Fault case-based steam turbine maintenance decision optimization system and method

Abstract

The invention discloses a steam turbine maintenance decision optimization system and method based on fault cases, and belongs to the field of steam turbine fault maintenance, and the system comprises a data collection and processing module, a dynamic graph construction module, a multi-source data alignment module, a cross-domain fault simulation module, a fault causal deduction module, a maintenance strategy generation module, and a strategy adaptation migration module. According to the method, effective integration of cross-modal information can be realized, the expression ability of fault diagnosis and maintenance knowledge is enhanced, the adaptability and accuracy of the system are improved, the physical authenticity and reliability of fault prediction are improved, maintenance personnel are helped to understand a fault propagation mechanism and a risk source, and the transparency and reliability of decision making are improved; the limitation of a single scheme is avoided, it is guaranteed that a maintenance decision can respond to environment changes in time, the adaptability and efficiency of maintenance execution are improved, the maintenance effect can be rehearsed in advance, the field trial and error risk is reduced, scientific assessment and quantitative management of the risk are achieved, and the safety guarantee of the maintenance scheme is improved.

Description

Technical Field

[0001] This invention relates to the field of steam turbine fault repair, and in particular to a steam turbine maintenance decision optimization system and method based on fault cases. Background Technology

[0002] In recent years, with the improvement of the intelligence and digitalization level of energy equipment, equipment health management methods based on the combination of data-driven and mechanism modeling have gradually become a research hotspot. As the core equipment of large-scale energy conversion equipment, the reliability and stability of steam turbines directly affect the economic benefits and safety of power plants, chemical plants, and ship propulsion systems. Due to long-term operation under high temperature, high pressure, and high-speed rotation conditions, steam turbine components are prone to complex multi-physics coupled faults, such as blade fatigue cracks, rotor dynamic balance deviations, bearing wear, and seal leaks. These faults not only affect equipment performance but may also trigger cascading damage, leading to large-scale shutdowns and high economic losses. Traditional maintenance decision-making methods rely heavily on manual experience and single data sources, making it difficult to effectively integrate multi-modal operating data such as acoustics, vibration, temperature, and pressure with historical maintenance data. Furthermore, they lack sufficient modeling of the causal relationships between complex faults and the cross-physical domain influence mechanisms, and lack quantitative analysis of the cost and risk of maintenance paths, making dynamic adjustments difficult.

[0003] Existing turbine maintenance decision optimization systems and methods cannot effectively integrate cross-modal information, have poor ability to express fault diagnosis and maintenance knowledge, reduce the adaptability and accuracy of the system, and hinder maintenance personnel from understanding fault propagation mechanisms and risk sources, thus reducing the transparency and reliability of decision-making. In addition, existing turbine maintenance decision optimization systems and methods cannot overcome the limitations of single solutions, reducing the adaptability and efficiency of maintenance execution, resulting in high risks of on-site trial and error, and failing to achieve scientific risk assessment and quantitative management. Therefore, we propose a turbine maintenance decision optimization system and method based on fault cases. Summary of the Invention

[0004] The purpose of this invention is to address the deficiencies in the existing technology by proposing a turbine maintenance decision optimization system and method based on failure cases.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The turbine maintenance decision optimization system based on fault cases includes a data acquisition and processing module, a dynamic graph construction module, a multi-source data alignment module, a cross-domain fault simulation module, a fault causal inference module, a maintenance strategy generation module, a strategy adaptation and migration module, a maintenance sandbox simulation module, a strategy dynamic optimization module, a hybrid reasoning interpretation module, and a maintenance knowledge evolution module. The data acquisition and processing module is used to collect and preprocess various modal data during the operation of the steam turbine. The dynamic map construction module establishes a dynamic multidimensional fault-maintenance map based on the preprocessed modal data. The multi-source data alignment module is used to align the text descriptions and physical signal features of each modal data after preprocessing. The cross-domain fault simulation module predicts the impact of each fault mode on the maintenance area based on the multiphysics field fault propagation law. The fault cause deduction module identifies the root cause and propagation path of the fault based on the fault-maintenance graph and the fault simulation results. The maintenance strategy generation module dynamically adjusts the maintenance strategy based on the identification results and generates multiple alternative solutions in real time. The strategy adaptation and migration module is used to migrate the adjusted maintenance strategy from the source turbine to the target turbine. The maintenance sandbox simulation module simulates and previews the implementation of maintenance strategies on each steam turbine in a virtual environment and identifies potential risks. The strategy dynamic optimization module reconstructs the maintenance strategy based on the identification results, and sends the reconstructed maintenance strategy back to the maintenance sandbox simulation module for iterative verification. The hybrid reasoning and interpretation module is used to collect the dynamic evolution characteristics of each turbine fault and generate interpretable causal reasoning. The maintenance knowledge evolution module collects fault characteristics, measures taken, and final effects during the maintenance process in real time, and evolves and innovates the current maintenance strategy.

[0006] As a further aspect of the present invention, the specific steps of the data acquisition and processing module in collecting and preprocessing the modal data during turbine operation are as follows: S1.1: Through sensor networks, distributed monitoring systems and historical databases, vibration signals, acoustic signals, temperature field images, operating condition time series data and historical maintenance text modal data are collected synchronously. S1.2: The vibration signal and acoustic signal are denoised using bandpass filters, and then the denoised vibration signal and acoustic signal are normalized. The specific calculation formula for the denoising process is as follows: In the formula, Representative moment Time-filtered vibration or acoustic signal; Representative moment The original vibration signal or acoustic signal at that time; The impulse response function representing a bandpass filter; Represents a delayed variable; The specific calculation formula for normalization is as follows: In the formula, Representative moment At that time, the first The normalized vibration or acoustic signal; Representing the The mean value of the vibration or acoustic signal after group filtering; Representing the The standard deviation of the vibration or acoustic signal after group filtering; S1.3: The temperature field image is grayscale normalized, and the PCA method is used to extract the feature information of the normalized temperature field image for feature dimensionality reduction. Then, the operating condition time series data is smoothed by moving average, and the energy features in the smoothed operating condition time series data are extracted. The specific calculation formula for grayscale normalization is as follows: In the formula, Representative moment At that time, the original temperature field image was in pixels The grayscale value at that location; Representative moment At that time, the standardized temperature field image at the pixel level Pixel value at; as well as Represents the minimum and maximum gray values ​​of the current temperature field image; S1.4: Remove stop words and punctuation marks from the historical maintenance text, and then perform word segmentation on the cleaned historical maintenance text. Then, use the Word2Vec word embedding model to process the segmented historical maintenance text into corresponding vector representations, and finally use the feature mapping function to unify the preprocessed modal data into the same feature space.

