Relay maintenance timing determination method and system considering explainability

Abstract

The application relates to the technical field of predictive maintenance, and particularly discloses a relay maintenance timing determination method and system considering explainability, wherein the relay maintenance timing determination method considering explainability comprises the following steps: acquiring performance index data of a relay to construct an initial confidence rule base model, applying an explainability constraint to the initial confidence rule base model to obtain a constrained confidence rule base model, adopting an optimized artificial bee colony algorithm to optimize parameters of the constrained confidence rule base model to obtain an optimized target confidence degree distribution, determining a performance state utility value of the relay according to the target confidence degree distribution, and calculating a quantitative maintenance timing of the relay based on the performance state utility value and by using a preset degradation mapping model; thus, the balance between evaluation accuracy and explainability is achieved in a small sample scene, and the relay can be used for pre-task performance optimization and whole life cycle state monitoring in various types of equipment.

Description

Technical Field

[0001] This invention belongs to the field of predictive maintenance technology, specifically relating to a method and system for determining relay maintenance timing that takes into account interpretability. Background Technology

[0002] Relays are key actuators in the electrical control and signal switching links of various equipment. The closing, releasing, and conducting states of their contacts directly affect the reliability of control command transmission. During equipment storage and use, accurately assessing the performance status of relays and determining scientific and reasonable maintenance times are crucial to ensuring the reliability of equipment tasks, preventing task failures due to contact degradation, and reducing the total life-cycle maintenance cost.

[0003] Currently, the BRB (Belief Rule Base) model is an effective means of relay performance evaluation and maintenance timing decision-making. The BRB model integrates domain expert knowledge and limited test data through a series of belief rules, enabling it to evaluate device performance under small sample conditions and provide corresponding maintenance timing suggestions based on the evaluation results. The BRB model is essentially a gray box model, and its reasoning process has natural transparency and traceability. It is a representative technical path that takes into account the advantages of knowledge and data in multi-source information hybrid driving methods.

[0004] However, traditional BRB models have a core problem in practical applications: during parameter optimization, the optimization algorithm blindly adjusts model parameters in pursuit of evaluation accuracy, which can easily lead to logical contradictions in the confidence distribution, causing the model to fundamentally lose its interpretability. For example, the optimized confidence distribution may exhibit a logical discontinuity of "either good or bad," which is neither in line with engineering practice nor with the laws of physical degradation, resulting in a lack of reliable decision-making basis for maintenance timing recommendations based on its output. At the same time, the limited number of relay test samples makes it difficult to support the modeling requirements of purely data-driven methods, and traditional BRB models lack effective constraint mechanisms to ensure their interpretability during parameter optimization, making it difficult for such methods to provide traceable and reliable maintenance timing decision support in practical engineering while ensuring evaluation accuracy. Summary of the Invention

[0005] This invention aims to at least address the technical problem of accurately determining the optimal maintenance timing for relays in related technologies. Therefore, the object of this invention is to provide a relay maintenance timing determination method, system, electronic device, and storage medium that considers interpretability.

[0006] In a first aspect, the present invention proposes a relay maintenance timing determination method considering interpretability, comprising the following steps: acquiring relay performance index data, wherein the performance index data includes pull-in voltage and pull-in time; constructing an initial confidence rule base model based on the performance index data, wherein the initial confidence rule base model includes multiple confidence rules, and the initial confidence rule base model is used to output the initial confidence distribution of the relay corresponding to each preset performance level based on the matching degree between the performance index data and preset performance levels; applying interpretability constraints to the initial confidence rule base model to obtain a constrained confidence rule base model, wherein the interpretability constraints include: setting the initial confidence distribution on parameters... The optimization process constrains logical consistency to ensure that the optimized confidence distribution conforms to a preset monotonic variation pattern or a preset unimodal distribution pattern. It also constrains the optimization range of rule weights and indicator weights in the initial confidence rule base model. An optimized artificial bee colony algorithm is used to optimize the parameters of the constrained confidence rule base model, resulting in an optimized target confidence distribution. This optimized artificial bee colony algorithm integrates multiple search strategies to enhance global search capabilities and convergence speed compared to the standard artificial bee colony algorithm. Based on the target confidence distribution, the performance state utility value of the relay is determined. Based on the performance state utility value, a preset degradation mapping model is used to calculate the quantitative maintenance timing of the relay.

[0007] According to an embodiment of the present invention, the relay maintenance timing determination method considering interpretability first acquires relay performance index data, then constructs an initial confidence rule base model based on the performance index data, next applies interpretability constraints to the initial confidence rule base model to obtain a constrained confidence rule base model, then uses an optimized artificial bee colony algorithm to optimize the parameters of the constrained confidence rule base model to obtain an optimized target confidence distribution, then determines the relay performance state utility value based on the target confidence distribution, and finally calculates the quantified maintenance timing of the relay based on the performance state utility value using a preset degradation mapping model; thus, by constructing a BRB model with triple interpretability constraints, the method addresses the issue from its root cause. This approach addresses the issue of interpretability loss during parameter optimization in traditional BRB models, achieving full traceability and interpretability of the evaluation and reasoning process. By integrating a multi-strategy search mechanism into the optimized artificial bee colony algorithm, it significantly improves the solution performance for non-convex optimization problems, with global optimization capability and convergence speed superior to standard artificial bee colony and particle swarm optimization algorithms. Furthermore, by establishing a degradation mapping model from performance utility values ​​to maintenance timing, it effectively connects performance evaluation and maintenance decisions, directly outputting quantified maintenance timing recommendations. Thus, it achieves a balance between evaluation accuracy and interpretability in small-sample scenarios, making it suitable for pre-mission performance optimization and full lifecycle status monitoring of relays in various equipment.

[0008] In addition, the relay maintenance timing determination method considering interpretability according to embodiments of the present invention may also have the following additional technical features: Furthermore, constraining the logical consistency of the initial confidence distribution during the parameter optimization process includes: limiting the optimized confidence distribution of each confidence rule to any one of a monotonically decreasing distribution, a monotonically increasing distribution, and a unimodal distribution; wherein, the monotonically decreasing distribution includes: the confidence corresponding to each preset performance level decreases sequentially from excellent to poor performance level; the monotonically increasing distribution includes: the confidence corresponding to each preset performance level increases sequentially from excellent to poor performance level; the unimodal distribution includes: the confidence distribution has a peak value, and the confidence values ​​on both sides of the peak value decrease sequentially.

[0009] Furthermore, the optimization range of the rule weights in the initial confidence rule base model is constrained, including: setting differentiated weight constraint intervals for each confidence rule based on the actual occurrence probability of the input combination of the performance index data corresponding to each confidence rule and the degree of influence on the task success rate; wherein, the input combination is a combination of pull-in voltage level and pull-in time level, and both the pull-in voltage level and pull-in time level include three levels: large, medium, and small, and the pull-in voltage level and pull-in time level are combined in pairs to form nine confidence rules; for confidence rules where both the pull-in voltage and the pull-in time are of medium level, the lower limit of the corresponding weight constraint interval is set to the highest level; for confidence rules where both the pull-in voltage and the pull-in time are of large level or both are of small level, the lower limit of the corresponding weight constraint interval is set to the second highest level; for confidence rules where one of the pull-in voltage and the pull-in time is of medium level and the other is of large level or small level, the lower limit of the corresponding weight constraint interval is set to the lowest level.

[0010] Furthermore, the index weights include pull-in voltage weights and pull-in time weights; the optimization range of the index weights in the initial confidence rule base model is constrained, including: based on the fact that the pull-in voltage has a higher capacity to characterize the state of the relay pull-in mechanism and the electromagnetic coil than the pull-in time has a higher capacity to characterize the contact action hysteresis, the lower limit of the pull-in voltage weight is set to be greater than the upper limit of the pull-in time weight.

[0011] Furthermore, in the optimized artificial bee colony algorithm, the leader bee and follower bees randomly select any one of the following strategies for position update during neighborhood search: standard difference search strategy, elite-guided search strategy, and optimal mutation search strategy. The standard difference search strategy includes: the current artificial bee randomly selects another individual in the neighborhood and updates its position using the position difference between the two individuals as a perturbation. The elite-guided search strategy includes: the current artificial bee uses the best individual in the current population as a guide direction and moves towards the region where the best individual is located to update its position. The optimal mutation search strategy includes: based on the position of the best individual in the current population, superimposing the position difference between two randomly selected other individuals as a perturbation to generate a new position.

[0012] Furthermore, after acquiring the relay performance index data and before constructing the initial confidence rule base model based on the performance index data, the method further includes: using an overlapping sliding window algorithm to augment the original test samples of the performance index data, including: truncating subsequences on the original sample sequence by sliding a window with a preset step size, and taking each subsequence as a new sample to obtain a sample dataset for training and testing the constraint confidence rule base model.

[0013] Furthermore, the degradation mapping model is an inverse Gaussian degradation process model, wherein the cumulative degradation mean of the inverse Gaussian degradation process model changes with time according to a power function law; the step of calculating the quantitative maintenance timing of the relay based on the performance state utility value using the preset degradation mapping model includes: linearly normalizing the performance state utility value to map it to the degradation domain to obtain the current relative degradation amount; calculating the conditional reliability function of the relay based on the current relative degradation amount using the inverse Gaussian degradation process model, wherein the conditional reliability function is used to characterize the probability that the relay will still maintain qualified performance after a specified remaining time, given that it has degraded to the current relative degradation amount; using a preset reliability threshold as the decision boundary, solving for the remaining time required for the conditional reliability function to decrease to the reliability threshold, and back-calculating the quantitative maintenance timing.

[0014] Furthermore, the relay performance state represented by the performance state utility value is positively correlated with the value of the performance state utility value; the current relative degradation is obtained by performing inverse linear normalization on the performance state utility value, the current relative degradation is negatively correlated with the performance state utility value, and the current relative degradation is positively correlated with the current degree of degradation of the relay.

