A diesel engine maintenance guarantee method based on deep learning and data analysis

Abstract

This invention discloses a diesel engine maintenance and support method based on deep learning and data analysis, specifically including: collecting diesel engine operating data, maintenance history, parts replacement records, IETM (Integrated Equipment Management System) and maintenance process data; establishing a parts association graph, calculating the edge weights for lifespan decay, and then writing them into the parts association graph; inputting the association graph into a DCRNN (Digital Relationship Graph Neural Network) to generate node health scores and maintenance priority sequences; establishing a flowchart and knowledge graph, creating nodes and associated edges according to part number, fault type, operation stage, and tool number; writing the task list and priority sequence into the flowchart-knowledge dual-graph neural network, matching it with work instructions, 3D models, operation videos, and tool descriptions; generating a maintenance execution path, pushing it to the terminal, collecting completion status, time consumption, number of views, and rework identifiers to form a feedback record; and dynamically adjusting the weights of parts association edges and flowchart-knowledge nodes based on the feedback. This invention achieves intelligent maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance technology for diesel engines, and in particular to a diesel engine maintenance and support method based on deep learning and data analysis. Background Technology

[0002] In current diesel engine maintenance management, traditional maintenance relies on manual experience and regular maintenance plans, mainly through manual recording of operating parameters, maintenance history, and parts replacement information. This approach suffers from problems such as scattered data, information lag, and difficulty in quantitatively assessing the health status of parts, leading to maintenance decisions that depend on experience-based judgment and lack scientific basis.

[0003] Existing methods have limited ability to analyze complex assembly relationships and the collaborative effects of multiple components, and cannot achieve priority ranking of maintenance tasks and optimal allocation of resources. In actual operation, the health status of components is time-varying and uncertain due to the influence of load, number of start-stop cycles, operating time and environmental factors. Relying solely on regular inspections is insufficient to detect potential faults in a timely manner, which can easily lead to sudden downtime or increased rework.

[0004] In addition, knowledge and information such as maintenance procedures, work instructions, tool usage and operation videos are scattered, and manual retrieval is inefficient and difficult to form a structured association. This leads to problems such as slow information retrieval, non-standard operation and insufficient risk warnings for maintenance personnel during the execution process.

[0005] Therefore, there is a need for an intelligent maintenance method that can integrate operational data, maintenance history, component status, and process knowledge to achieve component health assessment, priority maintenance sequencing, task path generation, and dynamic optimization, thereby improving diesel engine maintenance efficiency, reducing failure rate, and providing a scientific basis for maintenance decisions. Summary of the Invention

[0006] One objective of this invention is to propose a diesel engine maintenance and support method based on deep learning and data analysis. This invention introduces a component association graph and a process-knowledge dual-graph neural network to perform multi-dimensional feature encoding and health score calculation on diesel engine operating data, maintenance history, and component replacement records. Combined with DCRNN diffusion convolution and cyclic updates, it generates maintenance task recommendations and priority component sequences. Through process-knowledge graph matching, it achieves precise association between work instructions, 3D models, operation videos, and tool descriptions, constructing a dynamic closed-loop maintenance execution and feedback optimization process. It has the advantages of accurate maintenance task sorting, visualized execution path, timely response, and scientific component health assessment.

[0007] A diesel engine maintenance and support method based on deep learning and data analysis according to an embodiment of the present invention includes the following steps: Collect diesel engine operating data, maintenance history, parts replacement records, IETM content, and maintenance process data; Establish a component association graph, read the cumulative runtime, number of start-stops, load level, most recent replacement time and number of failures, calculate the lifespan decay edge weights, and write them into the component association graph; The component association graph is written into the DCRNN diffusing convolutional recurrent network, and diffusing convolution and recurrent updates are performed to generate maintenance task recommendations and priority component sequences. Establish flowcharts and knowledge graphs, and create association edges between process nodes and knowledge nodes according to component number, fault type, operation stage, and tool number; The recommended maintenance tasks and priority component sequences are written into the process-knowledge dual-graph neural network, and the target process nodes are matched with the corresponding work instructions, 3D models, operation videos, tool instructions and risk warnings. Generate a maintenance task execution path, push the matching content to the maintenance terminal, collect the task completion status, step time, number of times the content is viewed, rework identification and maintenance results, and form a feedback record; The weights of component associations and the weights of process-knowledge node associations are dynamically adjusted based on feedback records.

[0008] Optionally, the collection of diesel engine operating data, maintenance history, parts replacement records, IETM content, and maintenance process data specifically includes: Diesel engine operating data, including real-time parameters such as engine speed, load, temperature, pressure, vibration, and fuel consumption; maintenance history, including historical maintenance records, maintenance cycles, maintenance types, and maintenance personnel information; parts replacement records, including parts number, installation / removal time, service life, and reason for replacement; IETM content, including work instruction text, 3D models of parts, operation videos, and tool instructions; and maintenance process data, including work phase divisions, operation steps, process node numbers, and work tool codes.