[0007] As a further aspect of the present invention, the specific steps of the dynamic map construction module in establishing a dynamic multidimensional fault-maintenance map are as follows: S2.1: Use the modal features corresponding to each modal data after preprocessing as the initial nodes of the fault-maintenance graph. Each node corresponds to a fault feature, maintenance measure or working condition information. Based on the association between nodes in historical cases, construct the edges and weights of the graph. Then, use the constructed initial nodes. S2.2: Each node in the fault-maintenance graph receives information from its neighboring nodes and aggregates the received neighboring node information based on the edge weights connecting the nodes. Then, each node combines its own modal features with the aggregated neighboring information and updates the node through a non-linear activation function. S2.3: Repeat the aggregation of neighbor node information and node update until the preset number of iterations is reached, then stop the node update and take the finally updated nodes as a set of modal sequences. Then add the corresponding modal embedding and position encoding to each node. After that, generate the query, key and value corresponding to each node in the current modal sequence through three different linear mapping matrices. S2.4: Based on the query and key of each node, calculate the relevance score between each node, and normalize the relevance score between each node using the Softmax algorithm to obtain the attention weight between each node, so as to represent the importance of information interaction between each node. S2.5: The values ​​of all nodes are weighted and summed according to the attention weights. Then, the attention weights of each node are calculated and weighted summed repeatedly using multiple sets of queries, keys and values ​​with different parameters. The sums of each set are then concatenated and the final node representation is generated through nonlinear transformation. S2.6: Convert the final node representation into entity embeddings in the fault-maintenance graph, and dynamically adjust the weights of the corresponding edges according to the strength of the relationships between each node. Then, when a new fault case arrives, expand the fault-maintenance graph through an incremental update mechanism and update the node representations in the fault-maintenance graph.

[0008] As a further aspect of the present invention, the initial node described in S2.1 is specifically manifested as follows: In the formula, Representing the The initial representation vector of each node; Representing the Multimodal feature vectors of each node; as well as These represent the weight matrix and bias of the feature map, respectively; The specific calculation formula for the attention weight mentioned in S2.5 is as follows: In the formula, Representative node For nodes Attention weights; Representative node The query vector; Representative node The key vector; The dimension representing the query and key vector; Represents the node index value.

[0009] As a further aspect of the present invention, the specific steps of the multi-source data alignment module in aligning the text description and physical signal characteristics of each modal data after preprocessing are as follows: P1.1: Vibration signals, acoustic signals, and historical maintenance texts are collected, preprocessed, and mapped to signal features and semantic features in a unified feature space. Then, the cosine similarity between semantic features and signal features is calculated. P1.2: Text-signal pairs with cosine similarity higher than a preset threshold are taken as positive samples, and text-signal pairs with cosine similarity lower than a preset threshold are taken as negative samples. Based on the positive and negative samples, the loss function is learned by comparison, and the loss value between the positive and negative samples is calculated. P1.3: If the loss value is higher than the preset threshold, the comparative learning loss function is optimized by gradient descent, and the loss value calculation and optimization are repeated until the loss value converges to the preset range. At the same time, the semantic features and signal features are aligned, and cross-modal correlation is established. P1.4: The semantic and signal features aligned through contrastive learning are formed into mapping pairs. Identification information is added to each semantic-signal pair as a bridge between nodes in the fault-maintenance graph. New edges are established for corresponding nodes in the fault-maintenance graph according to the mapping table. At the same time, weight values ​​are set for the corresponding new edges based on the cosine similarity between nodes.

[0010] As a further aspect of the present invention, the specific steps of the cross-domain fault simulation module in predicting the impact of each fault mode on the maintenance area are as follows: S3.1: Divide each steam turbine into a physical domain, including a thermal field domain, a mechanical field domain, a vibration field domain, and a flow field domain, and identify the coupling boundaries between each physical domain. Then, establish corresponding partial differential equations for each physical domain, and establish corresponding physical domain models based on the preset boundary conditions and initial states of each physical domain. At the same time, set the coupling boundary conditions of the coupling regions of each physical domain. S3.2: Extract observation data corresponding to different types of faults from the historical fault case library, and record the response fields of different subsystems in the turbine in each physical field under the collected fault observation data. Based on the collected fault observation data, take the corresponding fault mode and its spatial distribution as input data, take the multi-physics response caused by the fault in the corresponding spatiotemporal range as output data, and establish corresponding input-output training sample pairs. S3.3: Extract the spatial frequency distribution features of each input data through Fourier transform, expand the channels of the spatial frequency distribution features, and perform normalization processing. Then, perform global calculations on the processed spatial frequency distribution features through Fourier convolution kernels of different sizes. S3.4: After each set of Fourier convolution kernels completes global computation, the generated Fourier feature map is nonlinearly mapped, and the mapped Fourier feature map is used as the input of the next Fourier convolution kernel. Then, the Fourier feature map output by the last set of Fourier convolution kernels is compressed into the corresponding frequency domain response features through linear mapping. S3.5: Convert the obtained frequency domain response features into the spatial domain through inverse Fourier transform to obtain the corresponding response distribution prediction results. Calculate the error value between each response distribution prediction result and the corresponding output data. If the error value exceeds the preset threshold, optimize each Fourier convolution kernel, nonlinear mapping and various parameters in the linear mapping process through backpropagation based on the current error value. S3.6: Repeat the frequency domain response feature extraction and parameter update until the error value converges to the preset range. Then, use unused input-output training sample pairs to recalculate the error value to confirm that the error value is still within the preset range on the unknown data. If it exceeds the preset range, re-update the parameters. S3.7: Collect the latest physical field data of the steam turbine, and extract the corresponding response distribution prediction results based on the updated Fourier convolution kernels, nonlinear mapping and linear mapping process. Then calculate the risk score corresponding to each response distribution prediction result, and establish a risk heat map of the current steam turbine. At the same time, identify the risk subsystems and their impact paths with risk scores higher than the preset threshold, and supplement the identification results into the fault-maintenance knowledge graph.