[0015] Further, determining the performance state utility value of the relay based on the target confidence distribution includes: obtaining a fixed utility value corresponding to each preset performance level, wherein the fixed utility value corresponds one-to-one with the preset performance level, and the fixed utility value increases as the preset performance level increases; and using the confidence level of each performance level in the target confidence distribution as a weight, performing a weighted summation with the corresponding fixed utility value to obtain the performance state utility value.

[0016] Secondly, based on the same inventive concept, this invention also proposes a relay maintenance timing determination system considering interpretability, comprising: a data acquisition module for acquiring relay performance index data, wherein the performance index data includes pull-in voltage and pull-in time; an initial model construction module for constructing an initial confidence rule base model based on the performance index data, wherein the initial confidence rule base model includes multiple confidence rules, and the initial confidence rule base model is used to output the initial confidence distribution of the relay corresponding to each preset performance level based on the matching degree between the performance index data and preset performance levels; and a constraint model construction module for applying interpretability constraints to the initial confidence rule base model to obtain a constraint confidence rule base model, wherein the interpretability constraints include: the initial confidence distribution... The logical consistency of the parameter optimization process is constrained to ensure that the optimized confidence distribution conforms to a preset monotonic variation law or a preset unimodal distribution law. Furthermore, the optimization range of rule weights and indicator weights in the initial confidence rule base model is constrained. A parameter optimization module is used to optimize the parameters of the constrained confidence rule base model using an optimized artificial bee colony algorithm to obtain the optimized target confidence distribution. This optimized artificial bee colony algorithm integrates multiple search strategies to enhance global search capabilities and convergence speed based on the standard artificial bee colony algorithm. A utility value determination module is used to determine the performance state utility value of the relay based on the target confidence distribution. A maintenance timing calculation module is used to calculate the quantitative maintenance timing of the relay based on the performance state utility value using a preset degradation mapping model.

[0017] The relay maintenance timing determination system considering interpretability provided in the second aspect has the same technical features as the relay maintenance timing determination method considering interpretability provided in the first aspect, and therefore also has similar technical effects as the first aspect, which will not be described in detail here.

[0018] Thirdly, based on the same inventive concept, the present invention also proposes an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store a computer program; and the processor, when executing the computer program stored in the memory, implements the steps of the relay maintenance timing determination method considering interpretability as described in the above embodiments of the present invention.

[0019] The electronic device provided in the third aspect has the same technical features as the relay maintenance timing determination method considering interpretability provided in the first aspect, and therefore also has similar technical effects as the first aspect, which will not be described in detail here.

[0020] Fourthly, based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the relay maintenance timing determination method considering interpretability as described in the above embodiments of the present invention.

[0021] The computer-readable storage medium provided in the fourth aspect has the same technical features as the relay maintenance timing determination method considering interpretability provided in the first aspect, and therefore also has similar technical effects as the first aspect, which will not be described in detail here. Attached Figure Description

[0022] Figure 1 This is a flowchart of a relay maintenance timing determination method considering interpretability according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the mapping relationship between performance utility values ​​and maintenance timing according to a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the constrained BRB model optimization process according to a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the optimization process of a conventional BRB model according to a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 1; Figure 6 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 2; Figure 7 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 3; Figure 8 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 4; Figure 9 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 5; Figure 10 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 6; Figure 11 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 7; Figure 12 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 8; Figure 13 This is a schematic diagram of the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 9; Figure 14 This is a schematic diagram of the relay evaluation test results according to a specific embodiment of the present invention; Figure 15 This is a schematic diagram of the relay maintenance timing assessment results according to a specific embodiment of the present invention; Figure 16 This is a schematic diagram illustrating the evaluation error of utility values ​​according to a specific embodiment of the present invention; Figure 17 This is a block diagram of a relay maintenance timing determination system that considers interpretability according to an embodiment of the present invention; Figure 18 This is a structural block diagram of an electronic device according to an embodiment of the present invention.

[0023] Figure label: 100 - Relay maintenance timing determination system considering interpretability; 110 - Data acquisition module; 120 - Initial model construction module; 130 - Constraint model construction module; 140 - Parameter optimization module; 150 - Utility value determination module; 160 - Maintenance timing calculation module. Detailed Implementation

[0024] The embodiments of the present invention are described in detail below, and the embodiments described with reference to the accompanying drawings are exemplary.

[0025] To address the problem of accurately determining the optimal maintenance time for relays in related technologies, this invention provides a relay maintenance timing determination method, system, electronic device, and storage medium that considers interpretability. The following references... Figures 1-18 A method, system, electronic device, and storage medium for determining relay maintenance timing with consideration for interpretability are described according to embodiments of the present invention.

[0026] In the embodiments of the present invention, the relays involved are, for example, industrial airtight sealed electromagnetic relays (sealed electromagnetic relays). The following embodiments are all described using sealed electromagnetic relays, and will not be repeated hereafter. It is understood that the relays involved in the present invention are not limited to or applicable to a single relay.

[0027] Figure 1 This is a flowchart of a relay maintenance timing determination method considering interpretability according to an embodiment of the present invention. Figure 1 As shown, a relay maintenance timing determination method considering interpretability according to an embodiment of the present invention includes the following steps: Step S1: Obtain the relay's performance data, including the pull-in voltage and pull-in time.

[0028] In a specific embodiment, a relay comprehensive parameter testing system can be used to collect the relay's PV (Pull-in Voltage) and PT (Pull-in Time). Pull-in voltage reflects the state of the relay's pull-in mechanism and electromagnetic coil, while pull-in time reflects the contact action hysteresis and mechanism response characteristics. These two indicators are two core features characterizing the relay's fault state, and relevant parameter definitions and performance labels can be obtained through expert knowledge. At the same time, an overlapping sliding window algorithm can be used to augment the original test samples, meeting the model training requirements while preserving the data distribution characteristics and physical laws. Specifically, subsequences can be extracted from the original sample sequence according to a preset step size sliding window, with each subsequence treated as a new sample, to obtain a sample dataset for training and testing subsequent models.

[0029] Specifically, this embodiment of the invention first acquires the performance index data of the relay, which specifically includes the pull-in voltage and pull-in time. These two indicators can comprehensively characterize the performance state of the relay from two dimensions: the state of the electromagnetic coil and the hysteresis of the contact action, providing key quantitative input for subsequent evaluation. At the same time, the sample is expanded by using an overlapping sliding window algorithm, which effectively alleviates the data scarcity problem in the small sample scenario of relays without changing the physical distribution characteristics of the original data, and provides sufficient sample support for the training of the subsequent constraint confidence rule base model.

[0030] Step S2: Construct an initial confidence rule base model based on the performance index data. The initial confidence rule base model includes multiple confidence rules. The initial confidence rule base model is used to output the initial confidence distribution of the relay corresponding to each preset performance level based on the matching degree between the performance index data and the preset performance level.

[0031] In a specific embodiment, relay performance can be divided into four levels: Excellent (E), Good (G), Medium (M), and Poor (P) as model outputs. Pull-in voltage and pull-in time can each be divided into three evaluation levels: High (D), Medium (Z), and Low (X). These indicators are combined to construct nine If-Then confidence rules. An initial BRB (Belief Rule Base) model containing these nine confidence rules can be constructed by incorporating domain expert knowledge, completing the quantitative embedding and rule-based expression of expert knowledge. In the initial BRB model, the confidence distribution of each rule can be set based on expert experience, and the rule weights and indicator weights can also be assigned initial values ​​based on actual engineering conditions. Specifically, for example, the initial weight of pull-in voltage can be set to 0.55, and the initial weight of pull-in time can be set to 0.45.

[0032] Specifically, this embodiment of the invention constructs an initial confidence rule base model based on the acquired performance index data. The initial confidence rule base model transforms fuzzy expert knowledge into structured knowledge representation through If-Then confidence rules, effectively utilizing expert experience to compensate for insufficient data under small sample conditions. At the same time, the BRB model is essentially a white-box or gray-box model, and its reasoning process has natural transparency and traceability, laying the model foundation for subsequent interpretable maintenance timing decisions. By dividing the input indicators into three levels and combining them into nine rules, the rule base is ensured to cover various input combinations, thereby guaranteeing a comprehensive evaluation capability for devices in different performance states.

[0033] Step S3: Apply interpretability constraints to the initial confidence rule base model to obtain a constrained confidence rule base model. The interpretability constraints include: constraining the logical consistency of the initial confidence distribution during the parameter optimization process so that the optimized confidence distribution conforms to a preset monotonic change law or a preset unimodal distribution law; and constraining the optimization range of rule weights and indicator weights in the initial confidence rule base model.

[0034] In a specific embodiment, the parameters to be optimized in the BRB model include rule weights, index weights, and evaluation level confidence. To address the problem of lost interpretability caused by logical contradictions in the confidence distribution during the optimization of the traditional BRB model, three types of interpretability constraints that conform to engineering cognition can be introduced to limit the confidence distribution of each rule to only present three forms: (1) Monotonically decreasing distribution: the confidence decreases sequentially from excellent to poor with the performance level, corresponding to the scenario where the input index performance is excellent, and the relay performance is excellent with high confidence; (2) Monotonically increasing distribution: the confidence increases sequentially from excellent to poor with the performance level, corresponding to the scenario where the input index performance is deteriorated, and the relay performance is poor with high confidence; (3) Unimodal distribution: the confidence distribution has only one peak, and the values ​​on both sides of the peak decrease sequentially, and the relay performance is good or average with high confidence.