[0009] Optionally, establishing the component association diagram includes the following steps: Read the component numbers and their corresponding assembly relationships, and construct a set of component nodes; Obtain the cumulative runtime, number of start-stop cycles, load level, last replacement time, and number of failures for each component; Calculate the component life decay weights based on operation and maintenance information to form an inter-node edge weight matrix; Write the node set and edge weight matrix into the component association graph to complete the graph structure initialization.

[0010] Optionally, the step of writing the component association graph into the DCRNN diffusing convolutional recurrent network and performing diffusing convolution and recurrent updates specifically includes: The component node set and edge weight matrix are segmented and mapped to the DCRNN input layer nodes. The node features include cumulative runtime, number of start-stop times, load level, recent replacement time, and number of failures. The edge weight matrix defines the strength of the relationship between nodes. The graph convolution weight matrix and the recurrent unit state vector are initialized. The node state vector and the adjacent edge weight are weighted and summed to form the initial node hidden representation. Iteratively perform diffusion convolution. In each round of convolution, the hidden representation of the node is weighted and aggregated according to the adjacent edge weights. The aggregation result is concatenated with the node's own features and input into the nonlinear transformation unit to generate the updated node representation. Repeat the diffusion convolution operation to reach the set number of layers. The result of the diffusion convolution is input into the recurrent unit, and the node state is updated by combining it with time series information. The weighted fusion of historical state and current feature is controlled by recurrent gating to generate the temporal evolution hidden state of each node. The temporal evolution hidden state is input into the output mapping layer to generate the node health score vector and maintenance priority score sequence. After each iteration, the node state vector and edge weight matrix are updated, and the information transmission strength between nodes is adjusted. At the end of the iteration, the comprehensive health score vector of all nodes, edge weight matrix update information, and maintenance task sorting sequence are output to form maintenance task recommendations and priority execution component sequences.

[0011] Optionally, the process of establishing flowcharts and knowledge graphs, and creating association edges between process nodes and knowledge nodes according to component number, fault type, operation stage, and tool number, includes the following steps: The component number, fault type, operation stage, and tool number are combined into a set of node attributes. Construct flowchart nodes, each node corresponding to the operation steps of a specific component in a specific work stage. Node attributes include work stage identifier, step number, and tool number. Construct knowledge graph nodes, with each node corresponding to the content of the work instruction book, 3D model, operation video and tool instructions. Node attributes include document number, model number, video number and tool number. Establish sequential edges between process nodes based on component numbers and work stages; these sequential edges reflect the order of operation steps. Establish association edges between process nodes and knowledge nodes based on component number, fault type, and tool number. The association edges identify the guidance documents, models, videos, and tool information corresponding to each operation step. Perform attribute validation on all nodes and edges; Establish the connection edges between the flowchart and the knowledge graph to form a graph structure for matching the dual-graph neural network.

[0012] Optionally, the step of writing maintenance task recommendations and priority component sequences into the process—knowledge dual-graph neural network specifically includes: Map the recommended maintenance tasks to flowchart nodes, and mark the node indexes according to the part number, work stage, and step number; Map priority component sequences to knowledge graph nodes, and label node indices according to component number, fault type, and tool number; Establish a mapping matrix between flowchart nodes and knowledge graph nodes, with each mapping element recording the flowchart node index, knowledge node index, and task model pairing identifier; The mapping matrix and the hidden node representation are written into the input layer of the dual-graph neural network. The node features include health score vector, priority repair score and node attribute set. Initialize the parameters of the dual-graph neural network, including the message passing weight matrix, node state vectors, and edge feature matrix; The message passing path is defined according to the node index and mapping matrix. The process node information is passed to the knowledge node according to the mapping relationship. The edge weight is adjusted by the task recommendation score and priority component sorting. The dual-graph neural network is iteratively updated. In each iteration, the node receives information along the path of the mapping matrix. The received features of neighboring nodes are concatenated with the node's own features and input into the nonlinear transformation unit to generate the updated node state. Repeat the iteration until the node state converges, update the node hidden representation and the mapping edge weights to form the final matching representation of the process-knowledge node; After the iteration is completed, the output node matching results and task allocation sequence are generated. The matching nodes are indexed and bound with the corresponding work instructions, 3D models, operation videos, tool instructions and risk warnings to form a task execution table.

[0013] Optionally, the formation of the feedback record includes the following steps: The maintenance task execution path generates a task execution table on the maintenance terminal, marking the part number, operation stage and operation step number corresponding to each task node; Collect the completion status of each task node, and record whether the operation is completed, the completion time, and the time consumed by each step; Count the number of times each task node is viewed and mark whether the work instructions, 3D models, operation videos and tool descriptions have been accessed; Mark rework indicators, set rework marks for failed or abnormal operation steps, and record the number of reworks and the time; Record maintenance results, including changes in component status, troubleshooting details, and task completion indicators; The completion status, step time, number of views, rework identifier, and repair results are generated into a structured feedback record table according to the task node index; Verify the consistency between the verification feedback records and the recommended maintenance tasks and priority component sequence; Structured feedback records are written into the database to provide input data for updating the associated edge weights of components and the associated weights of process-knowledge nodes.

