A method and system for generating a wind turbine multi-component risk constraint joint maintenance action

Abstract

The application discloses a kind of wind turbine multi-component risk constraint joint maintenance action generation method and system, comprising: obtaining operation monitoring data, historical state data, historical maintenance data and current maintenance resource data and constructs multi-component belief state, determines component risk, system risk and maintenance cost;Multi-component belief state is input into pre-trained offline strategy network, obtains the action tendency value corresponding to each to-be-maintained component inspection action, predictive maintenance action and replacement action output channel;According to action tendency value, component risk, belief state and maintenance action cost, determine the priority of each type of maintenance action candidate and generate candidate action set;The consistency correction of maintenance effect is carried out to predictive maintenance candidate action, and the action mutual exclusion relationship and resource demand are determined again;Resource feasible region is constructed, and combined optimization enabling index is determined according to system risk, candidate action conflict degree and maintenance resource tension degree, to generate system-level joint maintenance action and maintenance task data.

Description

Technical Field

[0001] This invention relates to the fields of wind turbine condition monitoring, intelligent operation and maintenance, and predictive maintenance decision-making, and more specifically to a method and system for generating joint maintenance actions for multiple components of a wind turbine under risk constraints. Background Technology

[0002] As the single-unit capacity of wind turbines and the installed capacity of wind farms continue to expand, multiple heterogeneous components such as gearboxes, generators, and main bearings are subjected to random wind loads, alternating loads, temperature fluctuations, changes in operating conditions, and transient shocks over a long period of time. This can easily lead to cumulative degradation, fatigue wear, random failures, and coupling degradation between components. How to rationally allocate limited inspection and maintenance resources while controlling the risk of unplanned downtime is an important technical issue in the intelligent operation and maintenance of wind turbines.

[0003] Currently, some existing methods identify components to be maintained based on health indicators, remaining service life, failure probability, or component risk value, or determine the maintenance scope based on the combined risk of multiple components. These methods can identify components that need attention, but cannot further determine whether the corresponding components should be inspected, predictive maintained, or replaced. They also cannot directly solve the problem of mutually exclusive actions and resource competition when multiple components generate different maintenance needs at the same time.

[0004] Other wind farm operation and maintenance optimization methods take existing fault tasks, early warning tasks, or planned maintenance tasks as inputs, and schedule tasks and allocate resources based on constraints such as maintenance personnel, repair equipment, spare parts, transportation vehicles, and maintenance time windows. These methods mainly solve the problem of how to execute predetermined maintenance tasks. They are usually not responsible for generating inspection, predictive maintenance, and replacement tasks based on the degradation status of components, nor can they solve the problem of mismatch of action types before the maintenance task is formed.

[0005] Some machine learning or reinforcement learning maintenance decision-making methods directly output maintenance actions or combined maintenance strategies based on equipment status. However, when a wind turbine contains multiple components to be maintained, and each component may correspond to multiple types of actions such as inspection, predictive maintenance, and replacement, the system-level joint action space will expand rapidly with the number of components. Directly outputting the final system-level joint action by the strategy model can easily lead to unexecutable results such as assigning multiple actions to the same component at the same time, exceeding maintenance resource limits, and assigning predictive maintenance actions to failed components.

[0006] Furthermore, the degradation states of wind turbine components have different physical compositions. Reversible degradation can be partially reduced through predictive maintenance operations such as lubrication, cleaning, adjustment, tightening, parameter calibration, or general repair. However, irreversible wear is caused by cumulative loads, fatigue damage, long-term wear, or random impacts, and usually requires replacement of parts or core sub-components for resetting. If predictive maintenance and replacement actions are determined solely based on uniform health indicators or action tendencies, predictive maintenance that cannot eliminate structural damage may continue to be performed when components are simultaneously in a high-risk and high-irreversible-wear state, resulting in residual risks after maintenance and inefficient use of limited maintenance resources. In addition, the actual degradation state and degradation parameters of wind turbine components are usually difficult to observe continuously and accurately. Some states can only be observed relatively accurately through on-site inspections, vibration detection, oil level detection, endoscopic examination, or electrical testing. When effective inspections are not performed, the uncertainty of state estimation will increase over time. Online reinforcement learning requires strategy trial and error through actual equipment interaction, which is not suitable for direct implementation on high-value wind turbines.

[0007] Therefore, under the conditions that the states of multiple components of a wind turbine are partially observable, degradation parameters are unknown, and maintenance resources are limited, how to transform the component-level inspection, predictive maintenance, and replacement action tendencies generated by the strategy model into system-level joint maintenance actions that are consistent with the component degradation mechanism, satisfy action mutual exclusion and resource constraints, and can directly form maintenance work orders is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of the above problems, this invention proposes a method and system for generating joint maintenance actions for multiple components of wind turbines under risk constraints, which overcomes or at least partially solves the above problems. The aim is to obtain the component-level action tendencies of inspection, predictive maintenance and replacement actions for each component without using actual wind turbines for online strategy trial and error. Then, based on the physical effects of maintenance actions, component risks, irreversible wear, action conflicts and limited maintenance resources, the component-level action tendencies are corrected for consistency of maintenance effects and decoded at the system level, converting them into system-level joint maintenance actions that are physically valid, resource-feasible and can directly form maintenance work orders.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for generating joint maintenance actions for multiple components of a wind turbine under risk constraints includes the following steps: S1. Obtain operation monitoring data, historical status data, historical maintenance data, and current maintenance resource data of multiple components of the wind turbine to be maintained, and construct the multi-component belief status of the wind turbine; S2. Establish a risk-constrained partially observable maintenance decision model, and determine the component risk, system risk, and maintenance cost based on the belief state of each component to be maintained; S3. Input the belief states of multiple components into a pre-trained offline policy network, and obtain the action tendency values ​​corresponding to each component to be maintained through the inspection action, predictive maintenance action, and replacement action output channels, respectively. S4. Determine the candidate priorities for each type of maintenance action based on the action tendency value, component risk, belief status, and maintenance action cost of each component to be maintained, and generate a set of candidate actions; S5. Perform maintenance action consistency correction on the predictive maintenance candidate actions, and redetermine the action mutual exclusion relationship and resource requirements of the corrected replacement candidate actions; S6. Construct the resource feasible region based on action mutual exclusion constraints, inspection resource constraints, maintenance resource constraints, failure component maintenance constraints, and action mask; S7. Determine the activation index for combined optimization based on system risk, the degree of conflict of candidate actions, and the degree of tension of maintenance resources, and use the combined action decoder to adaptively adopt the item-by-item resource projection or risk constraint combined optimization method to generate system-level joint maintenance actions within the resource feasible domain.

[0011] Preferably, in step S1, the multi-component belief state of the wind turbine includes the recoverable degradation state, irreversible wear state, state uncertainty, degradation parameter estimate, observation time, failure state, operating condition and remaining maintenance resources of each component to be maintained; The specific content of constructing the multi-component belief state of a wind turbine is as follows: Based on at least one operating condition variable in the operation monitoring data, determine the operating condition unit to which the component to be maintained belongs in the current maintenance decision cycle, and obtain normal operation samples to determine the normal state benchmark of the operation monitoring variables; Based on the normalized deviation of the operating conditions of each operating monitoring variable of the component to be maintained, determine the current degradation characterization value of the component to be maintained; Based on the current degradation characterization value, the state estimate of the previous maintenance decision cycle, the current operating conditions, the predictive maintenance action execution record, the component replacement record, and the state recovery amount before and after maintenance, the recursive state estimation method is used to determine the recoverable degradation state and irreversible wear state of the current maintenance decision cycle. The recoverable degradation state and the irreversible wear state together form the degradation state characterization of the component to be maintained. The state uncertainty is recursively determined based on the number of maintenance decision cycles since the most recent effective inspection and the status observation update after the effective inspection is performed. The degradation parameter estimates are updated recursively based on the deviation between the actual change in the condition of the component to be maintained and the preset baseline degradation rate. The observation timeframe is determined based on the time interval between the current moment and the most recent valid inspection moment; The failure state is determined based on the relationship between the current degradation characterization value of the component to be maintained and the corresponding failure boundary.

[0012] Preferably, the specific content of constructing the multi-component belief state of a wind turbine also includes: After performing the inspection action, the recoverable degradation state, degradation parameter estimates, and state uncertainty of the corresponding component to be maintained are updated based on the status observations obtained from the inspection. After performing predictive maintenance actions, the recoverable degradation state of the corresponding component to be maintained is reduced according to a preset recovery ratio or a recovery ratio determined based on historical maintenance results, while the irreversible wear state remains unchanged. After the replacement action is performed, the recoverable degradation state and irreversible wear state of the corresponding component to be maintained will be updated to the initial state after replacement. During maintenance decision cycles in which no effective inspections are performed, the state uncertainty of the corresponding component to be maintained is increased recursively according to the preset state uncertainty growth parameter.

[0013] Preferably, in step S2, the component risk value of each component to be maintained is:

[0014] in, Let i be the component risk value of the i-th component to be maintained. For risk smoothing parameters, Let be the total degradation amount of the i-th component to be maintained. Let t be the failure boundary of the i-th component to be maintained, t be the maintenance decision cycle, and i be the component number to be maintained.

[0015] in, Risk contribution parameters corresponding to irreversible wear conditions The parameter contributes to the risk corresponding to state uncertainty. To be in a reversible degenerative state, It is in an irreversible wear state. The state is uncertain; System risks of wind turbines for:

[0016] in, Risks associated with the aggregation of key components To mitigate the risks associated with component aggregation, A collection of key components. The assembly of supporting components includes at least one of a gearbox, a generator, a main bearing, and a generator bearing, and the supporting components include at least one of a pitch system, a yaw system, a hydraulic system, and a converter. Maintenance cost C per maintenance decision cycle t for:

[0017] in, , , , and These are inspection costs, predictive maintenance costs, replacement costs, failure event costs, and failure-held costs. , and These are inspection actions, predictive maintenance actions, and replacement actions; To add an invalidation flag, It is in a failed and unrepaired state. For downtime loss parameters, Losses due to downtime of critical components For joint scheduling cost parameters, To cover component scheduling costs, N represents the number of components to be maintained.