[0035] Meanwhile, to address the issue of limited test samples and the difficulty in iterating optimization algorithms to achieve optimal results, constraints can be added to the weights of the input indicator data and the rule weights of each If-Then confidence rule, based on expert knowledge, performance parameters, and the actual impact of the rules. Specifically, since the pull-in voltage's ability to characterize the relay pull-in mechanism and electromagnetic coil state is higher than the pull-in time's ability to characterize contact action hysteresis, the lower limit of the pull-in voltage weight can be set higher than the upper limit of the pull-in time weight. Based on the actual probability of occurrence of the input combinations corresponding to each confidence rule and their impact on the task success rate, differentiated weight constraint intervals can be set for each confidence rule. For example, for confidence rules where both pull-in voltage and pull-in time are of medium level, the lower limit of its weight constraint interval can be set to the highest level; for confidence rules with two extreme indicators (both large or both small), the lower limit of its weight constraint interval can be set to the second highest level; and for confidence rules with a single extreme indicator or a single medium indicator, the lower limit of its weight constraint interval can be set to the lowest level.

[0036] Specifically, this embodiment of the invention then applies interpretability constraints to the initial confidence rule base model to obtain a constrained confidence rule base model. By applying three morphological constraints conforming to physical degradation laws to the confidence distribution, it fundamentally eliminates logical contradictions such as either excellent or poor confidence distributions during the optimization process, ensuring that the confidence distribution of each rule conforms to engineering logic throughout the parameter optimization process, and achieving full traceability and interpretability of the evaluation reasoning process. By applying differentiated constraints conforming to engineering cognition to the indicator weights, namely, the lower limit of the pull-in voltage weight is higher than the upper limit of the pull-in time weight, it ensures that the importance ranking of the two indicators during the optimization process is consistent. Ultimately, it conforms to physical reality, avoiding the algorithm from reversing the primary and secondary relationships of indicators due to blindly pursuing accuracy. By applying differentiated constraint intervals to the rule weights based on the probability of input combination occurrence and task importance, it limits the reasonable optimization range of the weights while retaining the core experience of expert knowledge, significantly reducing the iterative search space of the optimization algorithm, improving optimization efficiency, and ensuring that the rule weights still have clear physical meaning after optimization. By introducing triple interpretability constraints, the constrained BRB model always operates within the engineering interpretability boundary throughout the parameter optimization process, fundamentally solving the technical problem of the traditional model's inability to achieve both accuracy and interpretability.

[0037] Step S4: The optimized artificial bee colony algorithm is used to optimize the parameters of the constraint confidence rule base model to obtain the optimized target confidence distribution. The optimized artificial bee colony algorithm integrates multiple search strategies to enhance global search capabilities and convergence speed on the basis of the standard artificial bee colony algorithm.

[0038] In a specific embodiment, all parameters to be optimized in the constrained BRB model (including rule weights, index weights, and evaluation level confidence) can be integrated into a unified set, and the location of the artificial bee can be encoded and mapped accordingly. This transforms the parameter optimization problem into an intelligent search process using the ABC (Artificial Bee Colony) algorithm. To address the issues of low search efficiency and premature convergence in the standard ABC algorithm, this method considers the elite guidance strategy of the particle swarm optimization algorithm and the optimal mutation strategy of the differential evolution algorithm. Two new search strategies are added to the standard ABC algorithm, constructing a search strategy pool containing three parallel strategies. When the leader bee and the follower bee search their neighborhood, they randomly select any one of the standard differential search strategy, the elite guidance search strategy, and the optimal mutation search strategy to update their positions. Specifically, the standard difference search strategy involves the current artificial bee randomly selecting another individual in its neighborhood and using the position difference between the two individuals as a perturbation to update its position; the elite-guided search strategy involves the current artificial bee using the best individual in the current population as a guide and moving towards the region where the best individual is located to update its position; and the optimal mutation search strategy involves generating a new position based on the position of the best individual in the current population and superimposing the position difference between two randomly selected other individuals as a perturbation.

[0039] Meanwhile, the optimization process aims to minimize the prediction error, using MSE (Mean Square Error) as the objective function, and retains the local extremum escape mechanism of the scout bee. When the number of optimization attempts of an individual reaches the preset limit and the fitness does not improve, the scout bee is switched to reset the site to avoid local optima.

[0040] Specifically, this invention employs an optimized artificial bee colony algorithm to optimize the parameters of a constrained confidence rule base model. For the non-convex optimization problem of BRB parameters with interpretability constraints, its solution space exhibits complex characteristics of multi-peak and nonlinearity, with numerous local optima. Traditional gradient-based optimization algorithms and standard particle swarm optimization algorithms are prone to getting trapped in local optima. The optimized artificial bee colony algorithm used in this invention effectively enhances population diversity through random switching of three parallel search strategies: standard difference search ensures basic exploration capabilities, elite-guided search accelerates convergence to the optimal region, and optimal mutation search takes into account the refined development of the optimal solution. This multi-strategy parallel mechanism effectively overcomes the shortcomings of the standard artificial bee colony algorithm, such as its singular search method and insufficient population diversity, significantly improving the algorithm's global search capability and convergence rate. Simultaneously, by combining the applied weight constraint range, the search space of the parameters to be optimized is significantly compressed, further improving the algorithm's iterative efficiency. By optimizing the parameters of the constrained BRB model using this optimized artificial bee colony algorithm, the solution can efficiently approximate the globally optimal parameter combination while ensuring that all interpretability constraints are met, thus achieving optimal model evaluation accuracy.

[0041] Step S5: Determine the performance state utility value of the relay based on the target confidence distribution.

[0042] In a specific embodiment, a fixed utility value corresponding to each preset performance level (excellent, good, medium, poor) can be obtained. The fixed utility value corresponds one-to-one with the preset performance level. For example, excellent corresponds to a utility value of 4, good corresponds to a utility value of 3, medium corresponds to a utility value of 2, and poor corresponds to a utility value of 1. The fixed utility value increases as the preset performance level increases. The confidence level of each performance level in the optimized target confidence distribution is used as a weight, and the corresponding fixed utility value is weighted and summed to obtain the performance status utility value. The performance status utility value range is, for example, [1, 4]. The higher the value, the better the overall performance status of the device.

[0043] Specifically, the embodiments of the present invention then determine the performance status utility value of the relay based on the optimized target confidence distribution. By compressing multi-dimensional, distributed confidence information into a single utility value scalar, it is not only convenient to intuitively compare and rank the performance status of different devices, but more importantly, it transforms the abstract confidence distribution into a quantitative indicator with clear physical meaning and engineering cognitive basis, providing a unified input interface for the quantitative calculation of subsequent maintenance timing, thereby helping to achieve a smooth connection from performance evaluation to maintenance decision-making.

[0044] Step S6: Based on the performance status utility value, calculate the quantitative maintenance timing of the relay using a preset degradation mapping model.

[0045] In a specific embodiment, the degradation mapping model is, for example, an inverse Gaussian degradation process model. The cumulative degradation mean of the inverse Gaussian degradation process model changes with time according to a power function. First, the performance state utility value can be linearly normalized and mapped to the degradation domain to obtain the current relative degradation value. The current relative degradation value is negatively correlated with the performance state utility value; the lower the utility value, the deeper the current degradation of the corresponding device. Then, based on the current relative degradation value, the conditional reliability function of the relay can be calculated using the inverse Gaussian degradation process model. This conditional reliability function is used to characterize the probability that the relay will still maintain acceptable performance after a specified remaining time, given that it has degraded to the current relative degradation value. Finally, using a preset reliability threshold (an engineering decision parameter representing the minimum acceptable mission reliability of the equipment) as the decision boundary, the remaining time required for the conditional reliability function to decrease to the reliability threshold is calculated, and the quantitative maintenance timing is obtained by reverse calculation. The unit of the quantitative maintenance timing is, for example, days.

[0046] Specifically, in this embodiment of the invention, the quantitative maintenance timing of the relay is calculated based on the performance state utility value using a preset degradation mapping model. An inverse Gaussian degradation process model is established to describe the stochastic evolution of the relay's performance degradation. The performance utility value is used as a quantitative representation of the current degradation state and input into the degradation mapping model. Then, the optimal maintenance timing is derived using a conditional reliability function and a preset reliability threshold, achieving a precise quantitative mapping from performance evaluation results to maintenance timing. This ensures that the determination of maintenance timing no longer relies on manual experience or fixed cycles, but is based on scientific quantitative calculations of the device's current actual performance state and reliability degradation patterns. This significantly reduces the cost and subjectivity of maintenance decisions in engineering, effectively avoiding over-maintenance or under-maintenance problems.

[0047] Therefore, the relay maintenance timing determination method considering interpretability according to embodiments of the present invention first acquires relay performance index data, then constructs an initial confidence rule base model based on the performance index data, next applies interpretability constraints to the initial confidence rule base model to obtain a constrained confidence rule base model, then uses an optimized artificial bee colony algorithm to optimize the parameters of the constrained confidence rule base model to obtain an optimized target confidence distribution, then determines the relay performance state utility value based on the target confidence distribution, and finally calculates the quantified maintenance timing of the relay based on the performance state utility value using a preset degradation mapping model; thus, by constructing a BRB model with triple interpretability constraints, from It fundamentally solves the problem of lost interpretability during the parameter optimization process of traditional BRB models, achieving full traceability and interpretability of the evaluation and reasoning process. By integrating a multi-strategy search mechanism into the optimized artificial bee colony algorithm, it significantly improves the solution effect of non-convex optimization problems, with global optimization capability and convergence speed superior to standard artificial bee colony and particle swarm algorithms. By establishing a degradation mapping model from performance utility value to maintenance timing, it achieves effective connection between performance evaluation and maintenance decision-making, and can directly output quantitative maintenance timing suggestions. Thus, it achieves a balance between evaluation accuracy and interpretability in small sample scenarios, and is suitable for pre-mission performance optimization and full life cycle status monitoring of relays in various equipment.

[0048] In one embodiment of the present invention, step S2 constructs an initial confidence rule base model based on performance index data, including: dividing the pull-in voltage and pull-in time into three levels: large, medium, and small; dividing the overall performance of the relay into four levels: excellent, good, medium, and poor; constructing nine If-Then confidence rules based on the pairwise combinations of the pull-in voltage level and the pull-in time level; and setting the initial confidence level distribution, rule weight, and index weight for each confidence rule by combining domain expert knowledge, thereby obtaining the initial confidence rule base model.