[0014] Optionally, the step of dynamically adjusting the component association weights and process-knowledge node association weights based on feedback records specifically includes: The completion status, step time, number of views, rework identifier and maintenance result of each task node in the structured feedback record table are indexed and matched with the nodes in the component association graph. The health decay correction value of the corresponding edge of the component node is calculated. The edge weight is adjusted according to the cumulative running time, number of failures and rework identifier. The corrected edge weight is written back to the edge weight matrix of the component association graph. Map the task node execution status in the feedback record to the process-knowledge node association matrix. Calculate the weight correction coefficient for each mapping edge according to the task completion status, step time consumption, and rework identifier. Update the process-knowledge node association weight and write it into the dual-graph neural network mapping matrix. Perform consistency checks on the updated component association diagram and process-knowledge node association matrix, check the integrity of node indexes and the validity of edge weight ranges, and complete the dynamic adjustment operation.

[0015] The beneficial effects of this invention are: By combining DCRNN diffraction convolution with cyclic updates and component association graphs, a quantitative assessment of the health status of each component is achieved, generating a priority maintenance component sequence to ensure reasonable sorting of maintenance tasks and improve task allocation accuracy.

[0016] Based on a process-knowledge dual-graph neural network, maintenance tasks are precisely linked with work instructions, 3D models, operation videos, and tool descriptions, enabling visualized retrieval of maintenance information, reducing manual search time, and lowering operational risks.

[0017] By collecting task execution feedback, including step time, number of views, and rework identifiers, the associated edge weights of components and the associated weights of process-knowledge nodes are dynamically adjusted to continuously optimize maintenance task recommendations and execution paths, thereby improving maintenance efficiency.

[0018] By comprehensively considering the operating parameters of components, maintenance history and replacement records, the system can calculate the health degradation of components and dynamically update the edge weights, effectively reducing the probability of sudden failures and rework, and improving the overall operational reliability of the diesel engine.

[0019] By inputting structured multi-source maintenance data into a neural network, quantifiable health scores and maintenance priorities can be achieved, providing managers with a scientific basis and supporting refined and intelligent diesel engine maintenance decisions. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a diesel engine maintenance and support method based on deep learning and data analysis proposed in this invention. Figure 2 This is a schematic diagram illustrating the DCRNN health score and maintenance priority generation of a diesel engine maintenance and support method based on deep learning and data analysis proposed in this invention. Figure 3 This is a schematic diagram of dual-graph matching and dynamic feedback optimization for a diesel engine maintenance and support method based on deep learning and data analysis proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A diesel engine maintenance and support method based on deep learning and data analysis includes the following steps: Collect diesel engine operating data, maintenance history, parts replacement records, IETM content, and maintenance process data; Establish a component association graph, read the cumulative runtime, number of start-stops, load level, most recent replacement time and number of failures, calculate the lifespan decay edge weights, and write them into the component association graph; The component association graph is written into the DCRNN diffusing convolutional recurrent network, and diffusing convolution and recurrent updates are performed to generate maintenance task recommendations and priority component sequences. Establish flowcharts and knowledge graphs, and create association edges between process nodes and knowledge nodes according to component number, fault type, operation stage, and tool number; The recommended maintenance tasks and priority component sequences are written into the process-knowledge dual-graph neural network, and the target process nodes are matched with the corresponding work instructions, 3D models, operation videos, tool instructions and risk warnings. Generate a maintenance task execution path, push the matching content to the maintenance terminal, collect the task completion status, step time, number of times the content is viewed, rework identification and maintenance results, and form a feedback record; The weights of component associations and the weights of process-knowledge node associations are dynamically adjusted based on feedback records.

[0023] In this embodiment, the collection of diesel engine operating data, maintenance history, parts replacement records, IETM content, and maintenance process data specifically includes: Diesel engine operating data, including real-time parameters such as engine speed, load, temperature, pressure, vibration, and fuel consumption; maintenance history, including historical maintenance records, maintenance cycles, maintenance types, and maintenance personnel information; parts replacement records, including parts number, installation / removal time, service life, and reason for replacement; IETM content, including work instruction text, 3D models of parts, operation videos, and tool instructions; and maintenance process data, including work phase divisions, operation steps, process node numbers, and work tool codes.

[0024] In this embodiment, establishing the component association diagram includes the following steps: Read the component numbers and their corresponding assembly relationships, and construct a set of component nodes; Obtain the cumulative runtime, number of start-stop cycles, load level, last replacement time, and number of failures for each component; Calculate the component life decay weights based on operation and maintenance information to form an inter-node edge weight matrix; Write the node set and edge weight matrix into the component association graph to complete the graph structure initialization.

[0025] Specifically, obtaining the cumulative runtime, start / stop count, load level, most recent replacement time, and number of failures for each component includes: Read the part number and corresponding assembly information of each component, extract parameters such as engine speed, load, temperature, pressure, vibration and fuel consumption from the real-time operation monitoring system, and calculate the total running time and start-stop times of each component. Statistical analysis of component load levels under different load ranges and operating conditions; query maintenance records to obtain the most recent replacement time and service life; Count the number of historical faults from the fault logs; The above data is integrated to form the state vector of each component, which is used for the initialization of the node features of the component association graph and the calculation of the edge weights for life decay.

