conditions closer to the root node affect more subsequent branches, and their impact on decision paths and prediction results is usually greater. Therefore, depth weights are calculated based on the node level, giving higher weights to triples closer to the root node, and avoiding treating identical rules near leaf nodes with relatively minor impacts as the same as key rules at the root node.

[0071] For common triples, the smaller of their corresponding depth weights in the two decision paths is taken as the matching weight. If a rule only appears in a deeper node of one model, it indicates that the rule has a relatively limited effect in the other model. Using a smaller weight can avoid overestimating the decision consistency between the two models due to the high dependence of the rule on one side of the model.

[0072] Weighted Jaccard similarity is used to calculate the similarity score of decision paths. The sum of matching weights for common triples represents the importance of the common discriminative criteria used by the two sub-models in the current mechanical feature data prediction process. The union weight of the depth weights of all triples in the two decision paths represents the overall discriminative criteria of the two sub-models. By introducing depth weights corresponding to the node layer number, key splitting rules closer to the root node have a higher impact on the similarity score. By using the smaller depth weight from the two decision paths as the matching weight for common triples, the discriminative consistency of the two sub-models on the same rule is avoided from being overestimated.

[0073] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0078] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0079] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0080] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

Claims

1. A method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost, characterized in that, include: S1 collects the multi-dimensional mechanical characteristic parameters of the current node of the high-voltage circuit breaker, preprocesses them, and then splices the multi-dimensional mechanical characteristic parameters to form mechanical feature data. S2 generates life stage labels according to the preset number of operations in the training samples. Based on the sequential stacking of several decision trees, an XGBoost life stage classification sub-model is constructed with mechanical feature data as input and life stage category probability distribution as output. Based on the sequential stacking of several decision trees, an XGBoost remaining life regression sub-model is constructed with mechanical feature data as input and the remaining number of operable operations of the circuit breaker to the rated mechanical life as output. The hyperparameter combination of the two models is optimized by pigeon flock optimization algorithm. S3 inputs the mechanical feature data into the optimized XGBoost life stage classification sub-model and XGBoost remaining life regression sub-model to obtain the current life stage category probability distribution and remaining operable times of the high-voltage circuit breaker. S4 calculates the prediction uncertainty of each model, and combines the prediction uncertainty to determine the consistency of the current life stage category probability and remaining operable times of the high-voltage circuit breaker. If the determination fails, the corresponding model result is output based on the prediction uncertainty and decision path similarity.

2. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 1, characterized in that, S1 includes: S101 collects the mechanical characteristic parameters of the current high-voltage circuit breaker under both opening and closing operations. S102 For each mechanical characteristic parameter, the minimum and maximum values ​​of that parameter are determined based on the training set samples, and the parameter is mapped to the [0,1] interval according to the minimum and maximum normalization method; S103 concatenates the mechanical characteristic parameters under the current operation in parameter dimension order to form mechanical characteristic data.

3. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 2, characterized in that, S101 includes: S1011 collects the opening time, average opening speed, asynchrony of opening, and maximum current of the opening coil of the high-voltage circuit breaker. S1012 collects data on the closing time, average closing speed, closing asynchrony, maximum current of the closing coil, energy storage time, and starting current of the energy storage motor of the high-voltage circuit breaker.

4. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 1, characterized in that, S2 includes: S201 uses multi-class cross-entropy as the loss function for the XGBoost lifetime stage classification sub-model, and iteratively updates the model parameters based on this loss function. The calculation formula is as follows: ， in, This represents the cross-entropy loss function for multi-class classification. Indicates the category number, Indicates the total number of categories. Indicates an indicator function, This indicates the true lifespan stage category of the training samples. This represents the feature vector of the training sample. Indicates input Below, the prediction category is The conditional probability, This indicates taking the logarithm of the probability; S202 uses the mean squared error as the loss function for the XGBoost remaining lifetime regression sub-model, and iteratively updates the model parameters based on this loss function. The calculation formula is as follows: ， in, This represents the mean squared error loss function. This indicates the number of samples involved in the mean squared error calculation. Indicates the sample index. This represents the actual remaining number of operations for the i-th sample. This indicates that the i-th sample is in the input At that time, the regression model predicts the remaining number of operations. This represents the feature vector of the i-th sample; S203 constructs the solution vector of hyperparameters to be optimized for the pigeon flock optimization algorithm by using the model learning rate, maximum depth, minimum leaf node weight, subsample ratio, and number of iterations. S204 uses the pigeon flocking optimization algorithm to perform a phased global search and a landmark operator phased directional convergence optimization in the map-compass phased update mechanism of the pigeon flocking optimization algorithm on the hyperparameter solution vector to be optimized, and obtains the optimal hyperparameter solution vector corresponding to the model.

5. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 4, characterized in that, S204 includes: S2041 initializes the pigeon flock algorithm, maps the position of each pigeon to a candidate hyperparameter solution vector, and uses the candidate hyperparameter solution vectors to train the XGBoost life stage classification sub-model and the XGBoost remaining life regression sub-model respectively, to obtain the classifier and regressor corresponding to each candidate solution vector; In the current iteration, S2042 uses five-fold cross-validation to obtain the average accuracy of any candidate hyperparameter solution vector in the pigeon flock as its fitness value, and updates the historical best position corresponding to the candidate hyperparameter solution vector and the global best position of the group based on the fitness value. This process is repeated until the fitness evaluation of the current set of candidate hyperparameter solution vectors is completed. In the current iteration, S2043 uses five-fold cross-validation to obtain the reciprocal of the five-fold average RMSE of the XGBoost residual lifetime regression submodel corresponding to any candidate hyperparameter solution vector in the pigeon flock, and updates the historical best position corresponding to the candidate hyperparameter solution vector and the global best position of the group based on the fitness value. This process is repeated until the fitness evaluation of the current set of candidate hyperparameter solution vectors is completed. After completing the fitness evaluation of the current iteration round and updating the historical and global optimal positions, S2044 executes the map-compass stage update mechanism of the pigeon flock optimization algorithm. It updates the speed and position of each pigeon based on its historical optimal position and the global optimal position of the group, generates a new set of candidate hyperparameter solution vectors, and enters the next iteration round until the termination condition is met. S2045 When the termination condition is met, the candidate hyperparameter solution vector with the highest fitness value is selected as the optimal hyperparameter solution vector. The classification optimal hyperparameter solution vector is used to construct the XGBoost lifetime stage classification sub-model, and the regression optimal hyperparameter solution vector is used to construct the XGBoost remaining lifetime regression sub-model.

6. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 1, characterized in that, S3 includes: S301 inputs mechanical feature data into the optimized XGBoost life stage classification sub-model. The model calculates the score of the leaf node of each decision tree in turn, and outputs the predicted probability distribution of each category after being transformed by the Softmax function. The category with the highest probability is taken as the life stage prediction result. S302 inputs mechanical characteristic data into the optimized XGBoost remaining life regression sub-model and outputs a continuous estimate of the remaining number of operable operations. S303 trims the continuous estimate of the remaining number of operations, constraining it to the interval [0, ... Within this range, the final number of remaining operations is obtained, calculated using the following formula: ， in, This indicates the final number of operations remaining after trimming. This represents the function that takes the minimum value. This represents the function that takes the maximum value. This indicates the maximum rated mechanical life of the circuit breaker.

7. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 1, characterized in that, S4 includes: S401 calculates the entropy of the predicted probability distribution of the current sample as its prediction uncertainty through the XGBoost lifetime stage classification sub-model, and calculates the prediction standard deviation based on the integrated prediction results through the XGBoost remaining lifetime regression sub-model as its prediction uncertainty. S402 If the entropy of the predicted probability distribution is greater than the upper limit of entropy, the classification is judged to be low confidence and manual review is triggered; S403 If the entropy of the predicted probability distribution is less than or equal to the upper limit of entropy, obtain the remaining lifetime reference interval corresponding to the lifetime stage prediction result, calculate the absolute deviation of the remaining operable counts from the center of the remaining lifetime reference interval, and calculate its ratio to the prediction standard deviation. The calculation formula is as follows: ， in, This represents the ratio of absolute deviation to the predicted standard deviation. Indicates the remaining number of operations. Indicates the center of the remaining lifetime reference interval. Indicates the standard deviation of the forecast; S404 determines the consistency between two models based on the remaining number of operands and the ratio of absolute deviation to the prediction standard deviation. If the consistency determination is passed, it outputs the remaining number of operands and its lifetime stage. If the consistency determination fails in S405, the prediction uncertainty of the two models is obtained. If the prediction uncertainty of both models is higher or lower than their corresponding uncertainty threshold, manual review is triggered. S406 If only one model's prediction uncertainty is higher than the corresponding uncertainty threshold, extract the actual decision paths from the root node to the leaf node of each decision tree in the life stage classification sub-model and the remaining life regression sub-model for the current mechanical feature data, calculate the similarity of the decision paths of the two models, and output the model result with lower prediction uncertainty if the decision paths are similar.

8. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 7, characterized in that, S401 includes: For the XGBoost lifetime stage classification sub-model, S4011 uses the entropy of the predicted probability distribution as the prediction uncertainty, calculated as follows: ， in, Entropy represents the predicted probability distribution. Indicates the category number, Indicates the total number of categories. This represents the feature vector of the test sample to be predicted after preprocessing. Indicates input Below, the prediction category is The conditional probability, This indicates taking the logarithm of the probability; S4012 performs predictions on several XGBoost remaining lifetime regression sub-models constructed from different sub-samplings during the inference phase, obtaining several estimates of remaining operable times. The standard deviation of these estimates is calculated as the prediction uncertainty, as shown in the following formula: ， in, The standard deviation represents the number of remaining operands. This indicates the number of XGBoost remaining lifetime regression sub-models used for ensemble prediction. This represents the remaining number of operations predicted by the m-th XGBoost remaining lifetime regression sub-model.

9. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 8, characterized in that, S404 includes: S4041 If the remaining number of operable attempts is within the remaining lifetime reference range, the results of the two models are determined to be consistent, and the remaining number of operable attempts and their lifetime stage are output. S4042 If the remaining operable counts are outside the remaining lifetime reference range, and the ratio of the absolute deviation to the predicted standard deviation is greater than the ratio threshold, then the regression is determined to be low confidence and manual review is triggered. S4043 If the remaining operable counts are outside the remaining lifetime reference range, and the ratio of the absolute deviation to the prediction standard deviation is less than or equal to the ratio threshold, then the model results are considered inconsistent.

10. The method for predicting the remaining life of a high-voltage circuit breaker based on PIO-XGBoost according to claim 9, characterized in that, S406 includes: S4061 For the current mechanical feature data, for each decision tree in each sub-model, extract the internal nodes that the current mechanical feature data actually passes through from the root node to the leaf node, and obtain the actual decision path of the sample corresponding to the decision tree. S4062 records the split feature identifier, split operator, and split threshold to form a triplet for each internal node of the actual decision path; S4063 records the node layer number, starting with the root node having a layer number of 0, and increasing layer by layer along the direction from the root node to the leaf node. The depth weight of each triplet is calculated using the following formula: ， in, Indicates depth weights, Indicates the number of node layers; S4064 identifies all common triples in two decision paths that have the same feature identifier, the same splitting operator, and a splitting threshold within a preset tolerance range. For each common triple, it obtains the node layer number and depth weight of the corresponding node in the two decision paths, and selects the smaller depth weight as the matching weight. S4065 calculates the similarity score of decision paths based on the sum of the depth weights of the triples in the two decision paths and the sum of the matching weights of the common triples. If the similarity score is greater than the similarity threshold, the decision paths are considered similar. The calculation formula is as follows: ， in, This represents the sum of the matching weights of the common triples. This represents the sum of depth weights in the decision path of the XGBoost lifetime stage classification submodel. This represents the sum of depth weights in the decision path of the XGBoost remaining lifetime regression sub-model.

Images