[0049] In specific embodiments, there are many indicators reflecting relay performance, such as pull-in voltage, release voltage, pull-in time, release time, contact resistance, and coil resistance. Among them, pull-in voltage (PV) reflects the state of the relay pull-in mechanism and electromagnetic coil, and pull-in time (PT) reflects the contact action hysteresis and mechanism response characteristics. These are two core characteristics characterizing the relay fault state, and relevant parameters and performance labels can be obtained through expert knowledge. Therefore, this method selects pull-in voltage (PV) and pull-in time (PT) as the performance indicators for relays, and the specific level classification is shown in Table 1.

[0050] Table 1 Reference values ​​for each level of PV and PT

[0051] Within the aforementioned evaluation model framework, two core performance indicators are selected for each relay, and each indicator is divided into three evaluation levels: High (D), Medium (Z), and Low (X). These indicators are combined to construct nine inference rules. Simultaneously, the overall performance of the relay is divided into four evaluation levels: Excellent (E), Good (G), Medium (M), and Poor (P). To facilitate subsequent quantitative calculations and maintenance timing mapping, the four performance levels are assigned fixed utility values: Poor corresponds to utility value 1, Medium to utility value 2, Good to utility value 3, and Excellent to utility value 4. Based on expert experience, an initial correspondence between performance utility values ​​and optimal maintenance timing is established. The worse the performance, the shorter the maintenance cycle; the better the performance, the longer the maintenance cycle. Specific level divisions, utility assignments, and maintenance timing correspondence standards are shown in Table 2.

[0052] Table 2 Correspondence between Relay Performance Status and Maintenance Timing

[0053] The BRB model is a knowledge representation and evidence reasoning model composed of a series of If-Then rules. It can effectively integrate fuzzy expert knowledge and quantitative test data. k Rules R k The general form of expression is: In the formula, Indicates relay M Each input metric t The value at time, in this specific embodiment M =2, which correspond to the pull-in voltage and pull-in time, respectively; Indicates the first k In the rules M Each input indicator corresponds to a reference value, and each rule is composed of a combination of reference levels for each input indicator. All rules cover all combinations of input levels. This indicates the performance evaluation results of the relay; Indicates relay N Performance levels, in this specific embodiment N =4, corresponding to four levels: Excellent (E), Good (G), Average (M), and Poor (P); express R k middle N The confidence level of each performance level, and at the same time, using To determine the rules k The integrity, when When this happens, the rule is called an incomplete rule, meaning that there is unknown information that makes it impossible to assign all confidence levels to all conclusion levels; express R k The rule weight is used to measure the relative importance of the rule in the entire rule base, and its value ranges from (0,1]. Indicates relay M The indicator weights of each input indicator are used to measure the relative importance of different input indicators to the final evaluation result, and the values ​​range from (0,1).

[0054] Then, based on the inputs, outputs, and level classifications of this example, a complete confidence rule base can be obtained: In the formula, Indicates a relay. Indicates the rule number. This represents the confidence level under the corresponding performance evaluation. This represents the rule weights of the 9 rules. Indicates input metrics PV , PT The weight of the indicators.

[0055] Finally, the initial BRB model for relay performance evaluation, obtained by combining the aforementioned confidence rule base and based on expert knowledge, is shown in Table 3.

[0056] Table 3 Initial BRB Model Diagram

[0057] Based on the initial BRB model constructed using the above steps, the final performance evaluation result and maintenance timing can be obtained for any type of input information through the following reasoning process: 1) Individual matching degree calculation, that is, the conversion of input information into reference status information.

[0058] In the BRB model, to achieve the actual input value from the relay to its reference state set The transformation involves establishing a match function to achieve the following function mapping relationship: In the formula, It is a relay input indicator The A reference state, For input values Relative to reference state Match degree. and These are the number of attributes of the input metric and the number of reference states, respectively. A matching degree function is defined; to accommodate input attributes with different distribution characteristics, various matching degree functions have been proposed, and this specific embodiment adopts the triangular matching degree function. The input value can be obtained using the following formula. Relative to the reference state set Match degree: In the formula, for Compared to The degree of matching, This represents the matching degree distribution.

[0059] 2) Activation weight solution, i.e., activation of confidence rules.

[0060] Based on membership distribution, combined with input indicator weights , can be obtained With the k The matching degree of the rule is: In the formula, It is the first k Rule number 1 The degree of matching of the reference state of each input indicator; combined with rule weights. After normalization, the activation level of each rule can be obtained: In the formula, yes Compared to The activation weight represents the activation weight of the device. Of all the rules that are activated The relative importance of.

[0061] 3) Fusion of confidence rules.

[0062] After calculating the activation weights, the weight values ​​of the corresponding rules are obtained. Then, the confidence scores of all rules are synthesized using the Evidential Reasoning Algorithm (ERA) to obtain the final confidence score. In the formula, For each rule, the conclusion level is determined. j The joint confidence level, V The confidence level that an inactive rule has an impact on it.

[0063] To address the discrepancy between evaluation results and actual values ​​caused by the interference of inactive rules on the conclusion level, after calculating the confidence level of the impact of inactive rules on the performance level, a normalization factor also needs to be calculated. To ensure that the sum of the confidence levels of the rules output after fusion is 1: .

[0064] After the fusion of all rules is completed, the confidence scores of inactive rules that affect the conclusion need to be further subtracted to obtain the final confidence scores for each level: In the formula, For the first j Each level of confidence in the results.

[0065] Furthermore, for the convenience of subsequent interpretable performance evaluation of the relays, the confidence level of the results is converted into a utility value form for evaluation. The corresponding utility value is For reference, the output function of the BRB model S(x) Expected utility value for: In the formula, It is the output compared to the evaluation value. The lower bound of the likelihood is, and the upper bound of the likelihood is. In the formula, The remaining confidence level that has not been assigned to any particular level.

[0066] Furthermore, considering the possibility of incomplete or inaccurate evaluation values, three utility value parameters are defined, namely: maximum utility value. Minimum utility value With average utility value : .

[0067] After obtaining the final utility value, a nonlinear mapping relationship between the performance utility value and the optimal maintenance timing is established based on the inverse Gaussian stochastic process model of relay performance degradation: In the formula, This indicates that the relay starts from its initial healthy state, operates, and stores data until a certain time. t The performance reliability at that time, i.e. t The probability that the device performance remains acceptable at any given time; The cumulative degradation mean function represents the inverse Gaussian degradation process, describing the mathematical expectation of device performance degradation over time, and the whole function follows a power function form. , The shape characteristic parameters representing the inverse Gaussian degradation process are calibrated using historical performance degradation test data of the relay, and respectively determine the baseline rate and the degree of nonlinearity in the time dimension of the degradation process; The diffusion parameter represents the inverse Gaussian degradation process, characterizing the random fluctuation amplitude of the degradation process, and is also statistically calibrated from the historical test data of the device; The performance failure threshold of a relay is the critical value of the amount of degradation corresponding to the performance degradation of the device to an unqualified state, which is determined by the equipment specifications and usage requirements. The cumulative distribution function represents the standard normal distribution and is the fundamental function for calculating the probability of the inverse Gaussian distribution; This represents the performance utility value output by the BRB model, with a value range of [1,4]. A higher value indicates a better overall performance of the device. Performance utility value The corresponding current relative degradation is converted to the degradation domain through a linear normalization mapping. The lower the utility value, the deeper the current degradation of the corresponding device. This represents the conditional reliability function, indicating that the device has now degraded to a certain level. Under the premise that, the remaining time t The probability of maintaining acceptable performance remains, and its functional form is consistent with the aforementioned inverse Gaussian reliability function, except that the initial degradation amount is replaced with the current value. ; This represents the reliability threshold for conditional maintenance. It is a preset engineering decision parameter that represents the minimum acceptable mission reliability of the equipment. Maintenance is required when the remaining reliability drops to this threshold. Represents the corresponding performance utility value The optimal maintenance time, measured in days, is the remaining time when the remaining reliability just drops to the maintenance threshold, obtained by inverse solving the conditional reliability equation.

[0068] Figure 2 This is a schematic diagram illustrating the mapping relationship between performance utility values ​​and maintenance timing according to a specific embodiment of the present invention. The horizontal axis represents the performance utility value, and the vertical axis represents the optimal maintenance timing. Figure 2 As shown, this illustrates the mapping relationship between specific performance utility values ​​and maintenance timing.

[0069] Specifically, this embodiment of the invention divides the pull-in voltage and pull-in time into three levels and combines them to form nine confidence rules, ensuring that the rule base can cover all possible input combinations and guaranteeing a comprehensive evaluation capability for devices in different performance states. By dividing the overall performance of the relay into four levels and assigning fixed utility values, a unified numerical benchmark is established for subsequent quantitative calculations and mappings. By combining expert knowledge to construct an initial BRB model, it can effectively utilize prior knowledge of the domain to compensate for the lack of data under limited sample conditions, providing an initial solution that conforms to engineering practice and has clear physical meaning for subsequent model optimization, avoiding the instability of pure data-driven methods during cold starts.

[0070] In one embodiment of the present invention, the logical consistency of the initial confidence distribution during the parameter optimization process is constrained, including: limiting the optimized confidence distribution of each confidence rule to any one of a monotonically decreasing distribution, a monotonically increasing distribution, and a unimodal distribution; wherein, the monotonically decreasing distribution includes: the confidence corresponding to each preset performance level decreases sequentially from excellent to poor performance level; the monotonically increasing distribution includes: the confidence corresponding to each preset performance level increases sequentially from excellent to poor performance level; the unimodal distribution includes: the confidence distribution has a peak value, and the confidence values ​​on both sides of the peak value decrease sequentially.