[0026] Specifically, the calculation of component lifespan degradation weights based on operation and maintenance information to form the inter-node edge weight matrix includes: For each component, read the cumulative runtime, number of start-stops, load level, last replacement time, and number of historical failures; map the cumulative runtime, number of start-stops, and load level to a standardized lifespan decay coefficient, and use a weighted function to calculate the impact of operating load on lifespan; The remaining service life ratio is assessed based on the most recent replacement time and service life; the health status is corrected by combining the number of historical failures, and a component health score is generated; the edge weights of mutually assembled or functionally dependent component nodes are calculated according to the combination relationship of the health scores of the two nodes to reflect the potential failure propagation risk between nodes; the edge weights of all components are integrated to form an inter-node edge weight matrix, which is used as the weight of the edges in the component association graph and input into the DCRNN network to realize the transmission and iterative update of life decay information in the graph structure.

[0027] In this embodiment, writing the component association graph into the DCRNN diffusing convolutional recurrent network and performing diffusing convolution and recurrent updates specifically includes: The component node set and edge weight matrix are segmented and mapped to the DCRNN input layer nodes. The node features include cumulative runtime, number of start-stop times, load level, recent replacement time, and number of failures. The edge weight matrix defines the strength of the relationship between nodes. The graph convolution weight matrix and the recurrent unit state vector are initialized. The node state vector and the adjacent edge weight are weighted and summed to form the initial node hidden representation. Iteratively perform diffusion convolution. In each round of convolution, the hidden representation of the node is weighted and aggregated according to the adjacent edge weights. The aggregation result is concatenated with the node's own features and input into the nonlinear transformation unit to generate the updated node representation. Repeat the diffusion convolution operation to reach the set number of layers. The result of the diffusion convolution is input into the recurrent unit, and the node state is updated by combining it with time series information. The weighted fusion of historical state and current feature is controlled by recurrent gating to generate the temporal evolution hidden state of each node. The temporal evolution hidden state is input into the output mapping layer to generate the node health score vector and maintenance priority score sequence. After each iteration, the node state vector and edge weight matrix are updated, and the information transmission strength between nodes is adjusted. At the end of the iteration, the comprehensive health score vector of all nodes, edge weight matrix update information, and maintenance task sorting sequence are output to form maintenance task recommendations and priority execution component sequences.

[0028] Specifically, mapping the component node set and edge weight matrix to DCRNN input layer nodes in segments includes: Based on the functional modules or assembly subsystems of each component of the diesel engine, the entire set of component nodes is divided into several segments, each segment corresponds to a local subgraph, and each subgraph contains several component nodes and their edge weight information. The node feature vectors of each subgraph segment, including cumulative runtime, number of starts and stops, load level, recent replacement time and number of failures, are mapped to the corresponding node positions in the DCRNN input layer. At the same time, the edge weights between nodes in the subgraph are mapped to the adjacency matrix of the input layer. For each subgraph node, initialize the hidden state vector, and generate an initial hidden representation by weighting the node connection relationship and edge weight to ensure that the local structural information is effectively preserved. All subgraphs are sequentially input into the DCRNN network according to the module order. Node information is aggregated within the local subgraphs through diffraction convolution, while recurrent units enable cross-segment information transfer and global state updates. During the mapping process, node features are normalized to ensure that features of different segments and dimensions have a uniform numerical scale, providing input for subsequent diffusion convolution and iterative updates to generate health scores and maintenance priority sequences.

[0029] Specifically, iterative diffusion convolution includes: For each component node and edge weight matrix input into DCRNN, the weighted aggregation operation of the adjacent node features is performed on each node in sequence according to the preset number of convolutional layers and iteration rounds; In each round of convolution, the node's own feature vector and the hidden state of its neighboring nodes are read, and the neighbor information is weighted and summed in combination with the corresponding weight values ​​in the edge weight matrix to generate a local aggregated representation of the node. The local aggregated representation is concatenated with the node's own features and then input into a nonlinear transformation unit, such as an activation function or a fully connected layer, to generate an updated hidden representation of the node. Repeat the above operation until the set number of convolutional layers is reached, ensuring that the hidden representation of each node contains both its own information and information from multiple neighboring layers, thus achieving a comprehensive expression of local and global features. After each iteration, the node state can be normalized or regularized as needed to ensure the stability of feature values ​​and provide high-quality input for subsequent cyclic unit updates and health score calculations.

[0030] Specifically, the weighted fusion of historical states and current features using cyclic gating includes: For each component node, after the diffusion convolution is completed, the current hidden representation vector is obtained, and the historical hidden state vector of the previous time step loop unit is read. By setting up a gating structure, including an input gate, a forget gate, and an output gate, the current features, historical states, and final output are weighted and controlled respectively. The input gate determines the contribution ratio of the current hidden representation in the update, the forget gate controls the degree to which the historical state is retained or decayed, and the output gate adjusts the strength of the final output's propagation to the next time step. After multiplying the gating weights by the corresponding vector elements one by one, the summation operation is performed to achieve a weighted fusion of historical states and current features, generating the temporal evolution hidden state of the node.