[0011] As a further aspect of the present invention, the specific steps of the fault cause-effect deduction module in identifying the fault root cause and propagation path are as follows: S4.1: Extract structured entities and their relationships from the fault-maintenance graph, obtain the corresponding response distribution prediction results from the fault simulation results, map the collected data sets into variable node sets required for causal analysis, and establish causal edges between nodes in the variable node sets based on the response distribution prediction results. S4.2: Calculate the conditional independence between each pair of variable nodes, and based on the preset set of conditional variables, verify whether each variable node is independent under any preset conditional variable. If they are independent under any preset conditional variable, remove the causal edge between the corresponding two sets of variable nodes to establish a preliminary maintenance-fault causal graph. S4.3: Establish a structural scoring function for the current maintenance-failure cause-effect graph based on BIC or AIC to evaluate the balance between the goodness of fit of the current maintenance-failure cause-effect graph to each modality and the structural complexity. Then, perform various operations such as adding edges, deleting edges, and reversing directions on the current maintenance-failure cause-effect graph. After each modification to the graph structure, calculate a new score. If the score improves, accept the modification; otherwise, revert to the previous state. S4.4: Repeatedly iterate and optimize the maintenance-fault cause-effect graph until the change value of the score after multiple iterations converges to the preset range. After the iteration ends, output the optimal maintenance-fault cause-effect graph and correct the causal direction of each variable node in the maintenance-fault cause-effect graph based on physical prior knowledge and time sequence information. S4.5: Use a regression model to evaluate the strength of each causal edge in the maintenance-failure causal graph. Based on the failure simulation results, start from the target failure node, traverse backward along the maintenance-failure causal graph to the source node, and select the root cause nodes of each group whose causal edge strength is higher than the preset threshold and satisfies the preset physical constraints. S4.6: Starting from each root cause node, traverse forward along the causal direction to generate the corresponding fault propagation path, and generate the corresponding path description for each fault propagation path based on the semantic information in the fault-maintenance graph.

[0012] A turbine maintenance decision optimization method based on failure cases is proposed, and the specific steps of this optimization method are as follows: I. Collect and preprocess modal data from the turbine's operation monitoring system, historical maintenance records, and sensor acquisition terminals; II. The preprocessed modal data are fused and aligned, and a fault-maintenance map is constructed based on the aligned modal data; III. Based on the fault-maintenance map, establish physical field models for each part of the turbine, predict the fault propagation path and its impact on the maintenance area, and generate an impact distribution map; IV. Based on the fault-maintenance diagram and impact distribution diagram, construct a fault causal network to identify the root causes and propagation paths of the faults; V. Based on the identification results, construct a maintenance decision tree and dynamically adjust the maintenance strategy, while outputting multiple candidate solutions according to different fault scenarios; VI. Transfer the optimal maintenance strategy of the source turbine to the target turbine, and simulate and preview the generated maintenance strategy in a virtual sandbox environment; VII. Based on simulation results, adjust and optimize maintenance paths in real time when resource or risk conditions change, and make natural selections based on actual performance and simulation results.

[0013] As a further aspect of the present invention, the specific steps of dynamically adjusting the maintenance strategy in step V are as follows: S5.1: Traverse the fault-maintenance cause-effect graph by depth-first search or breadth-first search, and obtain the set of all paths from the current fault node to the node of feasible maintenance measures, where each path represents a chain of cause-effect repair solutions; S5.2: Treat each path as a branch of the decision tree, where each decision node in the branch represents a maintenance measure selection point, and the leaf node represents the system state after maintenance is completed. Then, calculate the comprehensive evaluation value of each maintenance path and sort the maintenance paths according to the comprehensive evaluation value from smallest to largest. S5.3: When the operating environment, component status, available resources or security policies change, update the comprehensive evaluation value of each maintenance path, and reorder the path priorities based on the updated comprehensive evaluation value. At the same time, select the maintenance plan corresponding to the maintenance path ranked first as the optimal maintenance plan. S5.4: The maintenance plans corresponding to the maintenance paths whose comprehensive evaluation values ​​are within the preset range are taken as multi-path alternatives. After the maintenance plan is executed, the actual cost, risk events and maintenance results are recorded, and the recorded data are used in the next round of maintenance strategy generation module plan update.

[0014] As a further aspect of the present invention, the specific steps of simulating the generated maintenance strategy in a virtual sandbox environment as described in step VI are as follows: S6.1: Based on the structure, material properties, and sensor data of a real steam turbine, construct a digital model of the steam turbine that includes the geometric and physical properties of the rotor system, bearing support, and sealing components, and then create an isolated simulation environment; S6.2: Import the generated multi-path maintenance scheme into the simulation environment, simulate the execution process of the multi-path maintenance scheme, and calculate the rotor dynamic balance deviation in real time during the simulation process of the multi-path maintenance scheme. S6.3: Select a set of uncertain parameters with randomness from the multi-path maintenance scheme and the turbine operating environment. Based on historical data statistics or engineering experience, set the probability distribution of each uncertain parameter. Then, perform independent or related sampling from each uncertain parameter probability distribution to obtain the full parameter vector required for one simulation. Repeat the sampling multiple times to establish the corresponding test set. S6.4: Based on each full parameter vector in the test set, run maintenance operation simulations in the simulation environment and the steam turbine digital model to obtain the performance index values ​​under the full parameter vector, calculate the risk factor of each simulation output result, and then calculate the overall expected risk based on all sampled risk factors, and perform histogram statistics on each risk factor to obtain the probability density function corresponding to each risk factor. S6.5: Calculate the average dynamic balance deviation, overall expected risk, and corresponding maintenance time and resource consumption for each maintenance plan, and construct a performance evaluation table. If the overall expected risk of the maintenance plan exceeds the preset safety limit or the average dynamic balance deviation is higher than the preset threshold, adjust the maintenance steps or replace the plan, and feed the probability density function back to the maintenance strategy generation module for subsequent maintenance strategy updates.

[0015] As a further aspect of the present invention, the specific calculation formula for the dynamic balance deviation in S6.2 is as follows: In the formula, The root mean square value represents the rotor dynamic balance deviation; This represents the total duration of the simulation evaluation; Represents the rotor's center of mass in time radial displacement; Represents the rotor's center of mass in time The desired rotor radial trajectory; The specific calculation formulas for the risk factors mentioned in S6.4 are as follows: In the formula, Representing the Risk factors for the simulation results; Representing the Performance metrics from the simulation; This represents the lower limit of the performance index; This represents the upper limit of the performance index; The specific formula for calculating the overall expected risk described in S6.4 is as follows: In the formula, Represents overall expected risk; Representing the Risk factors for the simulation results; This represents the total number of samples taken.