[0071] In a specific implementation, the confidence level of BRB is limited during the optimization process. To maintain a unimodal, monotonically decreasing, and monotonically increasing confidence distribution, the above three confidence distributions can be achieved through the following constraints: In the formula, The set of interpretable confidence distributions includes three cases: monotonically decreasing, monotonically increasing, and unimodal decreasing.

[0072] Specifically, this invention, by imposing three types of morphological constraints on the confidence distribution that conform to the laws of physical degradation, fundamentally eliminates logically contradictory confidence distributions during the optimization process. Relay performance degradation is a continuous and directional physical process: when input indicators are excellent, device performance should be rated as excellent or good with a high probability; when input indicators deteriorate, device performance should be rated as average or poor with a high probability; when input indicators are at a moderate level, the confidence level of device performance should reach its peak at the intermediate level. These three distributions precisely correspond to the three scenarios mentioned above that conform to engineering understanding. By embedding these three types of constraints into the parameter optimization process, it ensures that each rule maintains logical consistency in the optimized confidence distribution, making the entire evaluation and reasoning process traceable and interpretable. This effectively solves the problem of lost interpretability in traditional BRB models after parameter optimization, where the result is either "excellent or poor."

[0073] In one embodiment of the present invention, the optimization range of the rule weights in the initial confidence rule base model is constrained, including: setting differentiated weight constraint intervals for each confidence rule based on the actual occurrence probability of the input combination of the performance index data corresponding to each confidence rule and the degree of influence on the task success rate; wherein, the input combination is a combination of the pull-in voltage level and the pull-in time level, and both the pull-in voltage level and the pull-in time level include three levels: large, medium, and small, and the pull-in voltage level and the pull-in time level are combined in pairs to form nine confidence rules; for confidence rules where both the pull-in voltage and the pull-in time are of medium level, the lower limit of the corresponding weight constraint interval is set to the highest level; for confidence rules where both the pull-in voltage and the pull-in time are of large level or both are of small level, the lower limit of the corresponding weight constraint interval is set to the second highest level; for confidence rules where one of the pull-in voltage and the pull-in time is of medium level and the other is of large level or small level, the lower limit of the corresponding weight constraint interval is set to the lowest level.

[0074] In specific embodiments, considering the manufacturing process characteristics and performance distribution patterns of relays, the probability of occurrence of rules corresponding to different input combinations varies significantly. The moderate performance of dual indicators is the most common performance state of mass-produced relays, corresponding to the highest reliability of rule 5. The combination of single extreme indicator and single moderate indicator has the next highest probability of occurrence. The extreme performance of dual indicators corresponds to the worst and best performance. Although the probability of occurrence in mass production is low, it has a great impact on the success or failure of the task, and its rule weight cannot be too low. At the same time, it should also be considered that under the same degree of extremeness, the weight of the input indicator pull-in voltage is slightly higher than that of the pull-in time, and the lower limit of the rule weight that dominates it should also be higher.

[0075] Based on the above performance distribution characteristics, and while retaining the core experience of expert weight setting, the following differentiated constraints are applied to the weights of the nine inference rules: In the formula, These are the rule weights of the 9 BR (Belief Rule) rules, and the detailed constraints are shown in Table 4.

[0076] Table 4. Relay BRB Model Rule Weight Constraint System

[0077] Specifically, this embodiment of the invention sets differentiated weight constraint ranges for each rule, thus limiting the reasonable optimization range of the weights while preserving the core experience of expert knowledge. Rules with moderate dual indicators correspond to the peak performance distribution of mass-produced devices, have the largest sample size, and the highest prior confidence, therefore they are assigned the highest weight lower limit to reflect their statistical credibility. Rules with extreme dual indicators, although having a low probability of occurrence, have a significant impact on the success or failure of the task, and are assigned the second highest weight lower limit to ensure that they are not underestimated in key decisions. Rules with extreme single indicators or moderate single indicators are assigned relatively low weight lower limits, allowing the algorithm to adjust their importance based on data within a certain range. This differentiated constraint design not only preserves the expert's prior judgment on the importance of different rules, but also limits the reasonable search space of the weight parameters, significantly reducing the iteration range of the optimization algorithm, improving optimization efficiency, and ensuring that the rule weights still have clear physical meaning and engineering interpretability after optimization.

[0078] In one embodiment of the present invention, the index weights include pull-in voltage weights and pull-in time weights; the optimization range of the index weights in the initial confidence rule base model is constrained, including: based on the fact that the pull-in voltage has a higher capacity to characterize the state of the relay pull-in mechanism and the electromagnetic coil than the pull-in time has a higher capacity to characterize the contact action hysteresis, the lower limit of the pull-in voltage weight is set to be greater than the upper limit of the pull-in time weight.

[0079] In a specific embodiment, based on expert knowledge and engineering practice, in relay performance evaluation, the pull-in voltage reflects the pull-in conditions of the electromagnetic coil, spring mechanism, and contacts, and is a fundamental indicator for judging whether the relay can operate reliably; the pull-in time reflects the contact action response speed and mechanism hysteresis characteristics, and has a significant impact on the timeliness of control commands during the task; considering both the relay's working mechanism and the stability of test data, the importance of the pull-in voltage is slightly higher than that of the pull-in time; based on the above engineering practice, the following constraints are applied to the weights of the two input indicators: In the formula, It is the pull-in voltage PV input weight. It is the absorption time PT input weight.

[0080] Specifically, this embodiment of the invention applies explicit ranking constraints to the pull-in voltage weight and pull-in time weight, ensuring that the importance ranking of the two indicators always conforms to physical reality during the optimization process. The pull-in voltage is a fundamental indicator reflecting the pull-in conditions of the electromagnetic coil, spring mechanism, and contacts, directly determining whether the relay can operate reliably. The pull-in time mainly reflects the contact action response speed and is a measure of command timeliness. From the perspective of engineering fault mechanisms, abnormal pull-in voltage is often a more fundamental sign of performance degradation, and its importance should be higher than that of pull-in time. The indicator weight constraints avoid the risk of the optimization algorithm reversing the primary and secondary relationships of indicators due to blindly pursuing fitting accuracy, so that the optimized model maintains physical interpretability while improving accuracy, and maintains the credibility of the evaluation results in engineering practice.

[0081] In one embodiment of the present invention, step S3 applies interpretability constraints to the initial confidence rule base model to obtain a constrained confidence rule base model, and further includes: optimizing the parameters of the constrained confidence rule base model.

[0082] In a specific embodiment, the initial BRB model is constrained by both the subjectivity of expert experience and the uncertainty of environmental interference, resulting in significant accuracy deficiencies in relay performance evaluation. Therefore, data-driven parameter optimization is implemented: aiming to minimize prediction error, the core parameters of the constrained BRB model are globally optimized using a training set composed of multi-dimensional relay performance parameters and environmental parameters. This corrects the irrationality of the initial rules, enabling the model to more accurately reflect the operational laws of the actual system, thereby improving its evaluation performance. To minimize the gap between the model's evaluated utility and the actual utility, the parameters of the constrained BRB model are optimized. The difference between the evaluated utility and the actual utility is represented by MSE, and its expression is: In the formula, For the first i The model-predicted utility value for each sample. For the first i The true utility value of a sample Let be the mean square error function. It is a column vector composed of BRB model parameters.

[0083] It is important to note that the following basic constraints must also be met during optimization: .

[0084] Based on this, a complete constraint system is established to ensure that the parameter optimization results meet the actual requirements of the engineering. Then, the three constraint models are introduced into the general optimization model to obtain the constraint BRB model to be optimized, which has both interpretability and optimization efficiency.

[0085] Specifically, this invention, by imposing interpretability constraints on the initial BRB model and then further optimizing the parameters, effectively overcomes the problem of insufficient evaluation accuracy caused by the subjectivity of expert experience and the uncertainty of environmental interference in the initial BRB model. Although the expert knowledge in the initial BRB model has engineering rationality, it is limited by the ambiguity of expert cognition and individual differences. Its rule weights, indicator weights, and confidence distributions are difficult to accurately match the complex degradation laws of the actual system, resulting in significant prediction bias in quantitative evaluation. To address this, this invention takes minimizing the prediction error in the training sample set as the optimization objective, uses the mean squared error as the error metric function, and constructs an optimization problem with the mean of the squared difference between the model's predicted utility value and the actual utility value as the objective function. At the same time, all adjustable parameters, such as rule weights, indicator weights, and confidence distributions, are integrated into a unified decision variable vector, enabling the optimization algorithm to search for the optimal decision variable globally. The optimal parameter combination is achieved by fully embedding three types of interpretability constraints into the optimization model during the optimization process: monotonically decreasing / monotonically increasing / unimodal distribution constraints of confidence levels, physical ranking constraints of indicator weights, and differential interval constraints of rule weights. This ensures that parameter optimization not only pursues error minimization in a mathematical sense but also guarantees that the optimized parameters maintain physical interpretability and logical consistency in an engineering sense. The resulting constrained BRB optimization model corrects the inaccuracies in the initial expert knowledge through data-driven methods, enabling the model to more accurately fit the input-output mapping relationship of the actual system. Furthermore, thanks to the full effect of the triple interpretability constraints, it fundamentally avoids the problem of logical contradictions in the confidence distribution after optimization in traditional BRB models. While improving the evaluation accuracy, it fully preserves the traceability of the reasoning process, providing an evaluation result with both accuracy and reliability for subsequent maintenance timing decisions.

[0086] In one embodiment of the present invention, the leader bee and follower bee of the optimized artificial bee colony algorithm randomly select any one of the following strategies for position update during neighborhood search: standard difference search strategy, elite-guided search strategy, and optimal mutation search strategy. The standard difference search strategy includes: the current artificial bee randomly selects another individual in the neighborhood and updates its position using the position difference between the two individuals as a perturbation. The elite-guided search strategy includes: the current artificial bee uses the best individual in the current population as a guide direction and moves towards the region where the best individual is located to update its position. The optimal mutation search strategy includes: based on the position of the best individual in the current population, superimposing the position difference between two randomly selected other individuals as a perturbation to generate a new position.