[0018] Preferably, step S3, the specific content of the pre-trained offline policy network, is as follows: By using historical operating data of wind turbine units, the baseline degradation rate, operating condition transition patterns, probability of random impact, intensity of random impact, and degradation coupling relationship between components are calibrated for each component to be maintained, and a multi-component degradation-maintenance environment is constructed. By interacting with the degradation-maintenance environment using hybrid behavioral strategies, a full offline maintenance trajectory dataset is constructed. The hybrid behavioral strategies include at least two of the following behavioral modes: conservative maintenance, balanced maintenance, reactive maintenance, delayed maintenance, under-maintenance, proactive replacement, failure exposure, and random exploration. Each transition sample includes the current multi-component belief state, the corresponding maintenance action, the immediate evaluation value, the next multi-component belief state, the maintenance trajectory termination identifier, and the action mask. The offline maintenance trajectory is divided into policy network training data, combined action decoder parameter calibration data, and independent test data. Based on the training data of the policy network, the cumulative discounted return of each transfer sample is calculated according to the real-time evaluation values ​​of each transfer sample from the corresponding maintenance decision cycle to the end maintenance decision cycle of the maintenance trajectory. The sample weights are determined and truncated based on the difference between the cumulative discounted return of each transfer sample and the average return of the samples in the offline maintenance trajectory dataset. Using the belief states of multiple components as input, and the inspection actions, predictive maintenance actions, and replacement actions corresponding to each component to be maintained as training labels, the offline policy network is trained using reward-weighted multi-action binary cross-entropy loss. During training and action generation, action masks are used to block maintenance actions that violate the maintenance constraints of the failed component or do not meet the action execution conditions. After training, the parameters of the offline policy network are fixed. The offline policy network includes a state input layer, a shared state feature extraction layer, and action output channels corresponding to inspection actions, predictive maintenance actions, and replacement actions, respectively. Each action output channel outputs action tendency values ​​corresponding to multiple components to be maintained.

[0019] Preferably, the specific content of step S4 is as follows: The priority of inspection actions is determined based on the inspection action tendency value, state uncertainty, observation timeliness, component risk value, and inspection action cost. The priority of inspection action candidates increases with the increase of state uncertainty, observation timeliness, and component risk value, and decreases with the increase of inspection action cost. The candidate priority of predictive maintenance actions is determined based on the predictive maintenance action tendency value, recoverable degradation state, component risk value, irreversible wear state, and predictive maintenance action cost. The candidate priority of predictive maintenance actions increases with the increase of recoverable degradation state and component risk value, and decreases with the increase of irreversible wear state and predictive maintenance action cost. The candidate priority of replacement actions is determined based on the replacement action tendency value, component risk value, irreversible wear state, failure state, and replacement action cost. The candidate priority of replacement actions increases with the component risk value, irreversible wear state, and failure severity, and decreases with the replacement action cost. The candidate priorities of each type of maintenance action are compared with the corresponding candidate trigger thresholds. Maintenance actions that meet the corresponding candidate trigger conditions and whose action masks are in an executable state are added to the candidate action set.

[0020] Preferably, in step S6, the specific content of the resource feasible domain is as follows: Within the same maintenance decision cycle, the same component to be maintained may be assigned at most one of the following actions: inspection, predictive maintenance, and replacement. The inspection resources consumed by the assigned inspection actions, including inspection personnel, inspection equipment, or inspection time, shall not exceed the available inspection resources corresponding to the current maintenance decision cycle. The maintenance resources consumed by the assigned predictive maintenance actions and replacement actions, including maintenance personnel, maintenance equipment, or maintenance time, shall not exceed the available maintenance resources corresponding to the current maintenance decision cycle. The number and type of spare parts required for the assigned replacement action shall not exceed the current number and type of spare parts available. Failed components are not assigned predictive maintenance actions and are placed in either the replacement candidate queue or the corrective maintenance candidate queue. All maintenance actions entering the resource feasible domain satisfy the corresponding action mask; When maintenance actions for different components require the use of the same non-parallel maintenance equipment or occupy conflicting maintenance time windows, the corresponding maintenance actions should be set to not be executed simultaneously.

[0021] Preferably, in step S7, the specific content of generating the system-level joint maintenance action using the combined action decoder is as follows: The system risk, the degree of action conflict, and the current maintenance resource shortage are normalized respectively, and then weighted and merged based on the corresponding indicator weights to obtain the combined optimized activation indicators; The degree of action conflict is determined based on the number of multiple candidate actions corresponding to the same component to be maintained, the degree to which candidate actions of different components to be maintained compete for the same maintenance resources, and the degree of maintenance time conflict between candidate actions; the current level of maintenance resource shortage is determined based on the relative relationship between the total resource demand of the candidate action set and the currently available maintenance resources. When the combined optimization activation index does not reach the preset activation threshold, the item-by-item resource projection method is adopted to check each candidate action in descending order of priority. If there is no conflict between the current candidate action and the selected action, and the resource feasibility domain is still satisfied after adding the current candidate action, the current candidate action is added to the system-level joint maintenance action, and the remaining inspection resources and remaining maintenance resources are updated. The candidate action check and remaining resource update are repeated until there are no more resource feasibility candidate actions that can be added. When the activation index of the combinatorial optimization reaches the preset activation threshold, the risk-constrained combinatorial optimization method is adopted. The expected risk reduction and irreversible wear mitigation are determined based on the differences in component risk and irreversible wear state before and after the execution of each candidate action. The comprehensive utility of the candidate action is determined based on the expected risk reduction, irreversible wear mitigation, action cost, candidate priority and resource consumption. Based on the component risk after each candidate action is executed, determine the prediction system risk after the combination of candidate actions is executed; When there are candidate action combinations where the predicted system risk is no higher than the preset system risk upper limit, the risk constraint is that the predicted system risk is no higher than the preset system risk upper limit. The combination optimization objective is to improve the overall risk reduction effect and irreversible wear mitigation effect of the selected candidate actions, and reduce the overall action cost and resource consumption. System-level joint maintenance actions are determined within the resource feasible domain. When there are no candidate action combinations that predict system risk no higher than the preset system risk upper limit, the priority optimization objective is to minimize the excess of the predicted system risk relative to the preset system risk upper limit. Among the candidate action combinations with the same or different risk excess amounts that do not exceed the preset tolerance, the system-level joint maintenance action is determined based on the overall utility of the candidate actions.

[0022] Preferably, the method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine further includes: offline calibration of the parameters of the combined action decoder using agent-assisted evolutionary optimization; generating maintenance task data based on system-level joint maintenance actions; outputting the data to the wind turbine operation and maintenance management terminal to generate maintenance work orders; and updating component status and maintenance resources based on the actual execution results of the maintenance tasks.

[0023] A joint maintenance action generation system for risk constraints of multiple components of a wind turbine, based on the aforementioned method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine, includes: a data acquisition module, a belief state construction module, a risk calculation module, an action tendency generation module, a candidate action generation module, a maintenance action consistency correction module, a resource feasible domain construction module, an adaptive combination decoding module, a maintenance task output module, and an execution feedback update module. The data acquisition module is used to acquire operation monitoring data, historical status data, historical maintenance data, and current maintenance resource data of multiple components of the wind turbine that need to be maintained. The belief state construction module is used to construct the belief states of multiple components of a wind turbine based on the acquired data. The risk calculation module is used to establish an observable maintenance decision model for risk constraints, and to determine the component risk, system risk and maintenance cost based on the belief state of each component to be maintained. The action tendency generation module is used to input the belief states of multiple components into a pre-trained offline policy network, and obtain the action tendency values ​​corresponding to each component to be maintained through inspection actions, predictive maintenance actions, and replacement action output channels, respectively. The candidate action generation module is used to determine the candidate priority of various maintenance actions based on the action tendency value, component risk, belief status and maintenance action cost of each component to be maintained, and to generate a set of candidate actions. The maintenance action consistency correction module is used to perform maintenance action consistency correction on predictive maintenance candidate actions, and redetermine the action mutual exclusion relationship and resource requirements of the corrected replacement candidate actions; The resource feasible region construction module is used to construct the resource feasible region based on action mutual exclusion constraints, inspection resource constraints, maintenance resource constraints, failure component maintenance constraints, and action mask. The adaptive combination decoding module is used to determine the activation index of combination optimization based on system risk, the degree of conflict of candidate actions and the degree of maintenance resource tension, and to use the combination action decoder to adaptively adopt the item-by-item resource projection or risk constraint combination optimization method to generate system-level joint maintenance actions within the resource feasible domain. The maintenance task output module is used to generate maintenance task data based on system-level joint maintenance actions and output it to the wind turbine operation and maintenance management terminal to generate maintenance work orders. The execution feedback update module is used to obtain the actual execution results of maintenance work orders and update the status data and current maintenance resource data of the corresponding components to be maintained.