[0031] Specifically, adjusting the information transmission strength between nodes includes: During the DCRNN diffusion convolution and recurrent update process, the weights of adjacent edges are dynamically adjusted based on the health score and state vector of each component node. For nodes with a health score below 0.7, the edge weights between them and key neighbor nodes are increased by 10% to 25% to enhance the transmission of potential fault information in the graph structure. For nodes with a health score higher than 0.9 and stable state, the edge weights between them and their neighboring nodes are reduced by 5% to 15% to reduce interference from irrelevant information and optimize information flow efficiency. By combining the functional dependencies and assembly relationships between nodes, when dynamically adjusting edge weights, it is ensured that critical path information is propagated first, and the adjustment of edge weights for non-critical nodes does not exceed 15%. After each iteration, the updated edge weight matrix is ​​written back to the component association graph for use in the next round of convolution and iterative iteration, thereby achieving continuous adaptive optimization of edge weights.

[0032] Specifically, the formation of maintenance task recommendations and priority component sequences includes: Based on the health score vector of each component node output by the DCRNN network, components with a health score less than 0.7 are listed as priority repair components; Calculate the criticality score of each component in the overall system, based solely on the health score and the edge weights between nodes: Criticality score = 1 - Health score + 0.5 × Average edge weight of the corresponding node in the edge weight matrix, where the lower the health score and the larger the node edge weight, the higher the criticality score; Components with a criticality score greater than 0.8 are prioritized and their execution order is generated by combining them with the health score. The sorting results will generate a recommended list of maintenance tasks. Each task will clearly record the part number, the module it belongs to, the health score, the critical score, and the maintenance sequence number. Prioritize the execution of component sequences and their correspondence with flowchart and knowledge graph nodes to ensure that each task clearly corresponds to the work instruction number, 3D model number, operation video number, and tool number. The output maintenance task recommendations and priority component sequences are provided for maintenance personnel to operate directly, making maintenance tasks more scientific, visual and efficient. This ensures that limited human resources are prioritized for handling components with health scores below 0.7 or criticality scores above 0.8, thereby reducing potential failure risks and improving the overall reliability of diesel engine operation. Based on actual feedback records, the task order and component priority are dynamically updated to achieve closed-loop optimization, and the recommended list is adjusted in real time as the component health scores change.

[0033] In this embodiment, establishing a flowchart and knowledge graph, and creating association edges between process nodes and knowledge nodes according to component number, fault type, operation stage, and tool number, includes the following steps: The component number, fault type, operation stage, and tool number are combined into a set of node attributes. Construct flowchart nodes, each node corresponding to the operation steps of a specific component in a specific work stage. Node attributes include work stage identifier, step number, and tool number. Construct knowledge graph nodes, with each node corresponding to the content of the work instruction book, 3D model, operation video and tool instructions. Node attributes include document number, model number, video number and tool number. Establish sequential edges between process nodes based on component numbers and work stages; these sequential edges reflect the order of operation steps. Establish association edges between process nodes and knowledge nodes based on component number, fault type, and tool number. The association edges identify the guidance documents, models, videos, and tool information corresponding to each operation step. Perform attribute validation on all nodes and edges; Establish the connection edges between the flowchart and the knowledge graph to form a graph structure for matching the dual-graph neural network.

[0034] Specifically, the attribute validation for all nodes and edges includes: Read each node and its attributes from the flowchart and knowledge graph, including part number, fault type, operation stage, tool number, operation instruction number, 3D model number, and operation video number; Check that the node number is unique, ensuring that the same number appears only once in the entire graph structure; Verify the integrity of node attributes to ensure that the required fields for each node, such as job stage identifier, step number, and tool number, are filled in and formatted correctly. Traverse all edges and check whether sequential edges are correctly connected according to the order of operation steps, and whether associated edges correspond to valid process nodes and knowledge nodes. Verify the numerical range of edge weights or associated identifiers to ensure that all edge weight values ​​are within the preset valid range, such as 0-1; For any inconsistencies or missing attributes found, a verification report is generated, and the corresponding nodes or edges are labeled for subsequent graph structure correction. After verification, the graph structure is confirmed to be complete, node attributes are complete, and edge weights are valid, ensuring the correct relationship between the flowchart and the knowledge graph, and providing reliable input for subsequent dual-graph neural network matching and task recommendation.