[0087] In a specific embodiment, for the aforementioned non-convex optimization problem of BRB parameters with interpretability constraints, its solution space exhibits complex characteristics of multi-peak and nonlinearity, with a large number of local optima. Traditional gradient-based optimization algorithms are easily affected by initial values ​​and are prone to falling into local optimum traps during iteration, making it difficult to converge to the global optimum and failing to simultaneously guarantee the model's fitting accuracy and interpretability. Therefore, this invention employs the ABC algorithm with a local extremum escape mechanism to solve this optimization problem.

[0088] By integrating all the parameters to be optimized into a unified set and using this set to encode and map the positions of artificial bees, the complex parameter optimization problem can be transformed into an intelligent search process using the artificial bee colony algorithm. Simultaneously, by constructing a fitness function specific to the ABC algorithm based on a predetermined optimization objective function, the model evaluation metric and algorithm iteration criteria can be unified. Based on this mapping logic, the first... b The location of artificial bees is defined by the following encoding: In the formula, represents artificial bees. b Spatial positioning, C This is the dimension of the space; the other parameters are interpreted the same way.

[0089] The fitness function of the artificial bee colony algorithm is: In the formula, Indicates the first b Only the adaptability of artificial bees.

[0090] Then, a complete iterative execution framework for optimizing the artificial bee colony algorithm is defined: In the formula, express b The artificial bee was in the first c Position on the dimension and They represent the first b The artificial bee was in the first c The lower and upper bounds of the search on the dimension, c Number the dimensions. The formula is a random number within the range (0,1). It is mainly used to start the algorithm, randomly generate the initial position of each individual bee in the colony, and complete the initial configuration of the population. This indicates that the follower bee chooses the leader bee. b The probability, Indicates the total number of leading bees. Indicates the leading bee b The corresponding fitness value, the formula represents sorting by fitness, with the top high-quality individual being the leader bee and the rest being follower bees, is used to calculate the selection probability; This signifies leading the bees to a new position. Another artificial bee in the colony e exist cThe position of the dimension is any value between 0 and 1. This formula means that the follower bee locks the leader bee according to the selection probability. The two types of bee colonies use differential offset local optimization to update the position. Artificial bees b In the c The new position on the dimension, the formula means that when the number of times an individual seeks optimization reaches the preset limit and the fitness does not improve, it switches to a scout bee and resets the position to avoid local optima.

[0091] However, it is worth noting that the standard artificial bee colony algorithm uses the random difference method, which has the following shortcomings: 1) The core reason for the algorithm's low search efficiency is that the search phase lacks the influence of population experience and superior individuals. Artificial bees learn indiscriminately from any individual in the population, making it difficult to guarantee that they will gravitate towards superior individuals; 2) The search method is relatively limited, making it difficult to maintain population diversity, which can easily lead to premature convergence and getting trapped in local optima. To solve these problems, this invention considers the elite guidance strategy of the particle swarm optimization algorithm and the optimal mutation strategy of the differential evolution algorithm, adding two new search strategies to the standard ABC algorithm, constructing a search strategy pool containing three parallel strategies. The leader bee and follower bees randomly select one of these strategies during neighborhood search, which can effectively improve the algorithm's global search capability and convergence rate. The two new search strategies are: In the formula, represents the optimal solution at . c Coordinates in dimensional space; , , It can be any value between 0 and 1; the other parameters are interpreted the same way.

[0092] Two new preset schemes have been added: the optimal search scheme and the Differential Evolution (DE) / best / 1 search scheme. During the artificial bee's position iteration process, it will randomly select one of the three preset schemes and use a greedy mechanism to complete the orientation coordinate update decision.

[0093] Specifically, this invention overcomes the shortcomings of the standard artificial bee colony algorithm, such as its singular search method and insufficient population diversity, by constructing three parallel search strategies and allowing random selection of artificial bees. Standard difference search ensures the algorithm's basic exploration capability; elite-guided search uses the best individual in the current population as the guiding direction, accelerating the convergence speed towards the optimal region; and optimal mutation search, based on the best individual superimposed with random difference perturbation, conducts fine-tuning near the optimal solution to find a better solution. The parallel and random switching of the three strategies enables individuals within the population to have diverse search behaviors, which can both utilize the information of the best individual to accelerate convergence and maintain population diversity to avoid premature convergence, thereby significantly improving the algorithm's global search capability and convergence rate, enabling it to efficiently solve complex non-convex optimization problems constrained by the BRB model.

[0094] In one embodiment of the present invention, after obtaining the performance index data of the relay and before constructing the initial confidence rule base model based on the performance index data, the method further includes: using an overlapping sliding window algorithm to augment the original test samples of the performance index data, including: truncating subsequences on the original sample sequence by sliding a window with a preset step size, taking each subsequence as a new sample, and obtaining a sample dataset for training and testing the constraint confidence rule base model.

[0095] In a specific embodiment, a single relay initially has only 25 test samples. Using an overlapping sliding window algorithm, the initial samples can be expanded to 1500 effective samples, with 1200 used for model training and 300 for model testing. Specifically, the overlapping sliding window expands the sample size effectively without altering the original data distribution characteristics and physical laws by sliding a window with a fixed step size over the original sample sequence, treating each subsequence within the window as a new sample.

[0096] Specifically, the embodiments of the present invention effectively solve the engineering pain point of extremely limited relay test samples by using an overlapping sliding window algorithm for data augmentation. Compared with traditional data augmentation methods, the overlapping sliding window makes full use of the local correlation of time series data, and achieves effective expansion of the sample size while preserving the original data distribution characteristics and physical laws. This provides sufficient data support for the subsequent training and optimization of the constrained BRB model, significantly reduces the demand for destructive fault test samples, and lowers engineering costs.

[0097] In one embodiment of the present invention, the degradation mapping model is an inverse Gaussian degradation process model, wherein the cumulative degradation mean of the inverse Gaussian degradation process model changes with time according to a power function law; based on the performance state utility value, the quantitative maintenance timing of the relay is calculated using the preset degradation mapping model, including: linearly normalizing the performance state utility value to map it to the degradation domain to obtain the current relative degradation amount; based on the current relative degradation amount, calculating the conditional reliability function of the relay using the inverse Gaussian degradation process model, wherein the conditional reliability function is used to characterize the probability that the relay will still maintain qualified performance after a specified remaining time, given that it has degraded to the current relative degradation amount; using a preset reliability threshold as the decision boundary, solving for the remaining time required for the conditional reliability function to decrease to the reliability threshold, and back-calculating the quantitative maintenance timing.

[0098] The relay performance status, as represented by the performance status utility value, is positively correlated with the value of the performance status utility value. The current relative degradation is obtained by inversely linearly normalizing the performance status utility value. The current relative degradation is negatively correlated with the performance status utility value and positively correlated with the current degree of degradation of the relay.

[0099] Specifically, this invention describes the stochastic evolution of relay performance degradation by establishing an inverse Gaussian degradation process model, achieving a precise quantitative mapping from performance utility value to maintenance timing. The inverse Gaussian degradation process possesses excellent mathematical properties and physical interpretability; its cumulative degradation mean follows a power function over time, effectively characterizing the nonlinear trend and random fluctuations of relay performance degradation. By linearly mapping the performance utility value to the current relative degradation amount, and then using the conditional reliability function and a preset reliability threshold to inversely solve for the optimal maintenance timing, the determination of maintenance timing no longer relies on manual experience or fixed cycles, but is based on scientific quantitative calculations of the device's current actual performance state and reliability degradation patterns. This significantly reduces the cost and subjectivity of maintenance decisions in engineering, effectively avoiding over-maintenance (premature maintenance leading to resource waste) or under-maintenance (late maintenance leading to mission failure), providing direct and quantitative decision support for condition-based equipment maintenance.

[0100] In one embodiment of the present invention, determining the performance state utility value of a relay based on a target confidence distribution includes: obtaining a fixed utility value corresponding to each preset performance level, wherein the fixed utility value corresponds one-to-one with the preset performance level, and the fixed utility value increases as the preset performance level increases; and using the confidence level of each performance level in the target confidence distribution as a weight, performing a weighted summation with the corresponding fixed utility value to obtain the performance state utility value.

[0101] In a specific embodiment, the relay performance is divided into four levels: Excellent (E), Good (G), Average (M), and Poor (P). The fixed utility values ​​corresponding to each level are: Poor corresponds to utility value 1, Average corresponds to utility value 2, Good corresponds to utility value 3, and Excellent corresponds to utility value 4. The confidence level of each performance level in the optimized target confidence distribution is used as a weight, and the corresponding fixed utility value is weighted and summed to obtain the performance status utility value. The performance status utility value ranges from [1,4], and the higher the value, the better the overall performance status of the device.

[0102] Specifically, the embodiments of the present invention transform the target confidence distribution into a weighted utility value scalar, compressing multi-dimensional confidence information into a single value, which facilitates intuitive comparison and ranking of the performance status of different devices, and also provides a unified input interface for the quantitative calculation of subsequent maintenance timing, realizing a smooth connection from performance evaluation to maintenance decision-making.

[0103] The advantages and benefits of the relay maintenance timing determination method considering interpretability according to the above embodiments of the present invention are explained below with reference to a specific embodiment. In this specific embodiment, comparative experiments and results between the present method and other existing evaluation methods are provided. Specifically, the model evaluation accuracy and interpretability are verified by comparing and analyzing the iterative effects of the constraint confidence rule base with the traditional confidence rule base and the confidence rule base with existing evaluation methods; at the same time, the corresponding maintenance timing results are generated based on the performance utility value output by the optimal evaluation model.

[0104] In this specific embodiment, before model validation is carried out, it is necessary to analyze the optimization effect of the optimization algorithm and verify the feasibility of optimizing the artificial bee colony algorithm. In this specific embodiment, a total of 1200 training groups participated in model optimization.

