A critical infrastructure maintenance decision method and system
By combining continuous-time Markov decision processes and dynamic programming with clustering dimensionality reduction techniques, an intelligent maintenance decision system is constructed, which solves the problems of response lag, inefficient resource allocation, and dimensionality curse in the maintenance of critical infrastructure, and achieves efficient maintenance decision-making and resource optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINHUANGDAO QINRE POWER GENERATION CO LTD
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional critical infrastructure maintenance strategies suffer from problems such as delayed response, inefficient resource allocation, dimensional disasters, and incomplete information, making it difficult to make optimal decisions in scenarios involving fixed-cycle maintenance and sudden outages.
A continuous-time Markov decision process and dynamic programming combined with clustering dimensionality reduction techniques are used to construct an intelligent maintenance decision system, including data acquisition, modeling, strategy fusion and optimization. Through MDP-DP collaborative decision-making, a seamless integration of routine maintenance and major outages is achieved.
It improves maintenance response efficiency, reduces maintenance costs, increases engineer utilization, enhances information utilization, and solves the dimensional explosion problem faced by traditional methods.
Abstract
Description
Technical Field
[0001] This invention relates to the field of critical infrastructure operation and maintenance management technology, and in particular to a critical infrastructure maintenance decision-making method and system. Background Technology
[0002] The reliable operation of critical infrastructure (such as the operation and maintenance of large power plant equipment, power systems, telecommunications networks, and transportation facilities) is crucial to socio-economic stability. Traditional maintenance strategies, primarily based on fixed-cycle maintenance or simple rule-based scheduling, suffer from the following technical shortcomings: 1. Response lag: Fixed-cycle maintenance cannot adapt to dynamic changes in equipment status, resulting in "over-maintenance" or "under-maintenance"; 2. Inefficient resource allocation: During major outages, the lack of scientific resource scheduling methods leads to redundant maintenance paths and response delays; 3. The curse of dimensionality: Infrastructure systems are massive in scale, and traditional optimization methods face the challenge of state space explosion; 4. Incomplete information: In actual operation and maintenance, fault information is often incomplete, making it difficult to make optimal decisions.
[0003] While some research has attempted to apply reinforcement learning and other methods in the existing technology, the following limitations still exist: A single model cannot adapt to both routine scheduling and sudden interruption scenarios simultaneously. It lacks an effective state space dimensionality reduction mechanism, resulting in high computational complexity; Ignoring the cost difference between preventative maintenance and corrective maintenance; The mechanism for competitive allocation of engineer resources is imperfect; Therefore, there is an urgent need for an intelligent maintenance decision-making method that can overcome the above-mentioned shortcomings. Summary of the Invention
[0004] The main objective of this invention is to provide a critical infrastructure maintenance decision-making method and system that integrates continuous-time Markov decision-making process and dynamic programming modeling, and combines clustering dimensionality reduction technology to create an intelligent maintenance decision-making solution applicable to the operation and maintenance management of critical infrastructure such as large power plants, telecommunications base stations, subway systems, and port equipment.
[0005] To achieve the above objectives, the present invention provides a critical infrastructure maintenance decision-making method, comprising the following steps: S100, Data Acquisition: Collect infrastructure status data, fault record data, and resource availability data. Facility status data includes facility location distribution, energy consumption status, task queue length, and regional congestion. Resource availability data includes the number and skill level of engineers and equipment availability information. S200, Modeling of Continuous-Time Markov Decision Processes: Based on the facility status data and resource availability data in step S100, a continuous-time MDP model is established for daily maintenance scheduling. The state space of the MDP model is defined as a quadruple. ,in, For facility location vectors, The energy consumption state and the value range is , The number of faults to be repaired. The regional congestion level and its value range is [value range missing]. ; The action space includes maintenance dispatch actions. Path replanning action Charging scheduling actions ; The reward function is constructed based on throughput, average latency, and conflict penalty, and the influence of each indicator is balanced by adjustable weighting coefficients. S300, Dynamic Programming and Clustering Dimensionality Reduction: Based on the facility status data in step S100, a dynamic programming model is established for major interruption response. The weighted affinity propagation clustering algorithm is used, and Mahalanobis distance is used as the similarity measure for clustering. The clustering features include facility geographical location, fault risk level, and maintenance urgency. Fault clusters are formed and clusters are used as decision units. S400, Preventive Maintenance Threshold Trigger: Based on the fault record data in step S100 and the energy consumption status in step S200, a threshold triggering strategy is designed based on the equipment deterioration degree. The equipment failure rate is calculated using the Weibull model, and then the real-time deterioration degree of the equipment is estimated. When the equipment deterioration degree meets the preset optimal threshold, preventive maintenance is performed. The optimal threshold is obtained by training with historical data from step S100. S500, Engineer Competitive Allocation Optimization: Based on the resource availability data in step S100 and the clustering results in step S300, a dynamic allocation mechanism for inspection and maintenance tasks is established. The allocation ratio is determined based on the information gain rate, risk coefficient, and unit delay cost. A two-stage decision rule is adopted: in the first stage, high-risk clusters are prioritized for inspection to obtain information, and in the second stage, maintenance scheduling is carried out based on the obtained information. S600, Strategy Integration and Optimization: By leveraging MDP-DP collaborative decision-making, we can integrate routine preventative maintenance with emergency response to major outages, and output preventative maintenance plans, emergency response scheduling schemes, and resource allocation schemes.