[0035] In this embodiment, the process of writing maintenance task recommendations and priority component sequences into the knowledge dual-graph neural network specifically includes: Map the recommended maintenance tasks to flowchart nodes, and mark the node indexes according to the part number, work stage, and step number; Map priority component sequences to knowledge graph nodes, and label node indices according to component number, fault type, and tool number; Establish a mapping matrix between flowchart nodes and knowledge graph nodes, with each mapping element recording the flowchart node index, knowledge node index, and task model pairing identifier; The mapping matrix and the hidden node representation are written into the input layer of the dual-graph neural network. The node features include health score vector, priority repair score and node attribute set. Initialize the parameters of the dual-graph neural network, including the message passing weight matrix, node state vectors, and edge feature matrix; The message passing path is defined according to the node index and mapping matrix. The process node information is passed to the knowledge node according to the mapping relationship. The edge weight is adjusted by the task recommendation score and priority component sorting. The dual-graph neural network is iteratively updated. In each iteration, the node receives information along the path of the mapping matrix. The received features of neighboring nodes are concatenated with the node's own features and input into the nonlinear transformation unit to generate the updated node state. Repeat the iteration until the node state converges, update the node hidden representation and the mapping edge weights to form the final matching representation of the process-knowledge node; After the iteration is completed, the output node matching results and task allocation sequence are generated. The matching nodes are indexed and bound with the corresponding work instructions, 3D models, operation videos, tool instructions and risk warnings to form a task execution table.

[0036] Specifically, the process of indexing and binding matching nodes with corresponding work instructions, 3D models, operation videos, tool descriptions, and risk warnings includes: Read the node matching results after the iterative update of the dual-graph neural network, including the flowchart node index and the knowledge graph node index; Based on the component number, fault type, operation stage, and tool number of each node, find the corresponding work instruction number, 3D model number, operation video number, and tool instruction number in the knowledge graph; A unique index record is generated for each matching node, and the flowchart node is bound to the knowledge graph node and its corresponding resource information to form a structured task execution table. Record the metadata of each binding item, including node index, resource number, binding timestamp, and binding status, so as to quickly retrieve and track it during maintenance. Perform a completeness check on all binding records to ensure that each process node has corresponding resource information, and mark missing or abnormal data for subsequent supplementation; Output the structured task execution table after completing index binding.

[0037] In this embodiment, generating a feedback record includes the following steps: The maintenance task execution path generates a task execution table on the maintenance terminal, marking the part number, operation stage and operation step number corresponding to each task node; Collect the completion status of each task node, and record whether the operation is completed, the completion time, and the time consumed by each step; Count the number of times each task node is viewed and mark whether the work instructions, 3D models, operation videos and tool descriptions have been accessed; Mark rework indicators, set rework marks for failed or abnormal operation steps, and record the number of reworks and the time; Record maintenance results, including changes in component status, troubleshooting details, and task completion indicators; The completion status, step time, number of views, rework identifier, and repair results are generated into a structured feedback record table according to the task node index; Verify the consistency between the verification feedback records and the recommended maintenance tasks and priority component sequence; Structured feedback records are written into the database to provide input data for updating the associated edge weights of components and the associated weights of process-knowledge nodes.

[0038] Specifically, the consistency between the verification feedback records and the recommended maintenance tasks and priority component sequences includes: Read the task execution feedback records collected by the maintenance terminal, including the completion status of each task node, operation time, number of views, rework identifier and maintenance result; According to the part number, operation stage and step number, the feedback record is matched with the maintenance task recommendation list and priority execution part sequence generated by the system; Check whether the execution results of each task node are consistent with the recommended task, such as whether the completion status meets the task requirements, whether the rework mark is reasonable, and whether the operation steps are complete. For task nodes that are inconsistent or abnormal, a verification report is generated, including the task node index, feedback data, recommended task information, and explanation of the differences, so as to facilitate subsequent processing and adjustment. Statistical indicators of overall consistency, such as completion matching rate, anomaly rate, and return-to-work rate, are used to evaluate the accuracy of maintenance execution and the rationality of task recommendations. The verification results are written into the system database to provide input data for the dynamic adjustment of the associated edge weights of components and the associated weights of process-knowledge nodes, thereby achieving closed-loop optimization.

[0039] In this implementation, dynamically adjusting the component association weights and process-knowledge node association weights based on feedback records specifically includes: The completion status, step time, number of views, rework identifier and maintenance result of each task node in the structured feedback record table are indexed and matched with the nodes in the component association graph. The health decay correction value of the corresponding edge of the component node is calculated. The edge weight is adjusted according to the cumulative running time, number of failures and rework identifier. The corrected edge weight is written back to the edge weight matrix of the component association graph. Map the task node execution status in the feedback record to the process-knowledge node association matrix. Calculate the weight correction coefficient for each mapping edge according to the task completion status, step time consumption, and rework identifier. Update the process-knowledge node association weight and write it into the dual-graph neural network mapping matrix. Perform consistency checks on the updated component association diagram and process-knowledge node association matrix, check the integrity of node indexes and the validity of edge weight ranges, and complete the dynamic adjustment operation.