[0105] First, to verify the effectiveness of the optimized artificial bee colony algorithm, three methods were set up: the improved ABC algorithm (in this invention), the standard ABC algorithm, and the particle swarm optimization algorithm. The maximum number of iterations was set to 500, and these methods were used to optimize the constrained BRB model and the traditional BRB model, respectively.

[0106] Figure 3 This is a schematic diagram of the constrained BRB model optimization process according to a specific embodiment of the present invention. Figure 4 This is a schematic diagram of the optimization process of a traditional BRB model according to a specific embodiment of the present invention, as shown below. Figure 3 and Figure 4 As shown in Table 5, the specific iterative process of different algorithms under different models is illustrated, and the corresponding iterative effects (objective function F value) are shown in Table 5.

[0107] Table 5 Comparison of Optimization Iteration Effects of Different Algorithms

[0108] according to Figure 3 and Figure 4 As can be clearly seen from Table 5, the overall optimization performance and optimization ability of the improved ABC algorithm are significantly better than those of the standard ABC algorithm and the traditional particle swarm optimization algorithm.

[0109] Then, the interpretability of the constrained BRB was analyzed; the traditional BRB model and the constrained BRB model were optimized using 1200 training sets, and the distribution of their 9 BR rules was compared.

[0110] Figure 5 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 1. Figure 6 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 2. Figure 7This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 3. Figure 8 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 4. Figure 9 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 5. Figure 10 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 6. Figure 11 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 7. Figure 12 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 8. Figure 13 This is a schematic diagram illustrating the confidence distribution of two BRB models according to a specific embodiment of the present invention under rule 9, as shown below. Figures 5-13 As shown in Table 6, the confidence distribution of the two BRB models under different rules is presented, and the corresponding interpretability analysis results are shown in Table 6.

[0111] Table 6. Confidence distribution of the nine rules in the two BRB models.

[0112] from Figures 5-13 Table 6 clearly shows two key conclusions: First, for both the constrained BRB model and the traditional BRB model, their confidence distributions on rules 1, 2, 4, 6, 7, and 9 are very similar. This indicates that the constrained BRB model, while ensuring its interpretability, fully preserves the core information of expert knowledge and data optimization, without excessive information loss. Furthermore, the optimal parameters of the two BRB models corroborate each other, further demonstrating the reliability of the results. Second, the traditional BRB model has significant problems; its rules... 3 and 5 contain serious logical contradictions. For example, the confidence distribution of rule 3 is (0.05, 0.42, 0.05, 0.55). The confidence level for "good" performance of the surface relay is 0.42, and the confidence level for "poor" performance is 0.55. The confidence levels for other levels, such as "excellent" and "medium", are close to 0. This creates a logical gap of "either good or poor", which does not conform to the actual evaluation rules. At the same time, it also verifies that the confidence distribution of all constrained BRB models strictly follows the division boundaries, and there are no logical conflicts, thus fully maintaining its original good interpretability.

[0113] Finally, the overall evaluation accuracy of the BRB model is analyzed.

[0114] Figure 14 This is a schematic diagram of a relay evaluation test result according to a specific embodiment of the present invention. Figure 15This is a schematic diagram illustrating the evaluation results of a relay maintenance timing according to a specific embodiment of the present invention, such as... Figure 14 and Figure 15 As shown, the focus is on analyzing the constrained BRB model. The evaluation result of a relay using the constrained BRB model is as follows. Figure 14 As shown in Table 7, the utility value and maintenance timing are then calculated based on the performance evaluation results. Figure 15 As shown.

[0115] Table 7. Results of an assessment of relay performance level distribution and maintenance timing.

[0116] To verify the overall evaluation effect of the BRB model, 1500 sets of samples were input into the constrained BRB model, the traditional BRB model, and the BP (Back Propagation) neural network to conduct experiments. Two sets of comparisons were constructed: one set compared the errors of the constrained BRB and the traditional BRB models, and the other set compared the overall error of the BRB model and the BP neural network. This was to verify the overall performance evaluation effect of the BRB model under the premise of accuracy loss of the constrained BRB.

[0117] The BP neural network built in this specific embodiment adopts a mature three-layer topology architecture. The network input end is set with two neural nodes, and the output end is only configured with one neural node. The entire network relies on 1200 sets of experimental samples to complete iterative training, and the gradient descent algorithm is used to complete the parameter iterative update. After the specific process is completed, the remaining 300 sets of samples are used to test the model and complete the simulation test and performance verification.

[0118] Figure 16 This is a schematic diagram illustrating the evaluation error of utility values ​​according to a specific embodiment of the present invention, such as... Figure 16 As shown, 10 samples were randomly selected to test the constrained BRB model, the traditional BRB model, and the BP neural network. The comparison of their utility value evaluation errors is shown in the figure. Figure 16 The specific errors are shown in Table 8.

[0119] Table 8 Output Error Table for Three Models

[0120] according to Figure 16Based on the results in Table 8, the average evaluation errors of the three models are 0.2528, 0.1167, and 0.1226, respectively. Combining these experimental results with other verification findings, the following conclusions can be drawn: 1) The improved ABC algorithm has the fastest convergence speed, the smallest final objective function value, and significantly better global optimization capabilities than the other two algorithms, enabling efficient parameter optimization of the constrained BRB model; 2) The confidence distribution of all rules in the constrained BRB model conforms to interpretability constraints, has no logical contradictions, and the entire evaluation and decision-making process is traceable; while rules 3 and 5 of the traditional BRB model exhibit logical breaks, losing interpretability, and their output maintenance suggestions lack reliable explanatory basis; 3) While maintaining interpretability, the constrained BRB model suffers an accuracy loss of only 5.06%, and its evaluation accuracy and maintenance timing determination accuracy are far superior to the BP neural network under small sample conditions. It achieves comprehensive interpretability assurance within an acceptable range of accuracy loss, verifying the effectiveness and engineering practicality of the method.

[0121] In summary, the relay maintenance timing determination method considering interpretability provided by this invention fundamentally solves the problem of interpretability loss during the parameter optimization process of traditional BRB models by constructing a confidence rule base model with triple interpretability constraints, achieving full traceability and interpretability of the evaluation reasoning process. By integrating an optimized artificial bee colony algorithm with a multi-strategy search mechanism, it significantly improves the solution effect of non-convex optimization problems, with global optimization capability and convergence speed superior to standard artificial bee colony and particle swarm algorithms. By establishing an inverse Gaussian degradation mapping model from performance utility values ​​to maintenance timing, it achieves effective connection between performance evaluation and maintenance decision-making, directly outputting quantitative maintenance timing suggestions. Thus, it achieves a balance between evaluation accuracy and interpretability in small sample scenarios, and is applicable to pre-mission performance optimization and full life-cycle status monitoring of relays in various equipment.

[0122] In other words, the relay maintenance timing determination method considering interpretability provided by the present invention has at least the following advantages compared with the prior art: 1. This method fundamentally solves the problem of lost interpretability of the BRB model. By introducing three types of interpretability constraints, this method ensures that the confidence distribution of each rule in the entire parameter optimization process conforms to logical consistency, completely avoiding the logical paradox of "either good or bad" in traditional models, and realizing full traceability and interpretability of the evaluation reasoning process. 2. Significantly improves the solution performance for non-convex optimization problems; This method utilizes an optimized artificial bee colony algorithm with multi-strategy parallel search and local extremum escape mechanism, resulting in a significantly faster convergence speed compared to the standard ABC algorithm. Its global optimization capability is significantly superior to the standard ABC algorithm and other traditional algorithms, enabling efficient solution of complex non-convex optimization problems constrained by BRB models. By adding constraints to the two types of weights, the optimization space is limited, the iteration range is greatly reduced, and the optimization efficiency is further improved.

[0123] 3. A balance between accuracy and interpretability is achieved in small sample scenarios. For the problems of limited relay samples and multiple error sources, this method improves the algorithm to enhance the efficiency of optimization iteration based on the initial BRB model obtained from the expert knowledge base. At the same time, it adds constraints according to logical facts, engineering practice and expert experience to compress its optimization space. By improving the accuracy of the algorithm with a small number of samples, the constraints ensure interpretability, which greatly reduces the amount of test samples required. 4. Effective integration of performance evaluation and maintenance decision-making is achieved. After constructing an interpretable constraint confidence rule base, this method provides corresponding specific implementation steps for performance evaluation. Based on this, a mapping function from performance utility value to maintenance timing is established, which can directly output quantitative maintenance timing suggestions, greatly reducing the cost and subjectivity of maintenance decisions in engineering.

[0124] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0125] Further embodiments of the present invention disclose a relay maintenance timing determination system that takes into account interpretability. Figure 17 This is a block diagram of a relay maintenance timing determination system that considers interpretability according to an embodiment of the present invention. Figure 17 As shown, in one embodiment of the present invention, the relay maintenance timing determination system 100 considering interpretability includes: a data acquisition module 110, an initial model construction module 120, a constraint model construction module 130, a parameter optimization module 140, a utility value determination module 150, and a maintenance timing calculation module 160.

[0126] Specifically, the data acquisition module 110 is used to acquire the performance index data of the relay, including the pull-in voltage and pull-in time.

[0127] The initial model building module 120 is used to build an initial confidence rule base model based on performance index data. The initial confidence rule base model includes multiple confidence rules. The initial confidence rule base model is used to output the initial confidence distribution of the relay corresponding to each preset performance level based on the matching degree between the performance index data and the preset performance level.

[0128] The constraint model construction module 130 is used to apply interpretability constraints to the initial confidence rule base model to obtain a constraint confidence rule base model. The interpretability constraints include: constraining the logical consistency of the initial confidence distribution during the parameter optimization process so that the optimized confidence distribution conforms to a preset monotonic change law or a preset unimodal distribution law; and constraining the optimization range of rule weights and indicator weights in the initial confidence rule base model.