[0016] As a further aspect of the present invention, the set of uncertain parameters in S6.3 includes component assembly clearance, bolt preload, and changes in lubricating oil viscosity, etc.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention maps multimodal data into fault, maintenance, and operating condition nodes. It then uses historical cases to construct weighted edges to form an initial fault-maintenance graph. Node representations are updated by iteratively aggregating neighbor information. Subsequently, based on a multi-head attention mechanism, entity embeddings are generated and edge weights are dynamically adjusted. By combining contrastive learning to align semantics and signal features, cross-modal associations are established. The turbine is then divided into multiple physical domains, corresponding partial differential equations and coupled boundary conditions are constructed, historical fault-response pairs are collected, and the fault-to-response mapping is learned in the frequency domain. Simultaneously, parameters are optimized in reverse to improve prediction accuracy. Based on the prediction results, a [database / system] is generated. Risk scoring and impact distribution maps identify high-risk subsystems and update the fault-maintenance graph. Subsequently, a maintenance-fault causal graph is constructed based on causal structure constraints and scoring optimization. The causal direction is corrected and the edge strength is quantified. The root cause is traced backward from the target fault, and then a propagation path and path description that conform to physical constraints are generated in the forward direction. This enables the effective integration of cross-modal information, enhances the expressive ability of fault diagnosis and maintenance knowledge, improves the adaptability and accuracy of the system, and enhances the physical authenticity and reliability of fault prediction. It helps maintenance personnel understand the fault propagation mechanism and risk sources, and improves the transparency and reliability of decision-making.

[0018] 2. This invention traverses the fault-maintenance cause-effect graph using depth-first or breadth-first search to obtain all paths from the fault node to the node of feasible maintenance measures. Each path is used as a branch of a decision tree. A comprehensive evaluation value is calculated for each path and the paths are ranked. The ranking is dynamically updated when the operating environment or resources change, prioritizing the execution of the top-ranked solution, with the others within a threshold as alternatives. Then, a digital model of the turbine is constructed based on the actual structure and sensor data. The multi-path solutions are imported into an isolated simulation environment for simulation execution. The rotor dynamic balance deviation is calculated in real time. A probability distribution is set, and Monte Carlo sampling is performed on the random operating and maintenance parameters. Multiple simulations are conducted to obtain performance indicators and risk factors. The probability density function and overall expected risk are statistically analyzed to generate a performance evaluation table. If the risk or deviation exceeds the limit, the solution is adjusted or replaced. After the maintenance solution is executed, the cost, risk, and effect are recorded, and the evaluation results are fed back to the strategy generation module. This avoids the limitations of a single solution, ensures that maintenance decisions can respond promptly to environmental changes, improves the adaptability and efficiency of maintenance execution, allows for advance simulation of maintenance effects, reduces the risk of trial and error on site, achieves scientific assessment and quantitative management of risks, and enhances the safety assurance of maintenance solutions. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0020] Figure 1 This is a system block diagram of the turbine maintenance decision optimization system based on fault cases proposed in this invention; Figure 2 This is a flowchart of the turbine maintenance decision optimization method based on fault cases proposed in this invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Example 1, referring to Figure 1 The turbine maintenance decision optimization system based on fault cases includes a data acquisition and processing module, a dynamic graph construction module, a multi-source data alignment module, a cross-domain fault simulation module, a fault causal inference module, a maintenance strategy generation module, a strategy adaptation and migration module, a maintenance sandbox simulation module, a strategy dynamic optimization module, a hybrid reasoning and interpretation module, and a maintenance knowledge evolution module. The data acquisition and processing module is used to collect and preprocess various modal data during the operation of the steam turbine.

[0023] Specifically, vibration signals, acoustic signals, temperature field images, operating condition time-series data, and historical maintenance texts are synchronously collected through sensor networks, distributed monitoring systems, and historical databases. Vibration and acoustic signals are denoised using bandpass filters, and then normalized. Temperature field images are standardized in grayscale, and PCA is used to extract feature information from the standardized images for feature reduction. Operating condition time-series data are then smoothed using a moving average, and energy features are extracted. Stop words and punctuation marks are removed from historical maintenance texts, which are then segmented. The Word2Vec word embedding model is used to convert the segmented historical maintenance texts into corresponding vector representations. Finally, a feature mapping function unifies the preprocessed modal data into the same feature space.

[0024] It should be further explained that the specific calculation formula for noise reduction is as follows: In the formula, Representative moment Time-filtered vibration or acoustic signal; Representative moment The original vibration signal or acoustic signal at that time; The impulse response function representing a bandpass filter; Represents a delayed variable; The specific calculation formula for normalization is as follows: In the formula, Representative moment At that time, the first The normalized vibration or acoustic signal; Representing the The mean value of the vibration or acoustic signal after group filtering; Representing the The standard deviation of the vibration or acoustic signal after group filtering; The specific formula for calculating grayscale normalization is as follows: In the formula, Representative moment At that time, the original temperature field image was in pixels The grayscale value at that location; Representative moment At that time, the standardized temperature field image at the pixel level Pixel value at; as well as These represent the minimum and maximum gray values ​​of the current temperature field image.

[0025] The dynamic map construction module establishes a dynamic, multi-dimensional fault-maintenance map based on the preprocessed modal data.

[0026] Specifically, the modal features corresponding to each preprocessed modal data are used as the initial nodes of the fault-maintenance graph. Each node corresponds to a fault feature, maintenance measure, or operating condition information. Based on the associations between nodes in historical cases, the edges and weights of the graph are constructed. Then, each node in the fault-maintenance graph receives information from its neighbor nodes, and the received neighbor node information is weighted and aggregated based on the edge weights connecting the nodes. Each node then combines its own modal features with the aggregated neighbor information and updates the node using a non-linear activation function. This process of neighbor node information aggregation and node updating is repeated until a preset number of iterations is reached. At this point, node updates are stopped, and the finally updated nodes are used as a modal sequence. A corresponding modal embedding and positional encoding are then added to each node. Finally, three different linear mapping matrices are used to generate the current... The query, key, and value corresponding to each node in the premodal sequence are used to calculate the relevance score between each node based on the query and key of each node. The relevance score between each node is normalized by the Softmax algorithm to obtain the attention weight between each node to represent the importance of information interaction between nodes. The values ​​of all nodes are weighted and summed according to the attention weight. Then, using multiple sets of queries, keys, and values ​​with different parameters, the attention weight calculation and weighted summation of each node are repeated. The obtained summation results are concatenated and then a nonlinear transformation is used to generate the final node representation. The final node representation is converted into entity embedding in the fault-maintenance graph, and the weight of the corresponding edge is dynamically adjusted according to the relationship strength between each node. When a new fault case arrives, the fault-maintenance graph is expanded through an incremental update mechanism, and the node representation in the fault-maintenance graph is updated.

[0027] In addition, it should be noted that the initial node is represented as follows: In the formula, Representing the The initial representation vector of each node; Representing the Multimodal feature vectors of each node; as well as These represent the weight matrix and bias of the feature map, respectively; The specific calculation formula for the attention weight mentioned in S2.5 is as follows: In the formula, Representative node For nodes Attention weights; Representative node The query vector; Representative node The key vector; The dimension representing the query and key vector; Represents the node index value.