[0040] Specifically, calculating the health decay correction value for the edge corresponding to the component node includes: Read the health score and historical operating data of each node in the component association diagram, including cumulative runtime, number of start-stop times, load level, last replacement time, and number of failures; For each edge between nodes, the attenuation impact coefficient is calculated based on the health scores of the two endpoints. For example, the formula is: Edge health attenuation correction value = 1 - (Node A health score × Node B health score); Based on the node's historical operation and fault information, the attenuation correction value is fine-tuned. For example, each time a fault or repair occurs, the edge attenuation value is increased by a fixed percentage to ensure that potential risks are reflected. The calculated health degradation correction value is weighted and updated with the original edge weight to obtain the corrected edge weight value, which is used to reflect the potential failure propagation risk between components. After performing a batch update on all edges, the corrected edge weight matrix is ​​written back to the component association graph to provide dynamic health information for subsequent iterative convolution and task recommendation. Example

[0041] To verify the feasibility of this invention, it was applied to a large-scale diesel engine maintenance management scenario. In this scenario, the daily operation of a diesel engine involves multiple engines, the operating states of each component are complex, and the wear of parts is time-varying. Traditional periodic maintenance methods struggle to accurately determine maintenance priorities, leading to low maintenance efficiency, delayed troubleshooting, and increased rework frequency. This invention collects diesel engine operating data, maintenance history, component replacement records, IETM content, and maintenance process data to establish a correlation graph for key components of each diesel engine. Edge weights are calculated based on cumulative runtime, start-stop count, load level, most recent replacement time, and number of failures. The correlation graph is input into a DCRNN diffusing convolutional recurrent network, and through multiple iterations, a health score and maintenance priority sequence for each component are generated, achieving scientific sorting of maintenance tasks.

[0042] During application, the recommended list of maintenance tasks and the sequence of priority components are mapped to nodes in the flowchart and knowledge graph. Combined with work instructions, 3D models, operation videos, and tool manuals, a process-knowledge dual-graph neural network is used for node matching to achieve precise association between work steps and resource information. Maintenance personnel operate according to the execution path generated by the system. The completion status, time consumed, number of views, and rework flags of each step are collected in real time and fed back to the system. This data is used to dynamically adjust the weights of component association edges and process-knowledge nodes, forming a closed-loop optimization mechanism.

[0043] In practical applications, the system continuously monitored and optimized the maintenance of 3,200 key components across 100 diesel engines. After 30 days of operation, the system-generated maintenance priority sequence, compared to manual experience-based prioritization, showed an 18% increase in task completion efficiency, a 12% reduction in the rework rate, and a 94% accuracy rate in component health scoring—23% higher than traditional periodic inspection methods. The maintenance execution path matching accuracy rate was 92%, and the average time operators spent consulting work instructions decreased from 5.5-6.8 minutes to 3.2-3.5 minutes. Overall maintenance response speed improved significantly, with average maintenance time reduced from 36-42 minutes to 33-36 minutes.

[0044] Table 1: Comparison of Traditional Diesel Engine Repair Methods and the Invention

[0045] The table shows a comparison of traditional periodic inspection methods and the present invention under the same number of diesel engines and key components, clearly demonstrating the advantages of the present invention in various maintenance indicators. Through the present invention, task completion efficiency is increased to 78%, an average improvement of approximately 16% to 20% compared to traditional methods. This means that maintenance personnel can complete more maintenance tasks in the same amount of time, increasing overall productivity. The rework rate is reduced to 11% to 14%, compared to 23% to 26% for traditional methods, showing that the system can rationally prioritize the maintenance of key components, promptly eliminate potential faults, and reduce the number of repeated repairs. The health score accuracy reaches 92% to 95%, significantly higher than the 68% to 71% of traditional methods, indicating that the DCRNN network can accurately reflect the health status of components, providing a reliable basis for maintenance decisions. The execution path matching accuracy is between 90% and 93%, ensuring a high degree of consistency between maintenance operations and work instructions, 3D models, and tool descriptions, effectively reducing operational errors. The average time for consulting work instructions is reduced from 6.5 to 6.8 minutes for traditional methods to 3.1 to 3.5 minutes, saving more than 50% of time and significantly improving operational efficiency. The average maintenance time was reduced from 36-42 minutes to 33-36 minutes, a decrease of approximately 10%-19% in operation time, thus improving overall maintenance efficiency. In summary, this invention, through intelligent analysis and closed-loop optimization, achieves scientific sequencing of maintenance tasks, precise matching of execution paths, and component health management, validating its effectiveness and practicality in improving maintenance efficiency, reducing rework rates, and optimizing operational resources.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A diesel engine maintenance and support method based on deep learning and data analysis, characterized in that, Includes the following steps: Collect diesel engine operating data, maintenance history, parts replacement records, IETM content, and maintenance process data; Establish a component association graph, read the cumulative runtime, number of start-stops, load level, most recent replacement time and number of failures, calculate the lifespan decay edge weights, and write them into the component association graph; The component association graph is written into the DCRNN diffusing convolutional recurrent network, and diffusing convolution and recurrent updates are performed to generate maintenance task recommendations and priority component sequences. Establish flowcharts and knowledge graphs, and create association edges between process nodes and knowledge nodes according to component number, fault type, operation stage, and tool number; The recommended maintenance tasks and priority component sequences are written into the process-knowledge dual-graph neural network, and the target process nodes are matched with the corresponding work instructions, 3D models, operation videos, tool instructions and risk warnings. Generate a maintenance task execution path, push the matching content to the maintenance terminal, collect the task completion status, step time, number of times the content is viewed, rework identification and maintenance results, and form a feedback record; The weights of component associations and the weights of process-knowledge node associations are dynamically adjusted based on feedback records.