[0024] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for generating joint maintenance actions for risk constraints of multiple components of wind turbine units, which has the following beneficial effects: By separating the generation of component-level action tendencies from the generation of system-level joint maintenance actions, the offline policy network generates the inspection, predictive maintenance, and replacement action tendencies of each component only through independent action output channels. The final maintenance actions are determined by the subsequent correction and combined decoding process, which can avoid the expansion of action space and unexecutable actions caused by the policy network directly outputting high-dimensional system-level joint actions. It can characterize recoverable degradation state and irreversible wear state respectively, and based on the different effects of predictive maintenance and replacement actions on the two types of states, it can replace and correct the predictive maintenance candidate actions for high-risk and highly irreversible wear parts, which can reduce inefficient predictive maintenance and residual risks after maintenance under irreversible damage conditions. After predictive maintenance candidate actions are converted into replacement candidate actions, the mutual exclusion relationship of actions and resource requirements are redefined. For replacement candidate actions that cannot be executed temporarily due to insufficient resources, the replacement mark and recommended execution time limit are retained, and they are not restored to predictive maintenance candidate actions, so as to maintain the consistency between maintenance actions and component degradation mechanisms. Based on system risk, the degree of conflict among candidate actions, and the tightness of maintenance resources, the method of adaptively selecting item-by-item resource projection or risk constraint combination optimization can take into account both the solution efficiency under low conflict state and the joint resource coordination capability under high risk and strong resource competition state. By constructing a resource feasible domain through mutual exclusion constraints of actions, limited maintenance resource constraints, failure component repair constraints, and action masks, it is possible to transform component-level action tendencies into resource-feasible system-level joint maintenance actions, and further form maintenance tasks that include component identification, action type, execution priority, resource requirements, and recommended execution time limits. The strategy network is trained based on offline maintenance trajectory, and the parameters of the combined action decoder are calibrated before deployment. During the actual deployment phase, the strategy network parameters and decoder parameters are fixed, eliminating the need for online strategy trial and error on actual wind turbine units. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of a method for generating joint maintenance actions for multiple components of a wind turbine, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the multi-component belief state construction method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the pre-training of the offline policy network provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the offline strategy network structure provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of predictive maintenance-replacement consistency correction provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the resource feasible domain construction and adaptive combination decoding process provided in an embodiment of the present invention; Figure 7 This is a comparison diagram of the overall maintenance effect of the present invention and the comparative method provided in the embodiments of the present invention; Figure 8 This is a schematic diagram illustrating the generation result of joint maintenance actions for a single maintenance decision cycle provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 This invention discloses a method for generating joint maintenance actions for multiple components of a wind turbine under risk constraints, such as... Figure 1 This includes the following steps: S1. Obtain operation monitoring data, historical status data, historical maintenance data, and current maintenance resource data of multiple components of the wind turbine to be maintained, and construct the multi-component belief status of the wind turbine; S2. Establish a risk-constrained partially observable maintenance decision model, and determine the component risk, system risk, and maintenance cost based on the belief state of each component to be maintained; S3. Input the belief states of multiple components into a pre-trained offline policy network, and obtain the action tendency values ​​corresponding to each component to be maintained through the inspection action, predictive maintenance action, and replacement action output channels, respectively. S4. Determine the candidate priorities for each type of maintenance action based on the action tendency value, component risk, belief status, and maintenance action cost of each component to be maintained, and generate a set of candidate actions; S5. Perform maintenance action consistency correction on the predictive maintenance candidate actions, and redetermine the action mutual exclusion relationship and resource requirements of the corrected replacement candidate actions; S6. Construct the resource feasible region based on action mutual exclusion constraints, inspection resource constraints, maintenance resource constraints, failure component maintenance constraints, and action mask; S7. Determine the activation index for combined optimization based on system risk, the degree of conflict of candidate actions, and the degree of tension of maintenance resources, and use the combined action decoder to adaptively adopt the item-by-item resource projection or risk constraint combined optimization method to generate system-level joint maintenance actions within the resource feasible domain.

[0029] In this embodiment, the components to be maintained include gearboxes, generators, generator bearings, main bearings, pitch systems, yaw systems, hydraulic systems, and converters. Operational monitoring data is obtained from wind turbine data acquisition and monitoring control systems, condition monitoring systems, or wind farm centralized control systems, including at least one of wind speed, active power, generator speed, main shaft speed, ambient temperature, gearbox oil temperature, gearbox bearing temperature, generator winding temperature, generator bearing temperature, main bearing temperature, vibration characteristics, pitch current, yaw current, hydraulic pressure, and converter temperature. Historical status data includes at least one of component historical degradation characterization values, historical inspection observation results, alarm records, fault records, failure states, and historical operating conditions. Historical maintenance data includes at least one of lubrication, cleaning, adjustment, tightening, parameter calibration, general repair, core sub-component replacement, overall component replacement, maintenance execution time, and state recovery amount before and after maintenance. Current maintenance resource data includes at least one of the following: number of currently available inspection personnel, number of inspection equipment, number of maintenance teams, status of maintenance equipment, status of lifting equipment, available maintenance time window, spare parts type, and spare parts quantity.

[0030] In this embodiment, to unify the numerical scale of different physical quantities, the operation monitoring variables are converted into normalized deviations of the corresponding operating conditions through the normal state benchmark and scale parameters before entering the degradation state calculation. The degradation characterization value, recoverable degradation state, irreversible wear state, state uncertainty, component risk and system risk formed by the normalized deviation of the operating conditions are all represented by dimensionless quantities. The maintenance cost is represented by normalized cost when used for offline strategy training, candidate action scoring and comparison of the effects of different maintenance strategies, and is a dimensionless quantity. Resource constraints such as the number of inspection personnel, the number of maintenance teams, the number of equipment and spare parts are measured according to the corresponding resource quantities, and the maintenance time is expressed by the number of maintenance decision cycles or the actual time unit. When physical quantities with different dimensions participate in weighted fusion, the corresponding physical quantities are normalized first.

[0031] To further implement the above technical solution, step S1, the multi-component belief state of the wind turbine includes the recoverable degradation state, irreversible wear state, state uncertainty, degradation parameter estimate, observation time, failure state, operating condition and remaining maintenance resources of each component to be maintained; The specific content of constructing the multi-component belief state of a wind turbine is as follows: Based on at least one operating condition variable in the operation monitoring data, determine the operating condition unit to which the component to be maintained belongs in the current maintenance decision cycle, and obtain the normal operating sample of the corresponding component to be maintained in the corresponding operating condition unit to determine the normal state benchmark of the operation monitoring variable. Based on the normalized deviation of the operating conditions of each operating monitoring variable of the component to be maintained, determine the current degradation characterization value of the component to be maintained; The normalized deviation of the operating condition is expressed as:

[0032] Among them, the i-th component to be maintained has One state monitoring variable, Let be the observed value of the j-th condition monitoring variable of the i-th component to be maintained during maintenance decision period t. For the current operating conditions, This serves as the baseline for the normal operating condition. For scale parameters, To prevent positive numbers with a denominator of zero, t is the maintenance decision period; The degradation characteristic value of the i-th component to be maintained is:

[0033] in, Let be the total number of condition monitoring variables included in the i-th component to be maintained, and j be the condition monitoring variable number. , The fusion weight corresponding to the j-th state monitoring variable satisfies:

[0034] in, For deviation mapping function, This means limiting the results to a preset range. This represents the upper limit of the degradation characteristic value; The relationship between the degradation characterization value of the i-th component to be maintained and the observed reversible degradation state and irreversible wear state is expressed as follows:

[0035] in, Let the residual characterize the degradation state of the i-th component to be maintained during maintenance decision cycle t. To be in a reversible degenerative state, Irreversible wear is defined as a reversible degradation state, which characterizes a state shift that can be partially recovered by at least one predictive maintenance operation, such as lubrication, cleaning, adjustment, tightening, parameter calibration, or component repair. Irreversible wear is defined as cumulative damage caused by accumulated loads, fatigue damage, long-term wear, or random impacts that cannot be eliminated by predictive maintenance operations. It should be noted that recoverable degradation state and irreversible wear state are hidden states to be estimated, and are not uniquely algebraically decomposed based on the degradation characterization value of a single maintenance decision cycle. Based on the current degradation characterization value, the state estimate of the previous maintenance decision cycle, the current operating conditions, the predictive maintenance action execution record, the component replacement record, and the state recovery amount before and after maintenance, the recursive state estimation method is used to determine the recoverable degradation state and irreversible wear state of the current maintenance decision cycle. The recoverable degradation state and the irreversible wear state together form the degradation state characterization of the component to be maintained. In this embodiment, reversible degradation state and irreversible wear state are quantified using a normalized degradation scale consistent with the current degradation characterization value; reversible degradation state is quantified according to... Recursively determined, irreversible wear state according to The process is determined recursively, where, To maintain the decision-making cycle Recoverable degraded state after performing maintenance actions To maintain the decision-making cycle to The recoverable degradation increment between them The increment of irreversible wear within the corresponding maintenance decision cycle; the increment of recoverable degradation and the increment of irreversible wear are determined according to the state recursive relationship in the subsequently constructed multi-component degradation-maintenance environment; The state uncertainty is recursively determined based on the number of maintenance decision cycles since the most recent effective inspection and the status observation update after the effective inspection is performed. In this embodiment, state uncertainty is used to characterize the decrease in the reliability of state estimation due to the lack of continuous and effective inspection observations. During a maintenance decision cycle in which no effective inspections are performed, the [missing information] is... The uncertainty of the condition of a component to be maintained is according to Recursion, where, For the first Each component requiring maintenance is in the maintenance decision cycle. The uncertainty of the state is characterized by the standard deviation of the normalized state estimation error. For the first The variance growth parameter of the state uncertainty of a component to be maintained when no effective inspection observations are obtained within a maintenance decision cycle: When the After each component to be maintained undergoes a valid inspection, the state estimate is updated based on the state observations obtained from the inspection, and the state uncertainty is reduced. ,in, Due to the uncertainty of the state after effective inspection, This represents the shrinkage coefficient of the state uncertainty after the inspection; since the state estimation is based on a normalized degradation scale, therefore... , and All are expressed using the corresponding normalized scale and are dimensionless quantities; and The calibration and determination are based on the variation pattern of state prediction error in historical inspection samples; Based on the deviation between the actual change in the condition of the component to be maintained and the preset baseline degradation rate, the estimated degradation parameters are recursively updated. ; The observation timeframe is determined based on the time interval between the current moment and the most recent valid inspection moment; The time since the most recent effective inspection for the i-th component to be maintained is represented as:

[0036] in, The most recent valid inspection time for the i-th component to be maintained; The failure state is determined based on the relationship between the current degradation characterization value of the component to be maintained and the corresponding failure boundary.

[0037] in, Let i be the failure boundary of the i-th component to be maintained. This is an indicator function.

[0038] In this embodiment, during the maintenance decision period t, the belief states of the multiple components of the wind turbine are represented as follows:

[0039] in, This refers to the recoverable degradation state of each component requiring maintenance. This refers to the irreversible wear condition of each component requiring maintenance. Due to the uncertainty of the condition of each component to be maintained, For the estimated degradation parameters of each component to be maintained, For the observation time of each component to be maintained, This refers to the failure state of each component requiring maintenance. For the current operating conditions, This represents the remaining maintenance resources.