[0028] The multi-source data alignment module is used to align the text descriptions and physical signal characteristics of each modal data after preprocessing.

[0029] It should be noted that after collecting vibration signals, acoustic signals, and historical maintenance texts, and preprocessing them, they are mapped to signal features and semantic features in a unified feature space. Then, the cosine similarity between semantic features and signal features is calculated. Text-signal pairs with a cosine similarity higher than a preset threshold are considered positive samples, and those with a cosine similarity lower than the preset threshold are considered negative samples. Based on the positive and negative samples, a contrastive learning loss function is used to calculate the loss value between the positive and negative samples. If the loss value is higher than the preset threshold, the contrastive learning loss function is optimized using gradient descent. The loss value calculation and optimization are repeated until the loss value converges to a preset range. At the same time, semantic features and signal features are aligned, and cross-modal associations are established. The semantic features and signal features aligned by contrastive learning are formed into mapping pairs. Identification information is added to each semantic-signal pair as a bridge between nodes in the fault-maintenance graph. New edges are established for corresponding nodes in the fault-maintenance graph according to the mapping table, and weight values ​​are set for the corresponding new edges based on the cosine similarity between each node.

[0030] The cross-domain fault simulation module predicts the impact of each fault mode on the maintenance area based on the multiphysics field fault propagation law.

[0031] Specifically, each steam turbine is divided into physical domains: thermal field, mechanical field, vibration field, and flow field. The coupling boundaries between these physical domains are identified. Then, corresponding partial differential equations are established for each physical domain. Based on the preset boundary conditions and initial states of each physical domain, a corresponding physical domain model is built. Simultaneously, coupling boundary conditions for the coupling regions of each physical domain are set. Observational data corresponding to different types of faults are extracted from a historical fault case database, and the response fields of different subsystems within the steam turbine in each physical field are recorded under the collected fault observation data. Based on the collected fault observation data, the corresponding fault modes and their spatial characteristics are... Using the spatial frequency distribution as input data, the multiphysics response caused by the fault within the corresponding spatiotemporal range is used as output data. Corresponding input-output training sample pairs are established. The spatial frequency distribution features of each input data are extracted through Fourier transform, and the spatial frequency distribution features are channel-expanded and normalized. Then, Fourier convolution kernels of different sizes are used to perform global operations on the processed spatial frequency distribution features. After each set of Fourier convolution kernels has completed global operations, the generated Fourier feature map is nonlinearly mapped, and the mapped Fourier feature map is used as the input of the next Fourier convolution kernel. Then, the Fourier feature maps output by the last set of Fourier convolution kernels are compressed into corresponding frequency domain response features through linear mapping. The obtained frequency domain response features are then converted to the spatial domain using inverse Fourier transform to obtain the corresponding response distribution prediction results. The error value between each response distribution prediction result and the corresponding output data is calculated. If the error value exceeds a preset threshold, based on the current error value, the parameters of each Fourier convolution kernel, nonlinear mapping, and linear mapping process are optimized through backpropagation. Frequency domain response feature extraction and parameter updates are repeated until the error value converges to a preset range. Afterwards, unused... The input-output training sample pairs are recalculated to confirm that the error value is still within the preset range on the unknown data. If it exceeds the preset range, the parameters are updated again, the latest physical field data of the turbine are collected, and the corresponding response distribution prediction results are extracted based on the Fourier convolution kernels, nonlinear mappings, and linear mapping processes after the parameter update. Then, the risk score corresponding to each response distribution prediction result is calculated, and the current turbine's impact distribution map is established. At the same time, the risk subsystems with risk scores higher than the preset threshold and their impact paths are identified, and the identification results are added to the fault-maintenance knowledge graph.

[0032] The fault cause deduction module identifies the root cause and propagation path of the fault based on the fault-maintenance graph and the fault simulation results.

[0033] Specifically, structured entities and their relationships are extracted from the fault-maintenance graph, and corresponding response distribution prediction results are obtained from the fault simulation results. The collected data sets are then mapped to variable node sets required for causal analysis. Based on the response distribution prediction results, causal edges are established between nodes in the variable node sets. The conditional independence between each pair of variable nodes is calculated, and based on a preset set of conditional variables, it is verified whether each variable node is independent under any preset conditional variable. If independent under any preset conditional variable, the causal edges between the corresponding two sets of variable nodes are removed to establish a preliminary maintenance-fault causal graph. A structural scoring function for the current maintenance-fault causal graph is established based on BIC or AIC to evaluate the balance between the goodness of fit and structural complexity of the current maintenance-fault causal graph for each modality of data. Finally, edge addition, edge deletion, and direction modification are performed on the current maintenance-fault causal graph. The process involves calculating a new score after each modification to the graph structure. If the score improves, the modification is accepted; otherwise, it is rolled back. This iterative optimization of the maintenance-fault causal graph continues until the score's change over multiple iterations converges to a preset range. After the iteration ends, the optimal maintenance-fault causal graph is output. Based on prior physical knowledge and temporal information, the causal direction of each variable node in the maintenance-fault causal graph is corrected. A regression model is used to evaluate the strength of each causal edge in the maintenance-fault causal graph. Based on the fault simulation results, starting from the target fault node, the process traverses backward along the maintenance-fault causal graph to the source node, selecting root cause nodes whose causal edge strength is higher than a preset threshold and satisfies preset physical constraints. Starting from each root cause node, the process traverses forward along the causal direction to generate corresponding fault propagation paths. Based on the semantic information in the fault-maintenance graph, corresponding path descriptions for each fault propagation path are generated.

[0034] The maintenance strategy generation module dynamically adjusts the maintenance strategy based on the identification results and generates multiple alternative solutions in real time; the strategy adaptation and migration module is used to migrate the adjusted maintenance strategy from the source turbine to the target turbine; the maintenance sandbox simulation module simulates the implementation of the maintenance strategy on each turbine in a virtual environment and identifies potential risks; the strategy dynamic optimization module reconstructs the maintenance strategy based on the identification results and sends the reconstructed maintenance strategy back to the maintenance sandbox simulation module for iterative verification. The hybrid reasoning explanation module is used to collect the dynamic evolution characteristics of each turbine fault and generate interpretable causal reasoning; the maintenance knowledge evolution module collects fault characteristics, measures taken and final effects in real time during the maintenance process, and evolves and innovates the current maintenance strategy.

[0035] Example 2, refer to Figure 2 A turbine maintenance decision optimization method based on failure cases is proposed. The specific steps of this optimization method are as follows: Data from various modes is collected and preprocessed from the turbine's operation monitoring system, historical maintenance records, and sensor acquisition terminals.