[0006] Preferably, facility location vector The range of values is Length of the task queue awaiting repair The range of values is , This represents the maximum number of faults to be repaired.
[0007] Preferably, the expression for the reward function is as follows: , The weighting coefficient is set as follows: These correspond to the impact weights of throughput, average latency, and conflict penalty, respectively.
[0008] Preferably, the formula for calculating Mahalanobis distance is as follows:
[0009] in, Let be the characteristic covariance matrix.
[0010] Preferably, the parameters of the AP clustering algorithm are set as follows: bias parameters Damping coefficient The maximum number of iterations is 1000, and the clustering quality threshold is... .
[0011] Preferably, clustering dimensionality reduction techniques reduce the state space dimension by 60-80%.
[0012] Preferably, the formula for estimating the real-time degradation degree of the equipment is as follows:
[0013] in, The cumulative operating time of the equipment, where the failure rate function is: The preset optimal threshold is The specific triggering conditions are as follows: Furthermore, the equipment operating time meets the requirement of 2500 hours < T < 5600 hours.
[0014] Preferably, the formula for calculating the engineer allocation ratio is as follows: ,in , ,
[0015]
[0016] Preferably, the MDP model is solved using a value iteration algorithm, with a convergence threshold. DP clustering re-clusters every 5 time steps, dynamically adjusting cluster boundaries.
[0017] The present invention also provides a critical infrastructure maintenance decision-making system, comprising: Data Acquisition Layer: Used to collect facility status data, fault log data, and resource availability data. Facility status data includes facility location distribution. Energy consumption status Task queue length Regional congestion The fault record data is historical fault information, and the resource availability data includes the number of engineers and their skill levels, as well as equipment availability information. Core decision engine: includes a routine maintenance scheduling module, a major interruption response module, and a strategy fusion and optimization module; The daily maintenance scheduling module adopts a continuous-time MDP model, and its state space is a quadruple. ,in For the facility's geographic coordinate vector, The energy consumption state and the value range is , The number of faults to be repaired. The regional congestion level and its value range is [value range missing]. The action space includes maintenance dispatch actions. Path replanning action Charging scheduling actions The reward function is constructed based on throughput, average latency, and conflict penalty. It balances the influence of each indicator through adjustable weight coefficients and outputs daily maintenance scheduling decisions based on state space data. The critical interruption response module adopts the DP model and the weighted AP clustering algorithm, using Mahalanobis distance as the similarity metric for clustering. Clustering features include facility geographical location, fault risk level, and maintenance urgency, forming fault clusters and using clusters as decision-making units to realize fault cluster generation, resource competition balancing, and information inspection trade-offs. The strategy fusion and optimization module is used to realize MDP-DP collaborative decision-making, integrating daily maintenance scheduling decisions and major outage response decisions; Decision output layer: This layer outputs preventive maintenance plans, emergency response scheduling schemes, and resource allocation schemes. The preventive maintenance plan is generated based on a threshold triggering mechanism. This mechanism calculates equipment failure rates and estimates real-time equipment degradation using a Weibull model, and then determines whether to perform preventive maintenance based on a preset optimal threshold, which is obtained through training with historical data. The emergency response scheduling scheme is generated based on clustering optimization results. The resource allocation scheme is generated based on an engineer competitive allocation mechanism. This mechanism determines the allocation ratio based on information gain ratio, risk coefficient, and unit delay cost, employing a two-stage decision rule. Execution and Feedback Layer: Used to perform maintenance operations, evaluate effectiveness through KPI analysis, and adjust model parameters based on the evaluation results.