[0040] To further implement the above technical solution, the specific content of constructing the multi-component belief state of the wind turbine also includes: After performing the inspection action, the recoverable degradation state, degradation parameter estimates, and state uncertainty of the corresponding component to be maintained are updated based on the status observations obtained from the inspection. After performing predictive maintenance actions, the recoverable degradation state of the corresponding component to be maintained is reduced according to a preset recovery ratio or a recovery ratio determined based on historical maintenance results, while the irreversible wear state remains unchanged. After performing predictive maintenance actions, the degraded state can be restored to an updated state. The irreversible wear state remains unchanged. ,in, Let be the predictive maintenance recovery ratio for the i-th component to be maintained, and satisfy . ; After performing the replacement, the recoverable degradation state and irreversible wear state of the corresponding component to be maintained will be updated to the initial state after replacement. ,in, and The initial state after the change can be set to 0 or a preset initial value close to 0 in the normalized implementation; During maintenance decision cycles in which no effective inspections are performed, the state uncertainty of the corresponding component to be maintained is increased recursively according to the preset state uncertainty growth parameter.

[0041] In one eight-component embodiment, the multi-component belief state includes the recoverable degradation state, irreversible wear state, state uncertainty, degradation parameter estimate, observation timeliness, most recently observed valid identifier, and global states such as operating conditions and remaining maintenance resources for each component, forming a total of 56-dimensional state input. The construction process of the multi-component belief state is as follows: Figure 2 As shown.

[0042] In this embodiment, the multi-component maintenance decision problem of wind turbines is represented as a risk-constrained partially observable Markov decision process, that is, a risk-constrained partially observable maintenance decision model is established:

[0043] in, For a multi-component belief state space, To jointly maintain the action space, For state transition mechanism, For immediate evaluation functions, For the system risk function, For the resource feasible domain, As a discount factor, To maintain the planning time domain; The maintenance action for the i-th component to be maintained in maintenance decision cycle t is represented as follows:

[0044] in, , and These respectively represent inspection actions, predictive maintenance actions, and replacement actions; The three types of actions for the same component to be maintained within the same maintenance decision cycle satisfy the action mutual exclusion constraint:

[0045] Inspection resource constraints are represented as follows:

[0046] Maintenance resource constraints are represented as follows:

[0047] in, , and These represent the resource requirements for the inspection, predictive maintenance, and replacement actions corresponding to the i-th component to be maintained. Currently available inspection resources, Currently available maintenance resources; When the corresponding implementation method sets spare parts resource constraints, it also satisfies:

[0048] in, The demand for spare parts of type k is calculated to replace the i-th component to be maintained. This represents the current available quantity of the k-th type of spare parts, where k is the spare parts category number. This represents the total number of spare parts types.

[0049] To further implement the above technical solution, in step S2, the component risk value of each component to be maintained is:

[0050] in, Let i be the component risk value of the i-th component to be maintained. For risk smoothing parameters, Let be the overall degradation amount of the i-th component to be maintained. The closer the overall degradation amount is to or exceeds the failure boundary, the higher the risk value of the component. Let t be the failure boundary of the i-th component to be maintained, t be the maintenance decision cycle, and i be the component number to be maintained.

[0051] in, Risk contribution parameters corresponding to irreversible wear conditions The parameter contributes to the risk corresponding to state uncertainty. To be in a reversible degenerative state, It is in an irreversible wear state. The state is uncertain; System risks of wind turbines for:

[0052]

[0053]

[0054] in, Risks associated with the aggregation of key components To mitigate the risks associated with component aggregation, A collection of key components. To support the assembly of components, For the assembly of supporting components The number of components to be maintained included. To support the high-risk assessment threshold of the components, the key components include at least one of the following: gearbox, generator, main bearing, and generator bearing; and the supporting components include at least one of the following: pitch system, yaw system, hydraulic system, and converter. Maintenance cost C per maintenance decision cycle t for:

[0055] in, , , , and These are inspection costs, predictive maintenance costs, replacement costs, failure event costs, and failure-held costs. , and These are inspection actions, predictive maintenance actions, and replacement actions; To add an invalidation flag, It is in a failed and unrepaired state. For downtime loss parameters, Losses due to downtime of critical components For joint scheduling cost parameters, To cover component scheduling costs, N represents the number of components to be maintained.

[0056] To further implement the above technical solution, step S3, as follows: Figure 3 The specific content of the pre-trained offline policy network is as follows: By using historical operating data of wind turbine units, the baseline degradation rate, operating condition transition patterns, probability of random impact, intensity of random impact, and degradation coupling relationship between components are calibrated for each component to be maintained, and a multi-component degradation-maintenance environment is constructed. In this embodiment, the operating condition transition rule is as follows:

[0057] in, To maintain the operating conditions during the decision-making cycle t, To maintain the decision-making cycle Operating conditions To be based on the current operating conditions Transfer to the next operating condition The conditional transition probability and operating condition transition law are determined by calibration based on historical operating data of the wind turbine. The operating conditions are determined through degradation acceleration factors. It affects the degradation rate of the i-th component to be maintained; The random impact event is indicated as follows:

[0058] in, To maintain the decision-making cycle A 0-1 random variable indicating whether a random shock will occur. This indicates that a random shock has occurred during the current maintenance decision-making cycle. This indicates that no random shock has occurred. Current operating conditions The probability of a random shock occurring; The amplitude of a random impact is expressed as:

[0059] in, This represents the impact amplitude corresponding to the occurrence of a random impact. This represents the random impact amplitude distribution under the current operating conditions; the probability of random impact occurrence and the random impact amplitude distribution are determined by calibration based on the characteristics of impact events in historical operating data. The degenerative coupling relationship of multiple components is represented by the coupling matrix as follows:

[0060] in, Indicates the first The intensity of the degradation impact of each component to be maintained on the i-th component to be maintained. Calibration is determined based on the degradation coupling relationships between components in historical operating data; The coupling degradation term of the i-th component to be maintained Represented as:

[0061] in, To exclude the first Other parts to be maintained besides the one to be maintained. For the first Each component requiring maintenance is in the maintenance decision cycle. The reversible degenerate state; Irreversible wear condition according to The incremental irreversible wear is expressed recursively as follows:

[0062] in, Let be the random impact sensitivity coefficient of the i-th component to be maintained. , , and Wear contribution parameters corresponding to the factors Noise from the wear process, For the i-th component to be maintained in operating condition The degradation acceleration factor below; Recoverable degenerative state according to Recursively, the reversible degradation increment is represented as:

[0063] in, Let i be the baseline degradation rate of the i-th component to be maintained. This is an estimate of the degradation rate. This is the coupling amplification factor. This is the irreversible wear amplification factor. This is noise from the degradation process; When the component to be maintained performs an inspection, a status observation with observation errors is obtained: ,in, To account for inspection observation errors; based on the deviation between the inspection observations and the predicted state, update the degradation ratio estimate:

[0064] in, This refers to the reversible degradation state predicted before the inspection. Update the step size for degradation parameters; When a component fails to undergo effective inspection, the state uncertainty is determined according to... Increase, among which, Let be the uncertainty growth parameter for the i-th component to be maintained; after performing the inspection, the state uncertainty is updated as follows: After performing predictive maintenance, the recoverable degradation state is reduced according to the predictive maintenance status update relationship, while the irreversible wear state remains unchanged; after performing the replacement action, the recoverable degradation state and the irreversible wear state are updated to the initial state after the corresponding maintenance component is replaced.

[0065] To avoid online strategy trial and error using actual wind turbine units, a full offline maintenance trajectory dataset is constructed by interacting with the degradation-maintenance environment using a hybrid behavior strategy. The hybrid behavior strategy includes at least two of the following behavior modes: conservative maintenance, balanced maintenance, reactive maintenance, delayed maintenance, under-maintenance, proactive replacement, failure exposure, and random exploration. Each transition sample includes the current multi-component belief state, the corresponding maintenance action, the immediate evaluation value, the next multi-component belief state, the maintenance trajectory termination identifier, and the action mask. The constructed hybrid behavior strategy is represented as follows:

[0066]

[0067] Where q is the behavioral strategy category number. This represents the total number of behavioral strategy categories in the hybrid behavioral strategy library. For the q-th type of behavioral strategy, The sampling weights are for the q-th type of behavior strategy; A single-step transfer sample is represented as:

[0068] in, For offline transfer sample numbering, For the first The maintenance decision cycle of an offline transfer sample in its respective maintenance trajectory; Given the current state of belief across multiple components, For joint maintenance operations, For immediate evaluation value, For the next multi-component belief state, To maintain the trajectory termination marker, For action mask; To introduce system risk constraints and new failure constraints during offline policy network training, the instantaneous evaluation value is represented as:

[0069]

[0070] in, To maintain the number of newly added failures of all components to be maintained within the maintenance decision period t, the system is configured according to the newly added failure identifier of each component to be maintained. Summation yields the result; This represents the upper limit of system risk constraints. To add an upper limit to the failure constraint, and These are the penalty weights for exceeding risk limits and exceeding new failure limits, respectively. ; The full offline maintenance trajectory dataset is represented as follows: ,in, This represents the number of samples transferred offline. The offline maintenance trajectory is divided into policy network training data, combined action decoder parameter calibration data, and independent test data. In this embodiment, the policy network training data is used for offline policy network pre-training; the combined action decoder parameter calibration data is used to perform offline calibration of the combined action decoder parameters after the offline policy network has completed pre-training and the network parameters have been fixed; the independent test data is only used for the final maintenance performance evaluation after both the offline policy network parameters and the combined action decoder parameters have been fixed, and does not participate in the offline policy network training and combined action decoder parameter selection. Based on the training data of the policy network, the cumulative discounted return of each transfer sample is calculated according to the real-time evaluation values ​​of each transfer sample from the corresponding maintenance decision cycle to the end maintenance decision cycle of the maintenance trajectory. The cumulative discounted return of the nth offline transfer sample from its corresponding maintenance decision period to the end maintenance decision period of its respective maintenance trajectory is expressed as:

[0071] in, For the cumulative discount return corresponding to the nth offline transfer sample, Let u be the maintenance decision cycle of the sample within its respective maintenance trajectory, and u be the index for summing the maintenance decision cycles in the cumulative discount return calculation. This refers to the termination maintenance decision cycle for the maintenance trajectory to which the nth sample belongs. To maintain the real-time evaluation value of decision cycle u, The discount factor is used because the real-time evaluation value consists of negative maintenance costs, system risk exceeding limits penalty, and new failure exceeding limits penalty. A larger cumulative discount corresponds to an offline transfer sample with better overall maintenance performance. The sample weights are determined and truncated based on the difference between the cumulative discounted return of each transfer sample and the average return of the samples in the offline maintenance trajectory dataset. The sample weights are:

[0072] in, The average cumulative return for offline sample discounts. For temperature parameters; To prevent a small number of extremely high-reward samples from dominating the training of the offline strategy network, the sample weights are truncated:

[0073] in, This represents the upper limit of sample weights; The offline policy network is trained using the belief states of multiple components as input, the inspection actions, predictive maintenance actions and replacement actions corresponding to each component to be maintained as training labels, and the reward-weighted multi-action binary cross-entropy loss. The offline policy network is represented as:

[0074] in, For offline policy network parameters, , and These represent the action tendency values ​​for performing inspection, predictive maintenance, and replacement actions on the i-th component to be maintained, respectively. The three types of action tendencies use independent binary output channels and do not directly constitute the final action of the three-way selection for the same component to be maintained. The same component to be maintained can have multiple higher action tendency values ​​at the same time. The action tendency values ​​are only used to generate candidate actions. The reward-weighted multi-action binary cross-entropy loss is:

[0075] in, Weighted multi-action binary cross-entropy loss for the reward of the offline policy network; The number of transfer samples used for training the offline policy network; Number the training samples; For the action channel category, and ,in , and These represent the action channels for routine inspection, predictive maintenance, and replacement, respectively. For the first In the offline transfer sample, the first Each component to be maintained corresponds to an action channel. Action tags, This is the corresponding action tendency value output by the offline policy network; During training and motion generation, motion masks are used to shield maintenance actions that violate maintenance constraints for failed components or do not meet the conditions for motion execution.

[0076] in, After the action mask is applied, the first In the offline transfer sample, the first Each component to be maintained corresponds to an action channel. Effective action tendency value; For the corresponding action mask; This is the original action tendency value output by the offline policy network; when the corresponding maintenance action violates the repair logic of the failed component or does not meet the action execution conditions, it is set to... When the conditions for action execution are met, let ; After training, the offline policy network parameters are fixed, and the offline policy network is not updated through online trial and error on actual wind turbines. The offline policy network includes a state input layer, a shared state feature extraction layer, and action output channels corresponding to inspection actions, predictive maintenance actions, and replacement actions, respectively. Each action output channel outputs action tendency values ​​corresponding to multiple components to be maintained. The shared state feature extraction layer can adopt a fully connected network, a temporal coding network, or other network structures that can extract state features of multiple components.

[0077] In one eight-component embodiment, the offline policy network input is a 56-dimensional multi-component belief state, and the output is 24 action tendency values. Eight outputs correspond to inspection actions, eight outputs correspond to predictive maintenance actions, and eight outputs correspond to replacement actions. Figure 4 .

[0078] In this embodiment, the hybrid behavior strategy is only used for offline maintenance trajectory generation and does not participate in maintenance decisions during the actual testing or deployment phases. Specifically, the conservative maintenance mode is used to increase the frequency of inspections and predictive maintenance; the balanced maintenance mode is used to maintain a moderate level of intervention between maintenance costs and system risks; the reactive maintenance mode is used to trigger maintenance after a significant increase in component risk or failure; the delayed maintenance mode is used to generate trajectories for delayed execution of maintenance tasks; the under-maintenance mode is used to cover high-risk and insufficient maintenance resource states; the proactive replacement mode is used to increase the sample of replacement actions under high irreversible wear states; the failure exposure mode is used to cover new failures and unrepaired failure states; and the random exploration mode is used to increase the diversity of state-action combinations.

[0079] To further implement the above technical solution, the specific content of step S4 is as follows: The priority of inspection actions is determined based on the inspection action tendency value, state uncertainty, observation timeliness, component risk value, and inspection action cost. The priority of inspection action candidates increases with the increase of state uncertainty, observation timeliness, and component risk value, and decreases with the increase of inspection action cost. The priority of the inspection action candidate for the i-th component to be maintained is represented as follows:

[0080] in, For the normalized observation time, To normalize the inspection cost, each For the corresponding scoring weights; The candidate priority of predictive maintenance actions is determined based on the predictive maintenance action tendency value, recoverable degradation state, component risk value, irreversible wear state, and predictive maintenance action cost. The candidate priority of predictive maintenance actions increases with the increase of recoverable degradation state and component risk value, and decreases with the increase of irreversible wear state and predictive maintenance action cost. The candidate priority of the predictive maintenance action for the i-th component to be maintained is represented as follows:

[0081] in, This represents the normalized predictive maintenance cost. The candidate priority of replacement actions is determined based on the replacement action tendency value, component risk value, irreversible wear state, failure state, and replacement action cost. The candidate priority of replacement actions increases with the component risk value, irreversible wear state, and failure severity, and decreases with the replacement action cost. The priority of the replacement action for the i-th component to be maintained is represented as follows:

[0082] in, The normalized replacement cost; The candidate priorities of each type of maintenance action are compared with the corresponding candidate trigger thresholds, and maintenance actions that meet the corresponding candidate trigger conditions and whose action masks are in an executable state are added to the candidate action set. The candidate action set is as follows:

[0083] in, To maintain the decision-making cycle The set of candidate actions; The part number to be maintained. The number of parts to be maintained; For the action channel category, ,in, , and These represent the inspection action channel, the predictive maintenance action channel, and the replacement action channel, respectively. For the first Each component requiring maintenance is in the maintenance decision cycle. Corresponding action channel Candidate priority, For action channel The corresponding candidate trigger threshold, For the corresponding action mask, This indicates that the maintenance action is in an executable state. The candidate action set only represents maintenance actions that meet the corresponding candidate triggering conditions and are in an executable state under the current state. They do not directly constitute the final action of the same component to be maintained, nor do they directly constitute system-level joint maintenance actions.

[0084] In this embodiment, step S5 performs maintenance action consistency correction on the predictive maintenance candidate actions in the candidate action set to identify situations where the physical action of the predictive maintenance action does not match the current degradation state of the corresponding component to be maintained, such as... Figure 5 As shown; For the predictive maintenance candidate action of the i-th component to be maintained that has entered the candidate action set, when the component risk of the component to be maintained satisfies And its irreversible wear state satisfies When the predictive maintenance candidate action is determined to be inconsistent with the current degradation state of the component to be maintained, predictive maintenance-replacement correction is performed.

[0085] in, As the risk correction threshold, The threshold for irreversible wear correction. To provide candidate actions for performing predictive maintenance on the i-th component to be maintained, To perform a replacement candidate action for the i-th component to be maintained, \ denotes the set difference operation. This indicates that the current set of candidate actions will be updated and assigned values. This represents the set union operation; After completing the consistency correction of the maintenance action, the corresponding converted predictive maintenance candidate action is deleted from the current candidate action set, and the newly generated replacement candidate action is added to the replacement candidate action set of the corresponding component to be maintained.

[0086] Because predictive maintenance actions and replacement actions have different action costs, resource requirements, and execution conditions, the following processing is re-executed after generating replacement candidate actions: Re-examine the mutual exclusion relationships between replacement candidate actions and inspection candidate actions, predictive maintenance candidate actions, and other maintenance candidate actions corresponding to the same component to be maintained; Based on at least one of the following: maintenance personnel requirements, maintenance time requirements, lifting equipment requirements, and spare parts requirements corresponding to the replacement candidate actions, the resource requirements for the replacement candidate actions are re-determined. Based on the redefined mutual exclusion relationships of actions and resource requirements, determine whether the replacement candidate action can enter the resource feasible domain corresponding to the current maintenance decision cycle; When the total resource requirement of multiple replacement candidate actions exceeds the current available maintenance resources, the overall utility of replacement is determined based on the expected risk reduction, irreversible wear mitigation, and replacement cost corresponding to each replacement candidate action. Replacement candidate actions that can enter the resource feasible domain within the current maintenance decision cycle are determined in descending order of overall replacement utility. For replacement candidate actions that cannot enter the resource feasible domain of the current maintenance decision cycle due to insufficient current maintenance resources, the pending replacement flag of the component to be maintained is retained, and the recommended execution time limit is determined based on at least one of the component risk value, irreversible wear status, predicted risk growth trend, and subsequent available maintenance resources. This embodiment redefines the mutual exclusion relationship of actions and the processing of resource requirements in order to avoid directly using the lower resource requirements corresponding to the original predictive maintenance candidate actions to the replacement candidate actions, which would result in an underestimation of the resource requirements of the replacement task or the inability to execute system-level joint maintenance actions. When the corresponding implementation also sets spare parts resource constraints, lifting equipment constraints, or maintenance time window constraints, replacement candidate actions that cannot be executed in the current maintenance decision cycle due to insufficient spare parts resources, lifting equipment, or maintenance time window will still retain their pending replacement flag and recommended execution time limit; replacement candidate actions that are temporarily suspended due to insufficient resources will not be restored as predictive maintenance candidate actions, so as to avoid reallocating predictive maintenance actions with mismatched physical effects when it has been determined that predictive maintenance cannot adequately mitigate the risk of irreversible wear.

[0087] In this embodiment, the candidate actions after completing the consistency correction of maintenance functions are sequentially denoted as follows: ,in, , Set a binary selection variable to represent the current number of candidate actions:

[0088] The resource feasible region for the current maintenance decision cycle is represented as follows:

[0089] in, To maintain the decision-making cycle The resource feasible domain; This represents the number of candidate actions in the current maintenance decision-making cycle after maintenance action consistency correction. Number the candidate actions; Select vectors for candidate actions, where, Indicates the selection of the first Candidate actions , This indicates that the candidate action is not selected; For the candidate action resource demand matrix, its first... List corresponding candidate actions Resource requirements; This represents the available resource vector for the current maintenance decision-making cycle. This is the mutual exclusion constraint matrix for candidate actions, used to characterize the mutual exclusion relationship between candidate actions that cannot be executed simultaneously; A vector of all 1s that matches the number of mutual exclusion constraints; Candidate actions The corresponding action mask, This indicates that the candidate action meets the execution conditions. This indicates that the candidate action is not executable; Candidate Action Resource Requirement Matrix It includes at least inspection resource requirements and maintenance resource requirements. When the corresponding implementation method sets constraints on spare parts, lifting equipment or maintenance time windows, the resource requirement matrix further includes the corresponding resource requirements. The mutual exclusion matrix is ​​used to limit the same component to be maintained to perform at most one maintenance action within the same maintenance decision cycle, and to limit the candidate actions that compete for the same non-parallel equipment or occupy conflicting maintenance time windows to not be executed at the same time.