[0036] The preprocessed modal data are fused and aligned, and a fault-maintenance map is constructed based on the aligned modal data.

[0037] Based on the fault-maintenance map, we establish physical field models for each part of the steam turbine, predict the fault propagation path and its impact on the maintenance area, and generate an impact distribution map.

[0038] Based on the fault-maintenance diagram and the impact distribution diagram, a fault causal network is constructed to identify the root causes and propagation paths of the faults.

[0039] Based on the identification results, a maintenance decision tree is constructed, and the maintenance strategy is dynamically adjusted. At the same time, multiple candidate solutions are output according to different fault scenarios.

[0040] Specifically, the fault-maintenance cause-effect graph is traversed using depth-first search or breadth-first search to obtain a set of all paths from the current fault node to the feasible maintenance measure node. Each path represents a chain of cause-effect repair solutions, and each path is treated as a branch of a decision tree. In this branch, each decision node represents a maintenance measure selection point, and the leaf node represents the system state after maintenance. Then, the comprehensive evaluation value of each maintenance path is calculated, and the maintenance paths are sorted from smallest to largest according to the comprehensive evaluation value. When the operating environment, component status, available resources, or security policies change, the comprehensive evaluation value of each maintenance path is updated, and the path priorities are reordered based on the updated comprehensive evaluation value. At the same time, the maintenance plan corresponding to the maintenance path ranked first is selected as the optimal maintenance plan, and the maintenance plans corresponding to the other maintenance paths whose comprehensive evaluation values ​​are within the preset range are used as multi-path alternatives. After the maintenance plan is executed, the actual cost, risk events, and maintenance effects are recorded, and the recorded data are used in the next round of maintenance strategy generation module plan updates.

[0041] The optimal maintenance strategy of the source turbine is transferred to the target turbine, and the generated maintenance strategy is simulated and previewed in a virtual sandbox environment.

[0042] Specifically, based on the structure, material properties, and sensor data of a real steam turbine, a digital model of the turbine is constructed, including the geometric and physical properties of the rotor system, bearing supports, and sealing components. Then, an isolated simulation environment is created, and the generated multi-path maintenance scheme is imported into the simulation environment. The execution process of the multi-path maintenance scheme is simulated simultaneously, and the rotor dynamic balance deviation is calculated in real time during the simulation. A set of uncertain parameters with randomness is selected from the multi-path maintenance scheme and the steam turbine operating environment. Based on historical data statistics or engineering experience, the probability distribution of each uncertain parameter is set. Then, independent or correlated sampling is performed from each uncertain parameter probability distribution to obtain the full parameter vector required for one simulation. This sampling is repeated multiple times to establish a corresponding test set. Based on each full parameter vector in the test set, maintenance operation simulations are run in the simulation environment and the turbine digital model to obtain the performance index values ​​under the full parameter vector. The risk factors of each simulation output result are calculated. Then, based on all sampled risk factors, the overall expected risk is calculated, and histogram statistics are performed on each risk factor to obtain the probability density function corresponding to each risk factor. The average dynamic balance deviation, overall expected risk, and corresponding maintenance time and resource consumption of each maintenance plan are statistically analyzed, and a performance evaluation table is constructed. If the overall expected risk of the maintenance plan exceeds the preset safety limit or the average dynamic balance deviation is higher than the preset threshold, the maintenance steps are adjusted or the plan is replaced. The probability density function is fed back to the maintenance strategy generation module for subsequent maintenance strategy updates.

[0043] It should be further explained that the specific formula for calculating the dynamic balance deviation is as follows: In the formula, The root mean square value represents the rotor dynamic balance deviation; This represents the total duration of the simulation evaluation; Represents the rotor's center of mass in time radial displacement; Represents the rotor's center of mass in time The desired rotor radial trajectory; The specific formula for calculating the risk factor is as follows: In the formula, Representing the Risk factors for the simulation results; Representing the Performance metrics from the simulation; This represents the lower limit of the performance index; This represents the upper limit of the performance index; The specific formula for calculating the overall expected risk described in S6.4 is as follows: In the formula, Represents overall expected risk; Representing the Risk factors for the simulation results; This represents the total number of samples taken.

[0044] In addition, it should be noted that the set of uncertain parameters includes component assembly clearance, bolt preload, and changes in lubricant viscosity.

[0045] Based on simulation results, adjustments are made in real time when resource or risk conditions change, maintenance paths are optimized in real time, and natural selection is made based on actual performance and simulation results.

Claims

1. A turbine maintenance decision optimization system based on failure cases, characterized in that, It includes a data acquisition and processing module, a dynamic graph construction module, a multi-source data alignment module, a cross-domain fault simulation module, a fault causal inference module, a maintenance strategy generation module, a strategy adaptation and migration module, a maintenance sandbox simulation module, a strategy dynamic optimization module, a hybrid reasoning and interpretation module, and a maintenance knowledge evolution module. The data acquisition and processing module is used to collect and preprocess various modal data during the operation of the steam turbine. The dynamic map construction module establishes a dynamic multidimensional fault-maintenance map based on the preprocessed modal data. The multi-source data alignment module is used to align the text descriptions and physical signal features of each modal data after preprocessing. The cross-domain fault simulation module predicts the impact of each fault mode on the maintenance area based on the multiphysics field fault propagation law. The fault cause deduction module identifies the root cause and propagation path of the fault based on the fault-maintenance graph and the fault simulation results. The maintenance strategy generation module dynamically adjusts the maintenance strategy based on the identification results and generates multiple alternative solutions in real time. The strategy adaptation and migration module is used to migrate the adjusted maintenance strategy from the source turbine to the target turbine. The maintenance sandbox simulation module simulates and previews the implementation of maintenance strategies on each steam turbine in a virtual environment and identifies potential risks. The strategy dynamic optimization module reconstructs the maintenance strategy based on the identification results, and sends the reconstructed maintenance strategy back to the maintenance sandbox simulation module for iterative verification. The hybrid reasoning and interpretation module is used to collect the dynamic evolution characteristics of each turbine fault and generate interpretable causal reasoning. The maintenance knowledge evolution module optimizes and innovates maintenance strategies based on sandbox simulation results and actual execution feedback, and feeds this feedback back to the dynamic graph construction module.