[0018] The beneficial effects of this invention are as follows: By using MDP-DP fusion modeling, the average response latency is shortened, thereby improving work efficiency; Preventing sudden failures and reducing maintenance costs by establishing a preventative maintenance threshold mechanism; Clustering dimensionality reduction techniques reduce the dimensionality of the state space, thereby mitigating the curse of dimensionality. Dynamic resource allocation mechanisms improve engineer utilization. Information inspection and balancing mechanisms improve the utilization rate of decision-making information. Detailed Implementation
[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0020] The maintenance decision system in this embodiment adopts a closed-loop architecture of data acquisition, core decision-making, output execution, and feedback optimization. Each level achieves data interoperability and collaboration through standardized interfaces, specifically including a data acquisition layer, a core decision engine, a decision output layer, and an execution and feedback layer.
[0021] The specific implementation steps of this embodiment are as follows: S100, Data Acquisition The data acquisition layer, through a multi-source sensing and management system, standardizes the collection of three core types of data, providing fundamental support for subsequent modeling: Data types and acquisition logic: Facility status data: Energy consumption status B (reflecting facility operating load, value range [0, 1]) and facility location distribution (geographic coordinates or regional identifiers) are collected through infrastructure built-in sensors, positioning systems, operation and maintenance management platforms, and regional status monitoring systems to obtain facility geographic coordinate vectors. ), Task queue length (to obtain the number of faults to be repaired) ), Regional congestion C (reflects the accessibility or availability of maintenance resource scheduling, with a value range of [0, 1]).
[0022] Fault record data: Standardized fault information is extracted from the historical operation and maintenance database, including fault type, occurrence scenario, scope of impact, repair process and related cost data, to form a structured fault dataset.
[0023] Resource availability data: Through the human resources management system and equipment spare parts management system, we collect the number of engineers, skill level distribution, and real-time availability and inventory information of maintenance equipment and spare parts.
[0024] Data preprocessing: The collected data is standardized using industry-standard methods for missing value imputation and outlier removal to ensure that the data quality meets the modeling requirements.
[0025] S200, Continuous-Time Markov Decision Process Modeling This module addresses routine maintenance scheduling scenarios by constructing a continuous-time MDP model to achieve dynamic adaptive scheduling. State space definition: using quadruples A comprehensive depiction of the facility's operational status and dispatch environment: For facility location vectors, The energy consumption state and the value range is , The number of faults to be repaired. The regional congestion level and its value range is [value range missing]. .
[0026] Action space execution: Based on state space features, three core types of actions are executed: Maintenance dispatching actions Based on facility failure risk and engineer skill level, achieve precise matching of maintenance personnel and failure tasks; Path replanning action Combined with regional congestion levels Dynamically adjust the routes of maintenance resources to reduce commuting delays; Charging scheduling actions For maintenance equipment or infrastructure that relies on electricity, coordinate charging timing and resources to ensure continuity of operation and maintenance.
[0027] Calculation of the reward function: using the formula Through adjustable weighting coefficients The impact of different metrics is balanced, with throughput reflecting maintenance efficiency, average latency reflecting response speed, and conflict penalty used to avoid scheduling failures caused by resource contention.
[0028] Model Solution: The MDP model is solved using a value iteration algorithm, with a convergence threshold set. (This embodiment is) After iterating until the model converges, the daily maintenance scheduling strategy is output.
[0029] S300, Dynamic Programming and Clustering Dimensionality Reduction This module addresses the problem of efficient scheduling for large-scale faults by employing dynamic programming (DP) modeling and clustering dimensionality reduction techniques to respond to critical outages. DP Model Construction: With "minimizing failure losses and shortest repair time" as the core objective, a state transition equation is constructed. The state dimensions include fault cluster priority, remaining resources, and repair progress, while the decision variables are the allocation ratio of various maintenance resources.