[0090] To further implement the above technical solution, step S6, as follows: Figure 6 The specific content of the resource feasible domain is as follows: Within the same maintenance decision cycle, the same component to be maintained may be assigned at most one of the following actions: inspection, predictive maintenance, and replacement. The inspection resources consumed by the assigned inspection actions, including inspection personnel, inspection equipment, or inspection time, shall not exceed the available inspection resources corresponding to the current maintenance decision cycle. The maintenance resources consumed by the assigned predictive maintenance actions and replacement actions, including maintenance personnel, maintenance equipment, or maintenance time, shall not exceed the available maintenance resources corresponding to the current maintenance decision cycle. The number and type of spare parts required for the assigned replacement action shall not exceed the current number and type of spare parts available. Failed components are not assigned predictive maintenance actions and are placed in either the replacement candidate queue or the corrective maintenance candidate queue. All maintenance actions entering the resource feasible domain satisfy the corresponding action mask; When maintenance actions for different components require the use of the same non-parallel maintenance equipment or occupy conflicting maintenance time windows, the corresponding maintenance actions should be set to not be executed simultaneously.

[0091] To further implement the above technical solution, in step S7, the specific content of generating system-level joint maintenance actions using the combined action decoder is as follows: The system risk, the degree of action conflict, and the current maintenance resource shortage are normalized respectively, and then weighted and merged based on the corresponding indicator weights to obtain the combined optimized activation indicators; The degree of action conflict is determined based on the number of multiple candidate actions corresponding to the same component to be maintained, the degree to which candidate actions of different components to be maintained compete for the same maintenance resources, and the degree of maintenance time conflict between candidate actions; the current level of maintenance resource shortage is determined based on the relative relationship between the total resource demand of the candidate action set and the currently available maintenance resources. The degree of action conflict is expressed as follows:

[0092] in, This refers to the number of conflicts among multiple candidate actions for the same component to be maintained. This represents the number of resource competitions among candidate actions. To maintain the number of time window conflicts, , and For the corresponding weights; The level of resource scarcity is expressed as follows:

[0093] in, The total inspection resource requirements for the current set of candidate actions. To meet the overall maintenance resource requirements, Currently available inspection resources, Currently available maintenance resources; When the corresponding implementation also sets spare parts resource constraints, the degree of maintenance resource shortage is expressed as follows:

[0094] in, Total spare parts requirements for candidate replacement actions. Currently available spare parts resources; The activation indicators for combined optimization are:

[0095] in, For normalized systemic risk, The normalized degree of action conflict, To determine the level of maintenance resource scarcity after normalization, , and For the corresponding indicator weights, ; When the combined optimization activation index does not reach the preset activation threshold hour, The method of projecting resources one by one is adopted, and each candidate action is checked in descending order of priority. If there is no conflict between the current candidate action and the selected action, and the resource feasibility domain is still satisfied after adding the current candidate action, the current candidate action is added to the system-level joint maintenance action, and the remaining inspection resources and remaining maintenance resources are updated. When the corresponding implementation method sets spare parts resource constraints, the remaining spare parts resources are further updated. If adding the current candidate action would cause the same component to be maintained to perform multiple actions at the same time, inspection resources to exceed limits, maintenance resources to exceed limits, failure of the component to perform predictive maintenance, conflict of non-parallel equipment, or conflict of maintenance time windows, the candidate action is skipped. The candidate action check and remaining resource update are repeated until there are no more resource-feasible candidate actions that can be added. When the combined optimization activation index reaches the preset activation threshold The risk-constrained combination optimization method is adopted to determine the expected risk reduction and irreversible wear mitigation based on the differences in component risk and irreversible wear state before and after each candidate action. The comprehensive utility of the candidate action is determined based on the expected risk reduction, irreversible wear mitigation, action cost, candidate priority and resource consumption.

[0096] in, Number the candidate actions, consistent with the aforementioned definition; To execute the first The expected risk reduction after each candidate action. This is to mitigate irreversible wear. To normalize candidate priority, To normalize the action cost, To normalize resource consumption, each The corresponding weights are as follows: For inspection and predictive maintenance actions, the irreversible wear mitigation amount is set to 0; for replacement actions, the irreversible wear mitigation amount is determined based on the difference between the irreversible wear state before replacement and the initial irreversible wear state after replacement; for inspection actions, the expected risk reduction amount can be determined based on the expected reduction in uncertainty of the state after inspection; for predictive maintenance actions, the expected risk reduction amount is determined based on the change in component risk after the reduction of recoverable degradation state; for replacement actions, the expected risk reduction amount is determined based on the change in component risk after the update of recoverable degradation state and irreversible wear state. Based on the component risk after each candidate action is executed, determine the prediction system risk after the combination of candidate actions is executed; The predicted system risk is represented as:

[0097] When there are candidate action combinations where the predicted system risk is no higher than the preset system risk upper limit, the risk constraint is that the predicted system risk is no higher than the preset system risk upper limit. The combination optimization objective is to improve the overall risk reduction effect and irreversible wear mitigation effect of the selected candidate actions, and reduce the overall action cost and resource consumption. System-level joint maintenance actions are determined within the resource feasible domain. The combinatorial optimization objective is expressed as: And satisfy: ,in, To pre-determine the system risk ceiling; when there are multiple candidate action combinations that satisfy the system risk ceiling in the resource feasible domain, the system-level joint maintenance action is determined from the candidate action combinations that satisfy the system risk constraints based on the overall utility of the candidate actions; When there are no candidate action combinations that predict system risk no higher than the preset system risk limit due to current maintenance resource constraints, the optimization objective is to minimize the excess of the predicted system risk relative to the preset system risk limit.

[0098] And among the candidate action combinations where the risk exceeds the same amount or the difference does not exceed the preset tolerance, the system-level joint maintenance action is determined based on the overall utility of the candidate actions.

[0099] Combinatorial optimization problems can be solved using one or a combination of integer programming, branch and bound, dynamic programming, heuristic search, and evolutionary optimization. This embodiment forms a continuous processing chain that includes component-level action tendency generation, maintenance action consistency correction, action mutual exclusion relationship and resource requirement redetering, resource feasible domain construction, and risk constraint adaptive combination decoding.

[0100] To further implement the above technical solution, a method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine also includes: offline calibration of the parameters of the combined action decoder using agent-assisted evolutionary optimization, generating maintenance task data based on system-level joint maintenance actions, outputting it to the wind turbine operation and maintenance management terminal to generate maintenance work orders, and updating component status and maintenance resources based on the actual execution results of maintenance tasks.

[0101] In this embodiment, the specific details of offline calibration of the combined action decoder parameters are as follows: The parameter to be calibrated is represented as follows:

[0102] in, This is a risk multiplier parameter used to adjust the relative impact of component risk and system risk in candidate action scoring or decoding evaluation; This is a system risk gating parameter used to adjust the impact of system risk on the combinatorial optimization activation process and to determine the combinatorial optimization activation threshold. ; This is the maximum component risk gating parameter; For state uncertainty gating parameters; For irreversible wear gating parameters; and These are the candidate trigger thresholds for predictive maintenance actions and replacement actions, respectively. This is a rescue trigger time parameter used to control the scheduling conditions for high-risk candidate actions that have not been processed for a long time; For the scoring scale parameters of candidate actions; This sets the upper limit on the number of replacement actions allowed to be performed within the preset maintenance trajectory.

[0103] The parameter search space for the parameter to be calibrated is determined as follows: ,in, and These are the lower and upper limits of each parameter to be calibrated; an initial set of candidate parameters is generated from the parameter search space: ,in, This represents the initial number of candidate parameters; For any candidate parameter In a preset verification scenario, the degradation-maintenance simulation program was actually run to obtain the average maintenance cost, average system risk, average number of new failures, number of predictive maintenance actions, number of replacement actions, and constraint feasibility. The parameters to be evaluated in actual simulation are selected based on the prediction results of the verification target, the prediction results of the constraint feasibility, and the degree of difference between the candidate parameters. The parameter validation objective is represented as:

[0104] in, To average maintenance costs, To average system risk, This represents the average number of new failures. The average number of replacement actions within the preset maintenance trajectory. To verify the upper limit of system risk during the verification phase, To add a failure limit during the verification phase, This sets the upper limit for the number of replacement actions allowed to be performed within the preset maintenance trajectory. , and For the corresponding penalty parameters; A degradation-maintenance simulation is performed on the actual simulation evaluation parameters to obtain the actual simulation evaluation results, and the surrogate model is updated based on the actual simulation evaluation results; Repeat the process of candidate parameter generation, surrogate model prediction, parameter screening for actual simulation evaluation, degradation-maintenance simulation and surrogate model update until the maximum number of iterations is reached or the verification target is not improved after a preset number of consecutive iterations. From the candidate parameters that have been evaluated in actual simulations and meet the system risk constraints, new failure constraints, and replacement action number constraints, select the candidate parameters with the lowest actual maintenance cost or the optimal preset comprehensive evaluation target as the target decoder parameters. The target decoder parameters are fixed during the wind turbine maintenance task generation phase, and the target decoder parameters are not reselected using maintenance performance test data; The final maintenance strategy is expressed as follows:

[0105] in, For the completed offline policy network, The target decoder parameters are determined using parameter calibration data. A risk-constrained adaptive combined action decoder; The generated maintenance task data includes: wind turbine identifier, component identifier to be maintained, maintenance action type, component risk value, recoverable degradation state, irreversible wear state, state uncertainty, execution priority, required maintenance resources, recommended execution time limit and task generation basis, as well as spare parts information when spare parts are required for the corresponding maintenance action; The maintenance task generation criteria are used to characterize the generation of the corresponding maintenance task through one of the following methods: directly retaining candidates based on offline policy network action tendencies and candidate priorities; obtaining candidates based on the comprehensive utility of candidate actions under resource contention conditions; converting predictive maintenance candidate actions into replacement actions after maintenance effect consistency correction; or generating replacement tasks or corrective maintenance tasks from the failure component repair logic.