2. The turbine maintenance decision optimization system based on fault cases according to claim 1, characterized in that, The specific steps of the data acquisition and processing module in collecting and preprocessing the modal data during turbine operation are as follows: S1.1: Through sensor networks, distributed monitoring systems and historical databases, vibration signals, acoustic signals, temperature field images, operating condition time series data and historical maintenance text modal data are collected synchronously. S1.2: The vibration signal and acoustic signal are denoised using bandpass filters, and then the denoised vibration signal and acoustic signal are normalized. The specific calculation formula for the denoising process is as follows: In the formula, Representative moment Time-filtered vibration or acoustic signal; Representative moment The original vibration signal or acoustic signal at that time; The impulse response function representing a bandpass filter; Represents a delayed variable; The specific calculation formula for normalization is as follows: In the formula, Representative moment At that time, the first The vibration or acoustic signal after group normalization; Representing the The mean value of the vibration or acoustic signal after group filtering; Representing the The standard deviation of the vibration or acoustic signal after group filtering; S1.3: The temperature field image is grayscale normalized, and the PCA method is used to extract the feature information of the normalized temperature field image for feature dimensionality reduction. Then, the operating condition time series data is smoothed by moving average, and the energy features in the smoothed operating condition time series data are extracted. The specific calculation formula for grayscale normalization is as follows: In the formula, Representative moment At that time, the original temperature field image was in pixels The grayscale value at that location; Representative moment At that time, the standardized temperature field image at the pixel level Pixel value at; as well as Represents the minimum and maximum gray values ​​of the current temperature field image; S1.4: Remove stop words and punctuation marks from the historical maintenance text, and then perform word segmentation on the cleaned historical maintenance text. Then, use the Word2Vec word embedding model to process the segmented historical maintenance text into corresponding vector representations, and finally use the feature mapping function to unify the preprocessed modal data into the same feature space.

3. The turbine maintenance decision optimization system based on fault cases according to claim 2, characterized in that, The specific steps for the dynamic graph construction module to establish a dynamic multidimensional fault-maintenance graph are as follows: S2.1: Use the modal features corresponding to each modal data after preprocessing as the initial nodes of the fault-maintenance graph. Each node corresponds to a fault feature, maintenance measure or working condition information. Based on the association between nodes in historical cases, construct the edges and weights of the graph. Then, use the constructed initial nodes. S2.2: Each node in the fault-maintenance graph receives information from its neighboring nodes and aggregates the received neighboring node information based on the edge weights connecting the nodes. Then, each node combines its own modal features with the aggregated neighboring information and updates the node through a non-linear activation function. S2.3: Repeat the aggregation of neighbor node information and node update until the preset number of iterations is reached, then stop the node update and take the finally updated nodes as a set of modal sequences. Then add the corresponding modal embedding and position encoding to each node. After that, generate the query, key and value corresponding to each node in the current modal sequence through three different linear mapping matrices. S2.4: Based on the query and key of each node, calculate the relevance score between each node, and normalize the relevance score between each node using the Softmax algorithm to obtain the attention weight between each node, so as to represent the importance of information interaction between each node. S2.5: The values ​​of all nodes are weighted and summed according to the attention weights. Then, the attention weights of each node are calculated and weighted summed repeatedly using multiple sets of queries, keys and values ​​with different parameters. The sums of each set are then concatenated and the final node representation is generated through nonlinear transformation. S2.6: Convert the final node representation into entity embeddings in the fault-maintenance graph, and dynamically adjust the weights of the corresponding edges according to the strength of the relationships between each node. Then, when a new fault case arrives, expand the fault-maintenance graph through an incremental update mechanism and update the node representations in the fault-maintenance graph.

4. The turbine maintenance decision optimization system based on fault cases according to claim 3, characterized in that, The specific form of the initial node described in S2.1 is as follows: In the formula, Representing the The initial representation vector of each node; Representing the Multimodal feature vectors of each node; as well as These represent the weight matrix and bias of the feature map, respectively; The specific calculation formula for the attention weight mentioned in S2.5 is as follows: In the formula, Representative node For nodes Attention weights; Representative node The query vector; Representative node The key vector; The dimension representing the query and key vector; Represents the node index value.

5. The turbine maintenance decision optimization system based on fault cases according to claim 3, characterized in that, The specific steps of the cross-domain fault simulation module in predicting the impact of each fault mode on the maintenance area are as follows: S3.1: Divide each steam turbine into a physical domain, including a thermal field domain, a mechanical field domain, a vibration field domain, and a flow field domain, and identify the coupling boundaries between each physical domain. Then, establish corresponding partial differential equations for each physical domain, and establish corresponding physical domain models based on the preset boundary conditions and initial states of each physical domain. At the same time, set the coupling boundary conditions of the coupling regions of each physical domain. S3.2: Extract observation data corresponding to different types of faults from the historical fault case library, and record the response fields of different subsystems in the turbine in each physical field under the collected fault observation data. Based on the collected fault observation data, take the corresponding fault mode and its spatial distribution as input data, take the multi-physics response caused by the fault in the corresponding spatiotemporal range as output data, and establish corresponding input-output training sample pairs. S3.3: Extract the spatial frequency distribution features of each input data through Fourier transform, expand the channels of the spatial frequency distribution features, and perform normalization processing. Then, perform global calculations on the processed spatial frequency distribution features through Fourier convolution kernels of different sizes. S3.4: After each set of Fourier convolution kernels completes global computation, the generated Fourier feature map is nonlinearly mapped, and the mapped Fourier feature map is used as the input of the next Fourier convolution kernel. Then, the Fourier feature map output by the last set of Fourier convolution kernels is compressed into the corresponding frequency domain response features through linear mapping. S3.5: Convert the obtained frequency domain response features into the spatial domain through inverse Fourier transform to obtain the corresponding response distribution prediction results. Calculate the error value between each response distribution prediction result and the corresponding output data. If the error value exceeds the preset threshold, optimize each Fourier convolution kernel, nonlinear mapping and various parameters in the linear mapping process through backpropagation based on the current error value. S3.6: Repeat the frequency domain response feature extraction and parameter update until the error value converges to the preset range. Then, use unused input-output training sample pairs to recalculate the error value to confirm that the error value is still within the preset range on the unknown data. If it exceeds the preset range, re-update the parameters. S3.7: Collect the latest physical field data of the steam turbine, and extract the corresponding response distribution prediction results based on the updated Fourier convolution kernels, nonlinear mapping and linear mapping process. Then calculate the risk score corresponding to each response distribution prediction result, and establish a risk heat map of the current steam turbine. At the same time, identify the risk subsystems and their impact paths with risk scores higher than the preset threshold, and supplement the identification results into the fault-maintenance knowledge graph.