[0030] Clustering dimensionality reduction implementation: Clustering feature selection: Three core features are selected: facility geographical location, fault risk level, and maintenance urgency, to comprehensively reflect the spatial distribution of faults and their handling priorities; Similarity metric: Mahalanobis distance is used as the feature similarity metric, and the formula is as follows: , where Σ is the feature covariance matrix, which effectively avoids clustering bias caused by the correlation between features; Clustering parameter configuration: The weighted affinity propagation (AP) clustering algorithm is adopted, and the bias parameter, damping coefficient, maximum number of iterations and clustering quality threshold are set to ensure the rationality and stability of the clustering results; Dimensionality reduction effect: By clustering, scattered fault points are integrated into several fault clusters. With the clusters as decision-making units, the dimensionality of the state space is significantly reduced, solving the curse of dimensionality problem faced by traditional methods.
[0031] S400, Preventive Maintenance Threshold Trigger Based on equipment degradation patterns, a threshold triggering mechanism is established to enable precise initiation of preventative maintenance. Failure rate and degradation degree calculation: The Weibull model is used to fit the equipment failure rate function λ(t), combined with the cumulative operating time of the equipment, and the result is obtained through the formula.
[0032] Estimate the real-time degradation degree of the equipment x(T), where the parameters of λ(t) are obtained by training with historical fault data; Threshold triggering logic: Preset optimal preventive maintenance threshold (Determined through training with historical data), combined with constraints on equipment runtime, when the real-time degradation degree of the equipment... Reaching or exceeding When the time is 1, preventive maintenance operations are automatically triggered to avoid serious losses caused by sudden failures.
[0033] S500, Engineer Competition Allocation Optimization Establish a dynamic resource allocation mechanism to achieve efficient matching of engineers and maintenance tasks: Allocation ratio calculation: Based on information gain ratio (IGR), risk factor (RF), and unit delay cost (UDC), using the formula... The engineer allocation ratio is determined, where IGR reflects the trade-off between the informational value and time cost of inspection tasks, RF quantifies the risk level through the logit transformation of failure probabilities, and UDC reflects the economic losses from delayed maintenance. ,
[0034]
[0035] Two-stage decision-making rules: In the first stage, engineers are given priority to check the information of high-risk fault clusters to clarify the fault type and handling requirements; in the second stage, based on the accurate information obtained from the check, maintenance resources are re-scheduled and optimized to ensure that resources are invested in the most valuable maintenance tasks.
[0036] S600, Strategy Integration and Optimization Through the MDP-DP collaborative decision-making mechanism, seamless integration of routine maintenance and emergency response strategies is achieved. Collaboration Logic: In daily scenarios, the scheduling scheme output by the MDP model is the main one to ensure efficient handling of routine faults; in the event of a sudden interruption, the system automatically switches to the DP model, suspends low-priority daily tasks, centrally allocates resources to major fault clusters, and coordinates with relevant support departments (such as power and transportation departments) to provide collaborative support. Output scheme type: Preventive maintenance plans are generated based on threshold triggering mechanisms, clearly defining maintenance targets, time windows, and operational procedures. Emergency response and dispatch plan: generated based on clustering optimization results, clarifying the priority of fault cluster handling, resource allocation path and repair sequence; Resource allocation scheme: generated based on the engineer competitive allocation mechanism, which clarifies the task allocation ratio and scheduling rules for engineers of various skill levels.
[0037] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
Claims
1. A critical infrastructure maintenance decision method characterized by, Includes the following steps: S100, Data Acquisition: Collect infrastructure status data, fault record data, and resource availability data. Facility status data includes facility location distribution, energy consumption status, task queue length, and regional congestion. Resource availability data includes the number and skill level of engineers and equipment availability information. S200, Modeling of Continuous-Time Markov Decision Processes: According to the facility state data and the resource availability data in step S100, a continuous-time MDP model is established for daily maintenance scheduling, and a state space of the MDP model is defined as a four-tuple , wherein is a facility position vector, is an energy consumption state and has a value range of , is a number of faults to be repaired, is a regional congestion degree and has a value range of ; The action space includes a repair dispatch action , a path re-planning action , a charging schedule action ; The reward function is constructed based on throughput, average latency, and conflict penalty, and the influence of each indicator is balanced by adjustable weighting coefficients. S300, Dynamic Programming and Clustering Dimensionality Reduction: Based on the facility status data in step S100, a dynamic programming model is established for major interruption response. The weighted affinity propagation clustering algorithm is used, and Mahalanobis distance is used as the similarity measure for clustering. The clustering features include facility geographical location, fault risk level, and maintenance urgency. Fault clusters are formed and clusters are used as decision units. S400, Preventive Maintenance Threshold Trigger: Based on the fault record data in step S100 and the energy consumption status in step S200, a threshold triggering strategy is designed based on the equipment deterioration degree. The equipment failure rate is calculated using the Weibull model, and then the real-time deterioration degree of the equipment is estimated. When the equipment deterioration degree meets the preset optimal threshold, preventive maintenance is performed. The optimal threshold is obtained by training with historical data from step S100. S500, Engineer Competitive Allocation Optimization: Based on the resource availability data in step S100 and the clustering results in step S300, a dynamic allocation mechanism for inspection and maintenance tasks is established. The allocation ratio is determined based on the information gain rate, risk coefficient, and unit delay cost. A two-stage decision rule is adopted: in the first stage, high-risk clusters are prioritized for inspection to obtain information, and in the second stage, maintenance scheduling is carried out based on the obtained information. S600, Strategy Integration and Optimization: By leveraging MDP-DP collaborative decision-making, we can integrate routine preventative maintenance with emergency response to major outages, and output preventative maintenance plans, emergency response scheduling schemes, and resource allocation schemes.
2. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, Facility location vector The value range of The value range of The value range of , The maximum number of faults to be repaired.
3. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, The expression for the reward function is: ， The weighting coefficient is set as follows: These correspond to the impact weights of throughput, average latency, and conflict penalty, respectively.
4. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, The formula for calculating Mahalanobis distance is: ； in, Let be the characteristic covariance matrix.
5. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, The parameters of the AP clustering algorithm are set as follows: bias parameters Damping coefficient The maximum number of iterations is 1000, and the clustering quality threshold is... .
6. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, Clustering dimensionality reduction techniques can reduce the dimensionality of the state space by 60-80%.
7. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, The formula for estimating the real-time degradation of equipment is as follows: ； in, The cumulative operating time of the equipment is given by the failure rate function. The preset optimal threshold is The specific triggering conditions are as follows: Furthermore, the equipment operating time meets the requirement of 2500 hours < T < 5600 hours.
8. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, The formula for calculating the engineer allocation ratio is as follows: ,in ， ， 9. The critical infrastructure maintenance decision-making method according to claim 1, characterized in that, The MDP model is solved using a value iteration algorithm, with a convergence threshold. DP clustering re-clusters every 5 time steps, dynamically adjusting cluster boundaries.
10. A critical infrastructure maintenance decision-making system, characterized in that, include: Data Acquisition Layer: Used to collect facility status data, fault record data, and resource availability data. Facility status data includes facility location distribution, energy consumption status, task queue length, and regional congestion. Fault record data is historical fault information. Resource availability data includes the number and skill level of engineers and equipment availability information. Core decision engine: includes a routine maintenance scheduling module, a major interruption response module, and a strategy fusion and optimization module; The daily maintenance scheduling module adopts a continuous-time MDP model, and its state space is a quadruple. ,in For the facility's geographic coordinate vector, The energy consumption state and the value range is , The number of faults to be repaired. The regional congestion level and its value range is [value range missing]. ; The action space includes maintenance dispatch actions. Path replanning action Charging scheduling actions The reward function is constructed based on throughput, average latency, and conflict penalty. It balances the influence of each indicator through adjustable weight coefficients and outputs daily maintenance scheduling decisions based on state space data. The critical interruption response module adopts the DP model and the weighted AP clustering algorithm, using Mahalanobis distance as the similarity metric for clustering. Clustering features include facility geographical location, fault risk level, and maintenance urgency, forming fault clusters and using clusters as decision-making units to realize fault cluster generation, resource competition balancing, and information inspection trade-offs. The strategy fusion and optimization module is used to realize MDP-DP collaborative decision-making, integrating daily maintenance scheduling decisions and major outage response decisions; Decision output layer: Used to output preventive maintenance plans, emergency response scheduling schemes and resource allocation schemes. The preventive maintenance plan is generated based on a threshold triggering mechanism. This mechanism calculates the equipment failure rate and estimates the real-time deterioration of the equipment through the Weibull model, and determines whether to perform preventive maintenance based on a preset optimal threshold. The optimal threshold is obtained by training with historical data. The emergency response dispatch plan is generated based on clustering optimization results; the resource allocation plan is generated based on an engineer competitive allocation mechanism, which determines the allocation ratio based on information gain rate, risk coefficient and unit delay cost, and adopts a two-stage decision rule. Execution and Feedback Layer: Used to perform maintenance operations, evaluate effectiveness through KPI analysis, and adjust model parameters based on the evaluation results.