[0106] Example 2 Based on the method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine in Embodiment 1, this embodiment selects historical operating data of a single wind turbine in a wind farm in Jiuquan, Gansu Province, and models eight types of components to be maintained, including gearbox, generator, generator bearing, main bearing, pitch system, yaw system, hydraulic system and converter. The historical operating data is used to calibrate the component reference degradation rate, operating condition transfer law, random impact characteristics and degradation coupling relationship between components.

[0107] By leveraging a hybrid behavioral strategy and interacting with a calibrated degradation-maintenance environment, 760 offline maintenance trajectories were generated. Each trajectory includes 160 maintenance decision cycles, resulting in a total of 121,600 offline transfer samples. The offline policy network takes a 56-dimensional multi-component belief state as input and generates 24 action tendency values ​​corresponding to eight components to be maintained through independent inspection, predictive maintenance, and replacement action output channels.

[0108] In this embodiment, the state variables, risk variables, action tendency values, candidate scores, and related thresholds are all determined according to the method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine in Embodiment 1. In four independent test scenarios—normal operation, accelerated degradation, resource-constrained operation, and shock coupling—based on real-world operational data calibration, the method of Embodiment 1 of this invention has an average normalized total cost of 2376.55, an average system risk of 0.0466, and an average number of new failures of 0. Each complete test trajectory executes an average of 210.00 component-level inspection actions, 124.01 component-level predictive maintenance actions, and 4.53 component-level replacement actions. The overall maintenance effect of the method of this invention compared to the comparative method is as follows: Figure 7 As shown, compared with the Optimized Risk Opportunity Maintenance strategy (OptRiskOM-GA), the average normalized total cost is reduced by 29.90%, the average system risk is reduced from 0.0832 to 0.0466, and the average number of new failures is reduced from 0.5458 to 0.

[0109] Taking a maintenance decision cycle in an impact coupling test scenario as an example, the results are as follows: Figure 8 As shown, following the state recursion method in Example 1, the normalized recoverable degradation state of the generator is 0.680476, and the normalized irreversible wear state is 0.412614. The component risk value is 0.681148 according to the component risk calculation method. The current multi-component belief state is input into the trained and fixed-parameter offline policy network, resulting in a predictive maintenance action tendency value of 0.406035 and a replacement action tendency value of 0.984794. According to the candidate action priority calculation method, the basic candidate score for predictive maintenance of the generator is 0.917124, and the basic candidate score for replacement is 0.856013. Offline parameter calibration is performed using combined action decoder parameter calibration data, determining the predictive maintenance candidate score threshold to be 0.542034 and the replacement candidate score threshold to be 0.477242.

[0110] Since the predictive maintenance baseline candidate score of 0.917124 is higher than the predictive maintenance candidate score threshold of 0.542034, and the replacement baseline candidate score of 0.856013 is higher than the replacement candidate score threshold of 0.477242, both the predictive maintenance action and the replacement action meet the candidate triggering conditions.

[0111] In this embodiment, offline parameter calibration is performed using combined action decoder parameter calibration data. The risk correction threshold is set to 0.66, and the irreversible wear correction threshold is set to 0.305333. Since the generator's component risk value of 0.681148 is higher than the risk correction threshold of 0.66, and the irreversible wear state of 0.412614 is higher than the irreversible wear correction threshold of 0.305333, continuing to perform predictive maintenance cannot fully eliminate the risk exposure caused by irreversible wear. Therefore, after maintenance action consistency correction, the candidate predictive maintenance action for this generator is excluded, and the candidate replacement action is retained.

[0112] After decoding the mutual exclusion of actions and maintenance resource constraints, the maintenance decision cycle ultimately performs a replacement action on the generator and an inspection action on the generator bearing. The remaining components to be maintained do not perform maintenance actions in the current maintenance decision cycle. This result shows that the present invention can perform maintenance action consistency correction on component-level action tendencies based on component risk and irreversible wear state, and generate system-level joint maintenance actions that match the component degradation mechanism.

[0113] Example 3 A joint maintenance action generation system for risk constraints of multiple components of a wind turbine, based on a joint maintenance action generation method for risk constraints of multiple components of a wind turbine in Embodiment 1, includes: a data acquisition module, a belief state construction module, a risk calculation module, an action tendency generation module, a candidate action generation module, a maintenance action consistency correction module, a resource feasible domain construction module, an adaptive combination decoding module, a maintenance task output module, and an execution feedback update module; The data acquisition module is used to acquire operation monitoring data, historical status data, historical maintenance data, and current maintenance resource data of multiple components of the wind turbine that need to be maintained. The belief state construction module is used to construct the belief states of multiple components of a wind turbine based on the acquired data. The risk calculation module is used to establish an observable maintenance decision model for risk constraints, and to determine the component risk, system risk and maintenance cost based on the belief state of each component to be maintained. The action tendency generation module is used to input the belief states of multiple components into a pre-trained offline policy network, and obtain the action tendency values ​​corresponding to each component to be maintained through inspection actions, predictive maintenance actions, and replacement action output channels, respectively. The candidate action generation module is used to determine the candidate priority of various maintenance actions based on the action tendency value, component risk, belief status and maintenance action cost of each component to be maintained, and to generate a set of candidate actions. The maintenance action consistency correction module is used to perform maintenance action consistency correction on predictive maintenance candidate actions, and redetermine the action mutual exclusion relationship and resource requirements of the corrected replacement candidate actions; The resource feasible region construction module is used to construct the resource feasible region based on action mutual exclusion constraints, inspection resource constraints, maintenance resource constraints, failure component maintenance constraints, and action mask. The adaptive combination decoding module is used to determine the activation index of combination optimization based on system risk, the degree of conflict of candidate actions and the degree of maintenance resource tension, and to use the combination action decoder to adaptively adopt the item-by-item resource projection or risk constraint combination optimization method to generate system-level joint maintenance actions within the resource feasible domain. The maintenance task output module is used to generate maintenance task data based on system-level joint maintenance actions and output it to the wind turbine operation and maintenance management terminal to generate maintenance work orders. The execution feedback update module is used to obtain the actual execution results of maintenance work orders and update the status data and current maintenance resource data of the corresponding components to be maintained.

[0114] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0115] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine, characterized in that, Includes the following steps: S1. Obtain operation monitoring data, historical status data, historical maintenance data, and current maintenance resource data of multiple components of the wind turbine to be maintained, and construct the multi-component belief status of the wind turbine; S2. Establish a risk-constrained partially observable maintenance decision model, and determine the component risk, system risk, and maintenance cost based on the belief state of each component to be maintained; S3. Input the belief states of multiple components into a pre-trained offline policy network, and obtain the action tendency values ​​corresponding to each component to be maintained through the inspection action, predictive maintenance action, and replacement action output channels, respectively. S4. Determine the candidate priorities for each type of maintenance action based on the action tendency value, component risk, belief status, and maintenance action cost of each component to be maintained, and generate a set of candidate actions; S5. Perform maintenance action consistency correction on the predictive maintenance candidate actions, and redetermine the action mutual exclusion relationship and resource requirements of the corrected replacement candidate actions; S6. Construct the resource feasible region based on action mutual exclusion constraints, inspection resource constraints, maintenance resource constraints, failure component maintenance constraints, and action mask; S7. Determine the activation index for combined optimization based on system risk, the degree of conflict of candidate actions, and the degree of tension of maintenance resources, and use the combined action decoder to adaptively adopt the item-by-item resource projection or risk constraint combined optimization method to generate system-level joint maintenance actions within the resource feasible domain.

2. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 1, characterized in that, Step S1: The multi-component belief state of the wind turbine includes the recoverable degradation state, irreversible wear state, state uncertainty, degradation parameter estimate, observation time, failure state, operating condition and remaining maintenance resources of each component to be maintained. The specific content of constructing the multi-component belief state of a wind turbine is as follows: Based on at least one operating condition variable in the operation monitoring data, determine the operating condition unit to which the component to be maintained belongs in the current maintenance decision cycle, and obtain normal operation samples to determine the normal state benchmark of the operation monitoring variables; Based on the normalized deviation of the operating conditions of each operating monitoring variable of the component to be maintained, determine the current degradation characterization value of the component to be maintained; Based on the current degradation characterization value, the state estimate of the previous maintenance decision cycle, the current operating conditions, the predictive maintenance action execution record, the component replacement record, and the state recovery amount before and after maintenance, the recursive state estimation method is used to determine the recoverable degradation state and irreversible wear state of the current maintenance decision cycle. The recoverable degradation state and the irreversible wear state together form the degradation state characterization of the component to be maintained. The state uncertainty is recursively determined based on the number of maintenance decision cycles since the most recent effective inspection and the status observation update after the effective inspection is performed. The degradation parameter estimates are updated recursively based on the deviation between the actual change in the condition of the component to be maintained and the preset baseline degradation rate. The observation timeframe is determined based on the time interval between the current moment and the most recent valid inspection moment; The failure state is determined based on the relationship between the current degradation characterization value of the component to be maintained and the corresponding failure boundary.

3. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 2, characterized in that, The specific content of constructing the multi-component belief state of a wind turbine also includes: After performing the inspection action, update the recoverable degradation state, degradation parameter estimates, and state uncertainty of the corresponding component to be maintained based on the status observations obtained from the inspection. After performing predictive maintenance actions, the recoverable degradation state of the corresponding component to be maintained is reduced according to a preset recovery ratio or a recovery ratio determined based on historical maintenance results, while the irreversible wear state remains unchanged. After the replacement action is performed, the recoverable degradation state and irreversible wear state of the corresponding component to be maintained will be updated to the initial state after replacement. During maintenance decision cycles in which no effective inspections are performed, the state uncertainty of the corresponding component to be maintained is increased recursively according to the preset state uncertainty growth parameter.

4. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 2, characterized in that, In step S2, the component risk value of each component to be maintained is: in, Let i be the component risk value of the i-th component to be maintained. For risk smoothing parameters, Let be the total degradation amount of the i-th component to be maintained. Let t be the failure boundary of the i-th component to be maintained, t be the maintenance decision cycle, and i be the component number to be maintained. in, Risk contribution parameters corresponding to irreversible wear conditions The parameter contributes to the risk corresponding to state uncertainty. To be in a reversible degenerative state, It is in an irreversible wear state. The state is uncertain; System risks of wind turbines for: in, Risks associated with the aggregation of key components To mitigate the risks associated with component aggregation, A collection of key components. The assembly of supporting components includes at least one of a gearbox, a generator, a main bearing, and a generator bearing, and the supporting components include at least one of a pitch system, a yaw system, a hydraulic system, and a converter. Maintenance cost C per maintenance decision cycle t for: in, , , , and These are inspection costs, predictive maintenance costs, replacement costs, failure event costs, and failure-held costs. , and These are inspection actions, predictive maintenance actions, and replacement actions; To add an invalidation flag, It is in a failed and unrepaired state. For downtime loss parameters, Losses due to downtime of critical components For joint scheduling cost parameters, To cover component scheduling costs, N represents the number of components to be maintained.

5. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 1, characterized in that, In step S3, the pre-training of the offline policy network specifically includes: By using historical operating data of wind turbine units, the baseline degradation rate, operating condition transition patterns, probability of random impact, intensity of random impact, and degradation coupling relationship between components are calibrated for each component to be maintained, and a multi-component degradation-maintenance environment is constructed. By interacting with the degradation-maintenance environment using hybrid behavioral strategies, a full offline maintenance trajectory dataset is constructed. The hybrid behavioral strategies include at least two of the following behavioral modes: conservative maintenance, balanced maintenance, reactive maintenance, delayed maintenance, under-maintenance, proactive replacement, failure exposure, and random exploration. Each transition sample includes the current multi-component belief state, the corresponding maintenance action, the immediate evaluation value, the next multi-component belief state, the maintenance trajectory termination identifier, and the action mask. The offline maintenance trajectory is divided into policy network training data, combined action decoder parameter calibration data, and independent test data. Based on the training data of the policy network, the cumulative discounted return of each transfer sample is calculated according to the real-time evaluation values ​​of each transfer sample from the corresponding maintenance decision cycle to the end maintenance decision cycle of the maintenance trajectory. The sample weights are determined and truncated based on the difference between the cumulative discounted return of each transfer sample and the average return of the samples in the offline maintenance trajectory dataset. Using the belief states of multiple components as input, and the inspection actions, predictive maintenance actions, and replacement actions corresponding to each component to be maintained as training labels, the offline policy network is trained using reward-weighted multi-action binary cross-entropy loss. During training and action generation, action masks are used to block maintenance actions that violate the maintenance constraints of the failed component or do not meet the action execution conditions. After training, the parameters of the offline policy network are fixed. The offline policy network includes a state input layer, a shared state feature extraction layer, and action output channels corresponding to inspection actions, predictive maintenance actions, and replacement actions, respectively. Each action output channel outputs action tendency values ​​corresponding to multiple components to be maintained.

6. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 2, characterized in that, The specific content of step S4 is as follows: The priority of inspection actions is determined based on the inspection action tendency value, state uncertainty, observation timeliness, component risk value, and inspection action cost. The priority of inspection action candidates increases with the increase of state uncertainty, observation timeliness, and component risk value, and decreases with the increase of inspection action cost. The candidate priority of predictive maintenance actions is determined based on the predictive maintenance action tendency value, recoverable degradation state, component risk value, irreversible wear state, and predictive maintenance action cost. The candidate priority of predictive maintenance actions increases with the increase of recoverable degradation state and component risk value, and decreases with the increase of irreversible wear state and predictive maintenance action cost. The candidate priority of replacement actions is determined based on the replacement action tendency value, component risk value, irreversible wear state, failure state, and replacement action cost. The candidate priority of replacement actions increases with the component risk value, irreversible wear state, and failure severity, and decreases with the replacement action cost. The candidate priorities of each type of maintenance action are compared with the corresponding candidate trigger thresholds. Maintenance actions that meet the corresponding candidate trigger conditions and whose action mask is in an executable state are added to the candidate action set.

7. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 1, characterized in that, Step S6, the specific content of the resource feasible domain is as follows: Within the same maintenance decision cycle, the same component to be maintained may be assigned at most one of the following actions: inspection, predictive maintenance, and replacement. The inspection resources consumed by the assigned inspection actions, including inspection personnel, inspection equipment, or inspection time, shall not exceed the available inspection resources corresponding to the current maintenance decision cycle. The maintenance resources consumed by the assigned predictive maintenance actions and replacement actions, including maintenance personnel, maintenance equipment, or maintenance time, shall not exceed the available maintenance resources corresponding to the current maintenance decision cycle. The number and type of spare parts required for the assigned replacement action shall not exceed the current number and type of spare parts available. Failed components are not assigned predictive maintenance actions and are placed in either the replacement candidate queue or the corrective maintenance candidate queue. All maintenance actions entering the resource feasible domain satisfy the corresponding action mask; When maintenance actions for different components require the use of the same non-parallel maintenance equipment or occupy conflicting maintenance time windows, the corresponding maintenance actions should be set to not be executed simultaneously.

8. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 1, characterized in that, In step S7, the specific content of generating system-level joint maintenance actions using the combined action decoder is as follows: The system risk, the degree of action conflict, and the current maintenance resource shortage are normalized respectively, and then weighted and merged based on the corresponding indicator weights to obtain the combined optimized activation indicators; The degree of action conflict is determined based on the number of multiple candidate actions corresponding to the same component to be maintained, the degree to which candidate actions of different components to be maintained compete for the same maintenance resources, and the degree of maintenance time conflict between candidate actions; the current level of maintenance resource shortage is determined based on the relative relationship between the total resource demand of the candidate action set and the currently available maintenance resources. When the combined optimization activation index does not reach the preset activation threshold, the item-by-item resource projection method is adopted to check each candidate action in descending order of priority. If there is no conflict between the current candidate action and the selected action, and the resource feasibility domain is still satisfied after adding the current candidate action, the current candidate action is added to the system-level joint maintenance action, and the remaining inspection resources and remaining maintenance resources are updated. The candidate action check and remaining resource update are repeated until there are no more resource feasibility candidate actions that can be added. When the activation index of the combinatorial optimization reaches the preset activation threshold, the risk-constrained combinatorial optimization method is adopted. The expected risk reduction and irreversible wear mitigation are determined based on the differences in component risk and irreversible wear state before and after the execution of each candidate action. The comprehensive utility of the candidate action is determined based on the expected risk reduction, irreversible wear mitigation, action cost, candidate priority and resource consumption. Based on the component risk after each candidate action is executed, determine the prediction system risk after the combination of candidate actions is executed; When there are candidate action combinations where the predicted system risk is no higher than the preset system risk upper limit, the risk constraint is that the predicted system risk is no higher than the preset system risk upper limit. The combination optimization objective is to improve the overall risk reduction effect and irreversible wear mitigation effect of the selected candidate actions, and reduce the overall action cost and resource consumption. System-level joint maintenance actions are determined within the resource feasible domain. When there are no candidate action combinations that predict system risk no higher than the preset system risk upper limit, the priority optimization objective is to minimize the excess of the predicted system risk relative to the preset system risk upper limit. Among the candidate action combinations with the same or different risk excess amounts that do not exceed the preset tolerance, the system-level joint maintenance action is determined based on the overall utility of the candidate actions.

9. The method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine as described in claim 5, characterized in that, Also includes: After the offline policy network is pre-trained and the network parameters are fixed, the parameters of the combined action decoder are calibrated offline using agent-assisted evolutionary optimization, and the decoder parameters obtained from the offline calibration are fixed as the target decoder parameters. Maintenance task data is generated based on system-level joint maintenance actions and output to the wind turbine operation and maintenance management terminal to generate maintenance work orders. The component status and maintenance resources are updated based on the actual execution results of the maintenance tasks.

10. A joint maintenance action generation system for risk constraints of multiple components of a wind turbine, characterized in that, A method for generating joint maintenance actions for risk constraints of multiple components of a wind turbine, based on any one of claims 1-9, includes: a data acquisition module, a belief state construction module, a risk calculation module, an action tendency generation module, a candidate action generation module, a maintenance action consistency correction module, a resource feasible domain construction module, an adaptive combination decoding module, a maintenance task output module, and an execution feedback update module. The data acquisition module is used to acquire operation monitoring data, historical status data, historical maintenance data, and current maintenance resource data of multiple components of the wind turbine that need to be maintained. The belief state construction module is used to construct the belief states of multiple components of a wind turbine based on the acquired data. The risk calculation module is used to establish an observable maintenance decision model for risk constraints, and to determine the component risk, system risk and maintenance cost based on the belief state of each component to be maintained. The action tendency generation module is used to input the belief states of multiple components into a pre-trained offline policy network, and obtain the action tendency values ​​corresponding to each component to be maintained through inspection actions, predictive maintenance actions, and replacement action output channels, respectively. The candidate action generation module is used to determine the candidate priority of various maintenance actions based on the action tendency value, component risk, belief status and maintenance action cost of each component to be maintained, and to generate a set of candidate actions. The maintenance action consistency correction module is used to perform maintenance action consistency correction on predictive maintenance candidate actions, and redetermine the action mutual exclusion relationship and resource requirements of the corrected replacement candidate actions; The resource feasible region construction module is used to construct the resource feasible region based on action mutual exclusion constraints, inspection resource constraints, maintenance resource constraints, failure component maintenance constraints, and action mask. The adaptive combination decoding module is used to determine the activation index of combination optimization based on system risk, the degree of conflict of candidate actions and the degree of maintenance resource tension, and to use the combination action decoder to adaptively adopt the item-by-item resource projection or risk constraint combination optimization method to generate system-level joint maintenance actions within the resource feasible domain. The maintenance task output module is used to generate maintenance task data based on system-level joint maintenance actions and output it to the wind turbine operation and maintenance management terminal to generate maintenance work orders. The execution feedback update module is used to obtain the actual execution results of maintenance work orders and update the status data and current maintenance resource data of the corresponding components to be maintained.

Images