[0129] The parameter optimization module 140 is used to optimize the parameters of the constraint confidence rule base model using the optimized artificial bee colony algorithm to obtain the optimized target confidence distribution. The optimized artificial bee colony algorithm integrates multiple search strategies to enhance global search capabilities and convergence speed on the basis of the standard artificial bee colony algorithm.

[0130] The utility value determination module 150 is used to determine the performance state utility value of the relay based on the target confidence distribution.

[0131] The maintenance timing calculation module 160 is used to calculate the quantitative maintenance timing of the relay based on the performance status utility value and using a preset degradation mapping model.

[0132] It should be noted that the specific implementation of the relay maintenance timing determination system 100 considering interpretability in the embodiments of the present invention is similar to the specific implementation described in the relay maintenance timing determination method considering interpretability in the above embodiments of the present invention, and therefore has similar technical effects. For details, please refer to the description of the relay maintenance timing determination method considering interpretability section. In order to reduce redundancy, it will not be repeated here.

[0133] Further embodiments of the present invention also disclose an electronic device, Figure 18 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Figure 18 As shown, in one embodiment of the present invention, the electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store computer programs; when the processor executes the computer program stored in the memory, it implements the relay maintenance timing determination method considering interpretability as described in any of the above embodiments of the present invention.

[0134] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0135] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0136] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0137] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0138] It should be noted that the specific implementation of the electronic device in the embodiments of the present invention is similar to the specific implementation described in the above embodiments of the present invention regarding the method for determining relay maintenance timing with consideration of interpretability, and therefore has similar technical effects. For details, please refer to the description in the section on determining relay maintenance timing with consideration of interpretability. To reduce redundancy, it will not be repeated here.

[0139] Further embodiments of the present invention disclose a computer-readable storage medium storing a computer program that, when executed by a processor, implements the relay maintenance timing determination method considering interpretability as described in any of the above embodiments of the present invention.

[0140] It should be noted that the specific implementation of the computer-readable storage medium in the embodiments of the present invention is similar to the specific implementation described in the above embodiments of the present invention regarding the method for determining relay maintenance timing with consideration of interpretability, and therefore has similar technical effects. For details, please refer to the description in the section on determining relay maintenance timing with consideration of interpretability. To reduce redundancy, it will not be repeated here.

[0141] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0142] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A method for determining relay maintenance timing that considers interpretability, characterized in that, Includes the following steps: Acquire the performance index data of the relay, wherein the performance index data includes pull-in voltage and pull-in time; An initial confidence rule base model is constructed based on the performance index data. The initial confidence rule base model includes multiple confidence rules. The initial confidence rule base model is used to output the initial confidence distribution of the relay corresponding to each preset performance level based on the matching degree between the performance index data and the preset performance level. An interpretability constraint is applied to the initial confidence rule base model to obtain a constrained confidence rule base model. The interpretability constraint includes: constraining the logical consistency of the initial confidence distribution during the parameter optimization process so that the optimized confidence distribution conforms to a preset monotonic change law or a preset unimodal distribution law; and constraining the optimization range of rule weights and indicator weights in the initial confidence rule base model. The optimized artificial bee colony algorithm is used to optimize the parameters of the constraint confidence rule base model to obtain the optimized target confidence distribution. The optimized artificial bee colony algorithm integrates multiple search strategies to enhance global search capability and convergence speed on the basis of the standard artificial bee colony algorithm. The performance state utility value of the relay is determined based on the target confidence distribution; Based on the performance status utility value, the quantitative maintenance timing of the relay is calculated using a preset degradation mapping model.

2. The method for determining relay maintenance timing considering interpretability according to claim 1, characterized in that, The constraint on the logical consistency of the initial confidence distribution during the parameter optimization process includes: The optimized confidence distribution of each confidence rule is limited to any one of the following forms: monotonically decreasing distribution, monotonically increasing distribution, and unimodal distribution. The monotonically decreasing distribution includes: the confidence level corresponding to each preset performance level decreases sequentially from excellent to poor as the performance level decreases; The monotonically increasing distribution includes: the confidence level corresponding to each preset performance level increases sequentially from excellent to poor as the performance level decreases; The unimodal distribution includes: the confidence distribution has a peak value, and the confidence values ​​on both sides of the peak value decrease sequentially.

3. The method for determining relay maintenance timing considering interpretability according to claim 1, characterized in that, Constraining the optimization range of rule weights in the initial confidence rule base model includes: Based on the actual occurrence probability of the input combination of the performance index data corresponding to each confidence rule and the degree of influence on the task success rate, a differentiated weight constraint range is set for each confidence rule; The input combination is a combination of pull-in voltage level and pull-in time level. Both the pull-in voltage level and the pull-in time level include three levels: large, medium, and small. The pull-in voltage level and the pull-in time level are combined in pairs to form nine confidence rules. For confidence rules where both pull-in voltage and pull-in time are of medium level, the lower limit of the corresponding weight constraint interval is set to the highest level. For confidence rules where both the pull-in voltage and the pull-in time are of a high level or both are of a low level, the lower limit of the corresponding weight constraint interval is set to the second highest level. For a confidence rule where one of the pull-in voltage and the pull-in time is of a medium level and the other is of a large or small level, the lower limit of the corresponding weight constraint interval is set to the lowest level.

4. The method for determining relay maintenance timing considering interpretability according to claim 1, characterized in that, The weights of the indicators include pull-in voltage weight and pull-in time weight; Constraining the optimization range of the indicator weights in the initial confidence rule base model includes: Since the pull-in voltage is more effective at characterizing the state of the relay pull-in mechanism and the electromagnetic coil than the pull-in time is effective at characterizing the contact action hysteresis, the lower limit of the pull-in voltage weight is set to be greater than the upper limit of the pull-in time weight.

5. The method for determining relay maintenance timing considering interpretability according to claim 1, characterized in that, The optimized artificial bee colony algorithm uses the leader bee and follower bee to randomly select any one of the standard difference search strategy, elite guided search strategy and optimal mutation search strategy to update their positions during neighborhood search. The standard difference search strategy includes: the current artificial bee randomly selects another individual in the neighborhood and uses the position difference between the two individuals as a perturbation to update its position; The elite-guided search strategy includes: the current artificial bee uses the best individual in the current population as the guiding direction and moves towards the area where the best individual is located to update its position; The optimal mutation search strategy includes: generating a new position by superimposing the positional difference between two randomly selected individuals as a perturbation based on the position of the best individual in the current population.

6. The method for determining relay maintenance timing considering interpretability according to claim 1, characterized in that, After acquiring the relay's performance index data, and before constructing an initial confidence rule base model based on the performance index data, the method further includes: The original test samples of the performance index data are augmented using an overlapping sliding window algorithm, including: Subsequences are extracted from the original sample sequence by sliding a window with a preset step size, and each subsequence is treated as a new sample to obtain a sample dataset for training and testing the constraint confidence rule base model.

7. The method for determining relay maintenance timing considering interpretability according to claim 1, characterized in that, The degradation mapping model is an inverse Gaussian degradation process model, wherein the cumulative degradation mean of the inverse Gaussian degradation process model changes with time according to a power function law; The step of calculating the quantified maintenance timing of the relay based on the performance state utility value using a preset degradation mapping model includes: The performance state utility value is linearly normalized and mapped to the degradation domain to obtain the current relative degradation. Based on the current relative degradation amount, the conditional reliability function of the relay is calculated using the inverse Gaussian degradation process model, wherein the conditional reliability function is used to characterize the probability that the relay will still maintain qualified performance after a specified remaining time, given that it has degraded to the current relative degradation amount. Using a preset reliability threshold as the decision boundary, the remaining time required for the conditional reliability function to decrease to the reliability threshold is calculated, and the quantitative maintenance timing is then derived.

8. The method for determining relay maintenance timing considering interpretability according to claim 7, characterized in that, The relay performance state characterized by the performance state utility value is positively correlated with the value of the performance state utility value; The current relative degradation amount is obtained by inversely linearly normalizing the performance state utility value. The current relative degradation amount is negatively correlated with the performance state utility value and positively correlated with the current degradation degree of the relay.

9. The method for determining relay maintenance timing considering interpretability according to claim 1, characterized in that, Determining the performance state utility value of the relay based on the target confidence distribution includes: Obtain a fixed utility value corresponding to each preset performance level, wherein the fixed utility value corresponds one-to-one with the preset performance level, and the fixed utility value increases as the preset performance level increases; The performance state utility value is obtained by weighting the confidence level of each performance level in the target confidence distribution with the corresponding fixed utility value.

10. A relay maintenance timing determination system that considers interpretability, characterized in that, include: The data acquisition module is used to acquire the performance index data of the relay, wherein the performance index data includes the pull-in voltage and pull-in time; An initial model building module is used to build an initial confidence rule base model based on the performance index data. The initial confidence rule base model includes multiple confidence rules. The initial confidence rule base model is used to output the initial confidence distribution of the relay corresponding to each preset performance level based on the matching degree between the performance index data and the preset performance level. The constraint model construction module is used to apply interpretability constraints to the initial confidence rule base model to obtain a constraint confidence rule base model. The interpretability constraints include: constraining the logical consistency of the initial confidence distribution during the parameter optimization process so that the optimized confidence distribution conforms to a preset monotonic change law or a preset unimodal distribution law; and constraining the optimization range of rule weights and indicator weights in the initial confidence rule base model. The parameter optimization module is used to optimize the parameters of the constraint confidence rule base model using an optimized artificial bee colony algorithm to obtain the optimized target confidence distribution. The optimized artificial bee colony algorithm integrates multiple search strategies to enhance global search capabilities and convergence speed on the basis of the standard artificial bee colony algorithm. The utility value determination module is used to determine the performance state utility value of the relay based on the target confidence distribution; The maintenance timing calculation module is used to calculate the quantitative maintenance timing of the relay based on the performance status utility value and using a preset degradation mapping model.

Images