6. The turbine maintenance decision optimization system based on fault cases according to claim 4, characterized in that, The specific steps of the fault cause deduction module in identifying the root cause and propagation path of the fault are as follows: S4.1: Extract structured entities and their relationships from the fault-maintenance graph, obtain the corresponding response distribution prediction results from the fault simulation results, map the collected data sets into variable node sets required for causal analysis, and establish causal edges between nodes in the variable node sets based on the response distribution prediction results. S4.2: Calculate the conditional independence between each pair of variable nodes, and based on the preset set of conditional variables, verify whether each variable node is independent under any preset conditional variable. If they are independent under any preset conditional variable, remove the causal edge between the corresponding two sets of variable nodes to establish a preliminary maintenance-fault causal graph. S4.3: Establish a structural scoring function for the current maintenance-failure cause-effect graph based on BIC or AIC to evaluate the balance between the goodness of fit of the current maintenance-failure cause-effect graph to each modality and the structural complexity. Then, perform various operations such as adding edges, deleting edges, and reversing directions on the current maintenance-failure cause-effect graph. After each modification to the graph structure, calculate a new score. If the score improves, accept the modification; otherwise, revert to the previous state. S4.4: Repeatedly iterate and optimize the maintenance-fault cause-effect graph until the change value of the score after multiple iterations converges to the preset range. After the iteration ends, output the optimal maintenance-fault cause-effect graph and correct the causal direction of each variable node in the maintenance-fault cause-effect graph based on physical prior knowledge and time sequence information. S4.5: Use a regression model to evaluate the strength of each causal edge in the maintenance-failure causal graph. Based on the failure simulation results, start from the target failure node, traverse backward along the maintenance-failure causal graph to the source node, and select the root cause nodes of each group whose causal edge strength is higher than the preset threshold and satisfies the preset physical constraints. S4.6: Starting from each root cause node, traverse forward along the causal direction to generate the corresponding fault propagation path, and generate the corresponding path description for each fault propagation path based on the semantic information in the fault-maintenance graph.

7. A turbine maintenance decision optimization method based on fault cases, used to implement the functions of the turbine maintenance decision optimization system based on fault cases as described in any one of claims 1-6, characterized in that, The specific steps of this optimization method are as follows: I. Collect and preprocess modal data from the turbine's operation monitoring system, historical maintenance records, and sensor acquisition terminals; II. The preprocessed modal data are fused and aligned, and a fault-maintenance map is constructed based on the aligned modal data; III. Based on the fault-maintenance map, establish physical field models for each part of the turbine, predict the fault propagation path and its impact on the maintenance area, and generate an impact distribution map; IV. Based on the fault-maintenance diagram and impact distribution diagram, construct a fault causal network to identify the root causes and propagation paths of the faults; V. Based on the identification results, construct a maintenance decision tree and dynamically adjust the maintenance strategy, while outputting multiple candidate solutions according to different fault scenarios; VI. Transfer the optimal maintenance strategy of the source turbine to the target turbine, and simulate and preview the generated maintenance strategy in a virtual sandbox environment; VII. Based on simulation results, adjust and optimize maintenance paths in real time when resource or risk conditions change, and make natural selections based on actual performance and simulation results.

8. The turbine maintenance decision optimization method based on fault cases according to claim 7, characterized in that, The specific steps for dynamically adjusting the maintenance strategy described in step V are as follows: S5.1: Traverse the fault-maintenance cause-effect graph by depth-first search or breadth-first search, and obtain the set of all paths from the current fault node to the node of feasible maintenance measures, where each path represents a chain of cause-effect repair solutions; S5.2: Treat each path as a branch of the decision tree, where each decision node in the branch represents a maintenance measure selection point, and the leaf node represents the system state after maintenance is completed. Then, calculate the comprehensive evaluation value of each maintenance path and sort the maintenance paths according to the comprehensive evaluation value from smallest to largest. S5.3: When the operating environment, component status, available resources or security policies change, update the comprehensive evaluation value of each maintenance path, and reorder the path priorities based on the updated comprehensive evaluation value. At the same time, select the maintenance plan corresponding to the maintenance path ranked first as the optimal maintenance plan. S5.4: The maintenance plans corresponding to the maintenance paths whose comprehensive evaluation values ​​are within the preset range are taken as multi-path alternatives. After the maintenance plan is executed, the actual cost, risk events and maintenance results are recorded, and the recorded data are used in the next round of maintenance strategy generation module plan update.

9. The turbine maintenance decision optimization method based on fault cases according to claim 8, characterized in that, The specific steps for simulating and rehearsing the generated maintenance strategy in the virtual sandbox environment, as described in step VI, are as follows: S6.1: Based on the structure, material properties, and sensor data of a real steam turbine, construct a digital model of the steam turbine that includes the geometric and physical properties of the rotor system, bearing support, and sealing components, and then create an isolated simulation environment; S6.2: Import the generated multi-path maintenance scheme into the simulation environment, simulate the execution process of the multi-path maintenance scheme, and calculate the rotor dynamic balance deviation in real time during the simulation process of the multi-path maintenance scheme. S6.3: Select a set of uncertain parameters with randomness from the multi-path maintenance scheme and the turbine operating environment. Based on historical data statistics or engineering experience, set the probability distribution of each uncertain parameter. Then, perform independent or related sampling from each uncertain parameter probability distribution to obtain the full parameter vector required for one simulation. Repeat the sampling multiple times to establish the corresponding test set. S6.4: Based on each full parameter vector in the test set, run maintenance operation simulations in the simulation environment and the steam turbine digital model to obtain the performance index values ​​under the full parameter vector, and calculate the risk factor of each simulation output result. Then, based on all sampled risk factors, calculate the overall expected risk, and perform histogram statistics on each risk factor to obtain the probability density function corresponding to each risk factor. S6.5: Calculate the average dynamic balance deviation, overall expected risk, and corresponding maintenance time and resource consumption for each maintenance plan, and construct a performance evaluation table. If the overall expected risk of the maintenance plan exceeds the preset safety limit or the average dynamic balance deviation is higher than the preset threshold, adjust the maintenance steps or replace the plan, and feed the probability density function back to the maintenance strategy generation module for subsequent maintenance strategy updates.

10. The turbine maintenance decision optimization method based on fault cases according to claim 9, characterized in that, The specific calculation formula for the dynamic balance deviation mentioned in S6.2 is as follows: In the formula, The root mean square value represents the rotor dynamic balance deviation; This represents the total duration of the simulation evaluation; Represents the rotor's center of mass in time radial displacement; Represents the rotor's center of mass in time The desired rotor radial trajectory; The specific calculation formulas for the risk factors mentioned in S6.4 are as follows: In the formula, Representing the Risk factors for the simulation results; Representing the Performance metrics from the simulation; This represents the lower limit of the performance index; This represents the upper limit of the performance index; The specific formula for calculating the overall expected risk described in S6.4 is as follows: In the formula, Represents overall expected risk; Representing the Risk factors for the simulation results; This represents the total number of samples taken.

Images