2. The diesel engine maintenance and support method based on deep learning and data analysis according to claim 1, characterized in that, The data collected, including diesel engine operating data, maintenance history, parts replacement records, IETM content, and maintenance process data, specifically includes: Diesel engine operating data, including real-time parameters such as engine speed, load, temperature, pressure, vibration, and fuel consumption; Maintenance history, including historical maintenance records, maintenance cycles, maintenance types, and information on maintenance personnel; Parts replacement records, including part number, installation / removal time, service life, and reason for replacement; IETM content, including work instruction text, 3D model of parts, operation video, and tool instructions; Maintenance process data includes work phase divisions, operation steps, process node numbers, and work tool codes.

3. The diesel engine maintenance and support method based on deep learning and data analysis according to claim 2, characterized in that, The process of establishing the component association diagram includes the following steps: Read the component numbers and their corresponding assembly relationships, and construct a set of component nodes; Obtain the cumulative runtime, number of start-stop cycles, load level, last replacement time, and number of failures for each component; Calculate the component life decay weights based on operation and maintenance information to form an inter-node edge weight matrix; Write the node set and edge weight matrix into the component association graph to complete the graph structure initialization.

4. The diesel engine maintenance and support method based on deep learning and data analysis according to claim 3, characterized in that, The step of writing the component association graph into the DCRNN diffusing convolutional recurrent network and performing diffusing convolution and recurrent updates specifically includes: The component node set and edge weight matrix are segmented and mapped to the DCRNN input layer nodes. The node features include cumulative runtime, number of start and stop times, load level, recent replacement time and number of failures. The edge weight matrix defines the strength of the relationship between nodes. The graph convolution weight matrix and the recurrent unit state vector are initialized. The node state vector and the adjacent edge weight are weighted and summed to form the initial node hidden representation. Iteratively perform diffusion convolution. In each round of convolution, the hidden representation of the node is weighted and aggregated according to the adjacent edge weight. The aggregation result is concatenated with the node's own features and input into the nonlinear transformation unit to generate the updated node representation. Repeat the diffusion convolution operation to reach the set number of layers. The result of the diffusion convolution is input into the recurrent unit, and the node state is updated by combining it with time series information. The weighted fusion of historical state and current feature is controlled by recurrent gating to generate the temporal evolution hidden state of each node. The temporal evolution hidden state is input into the output mapping layer to generate the node health score vector and maintenance priority score sequence. After each iteration, the node state vector and edge weight matrix are updated, and the information transmission strength between nodes is adjusted. At the end of the iteration, the comprehensive health score vector of all nodes, edge weight matrix update information, and maintenance task sorting sequence are output to form maintenance task recommendation and priority execution component sequence.

5. The diesel engine maintenance and support method based on deep learning and data analysis according to claim 4, characterized in that, The process of establishing flowcharts and knowledge graphs, and creating association edges between process nodes and knowledge nodes according to component number, fault type, operation stage, and tool number, includes the following steps: The component number, fault type, operation stage, and tool number are combined into a set of node attributes. Construct flowchart nodes, each node corresponding to the operation steps of a specific component in a specific work stage. Node attributes include work stage identifier, step number, and tool number. Construct knowledge graph nodes, each node corresponding to the content of the work instruction book, 3D model, operation video and tool instructions. Node attributes include document number, model number, video number and tool number; Establish sequential edges between process nodes based on component numbers and work stages; these sequential edges reflect the order of operation steps. Establish association edges between process nodes and knowledge nodes based on component number, fault type, and tool number. The association edges identify the guidance documents, models, videos, and tool information corresponding to each operation step. Perform attribute validation on all nodes and edges; Establish the connection edges between the flowchart and the knowledge graph to form a graph structure for matching the dual-graph neural network.

6. The diesel engine maintenance and support method based on deep learning and data analysis according to claim 5, characterized in that, The process of writing maintenance task recommendations and priority component sequences into the knowledge dual-graph neural network specifically includes: Map the recommended maintenance tasks to flowchart nodes, and mark the node indexes according to the part number, work stage, and step number; Map priority component sequences to knowledge graph nodes, and label node indices according to component number, fault type, and tool number; Establish a mapping matrix between flowchart nodes and knowledge graph nodes, with each mapping element recording the flowchart node index, knowledge node index, and task model pairing identifier; The mapping matrix and the hidden node representation are written into the input layer of the dual-graph neural network. The node features include health score vector, priority repair score and node attribute set. Initialize the parameters of the dual-graph neural network, including the message passing weight matrix, node state vectors, and edge feature matrix; The message passing path is defined according to the node index and mapping matrix. The process node information is passed to the knowledge node according to the mapping relationship. The edge weight is adjusted by the task recommendation score and priority component sorting. The dual-graph neural network is iteratively updated. In each iteration, the node receives information along the path of the mapping matrix. The received features of neighboring nodes are concatenated with the node's own features and input into the nonlinear transformation unit to generate the updated node state. Repeat the iteration until the node state converges, update the node hidden representation and the mapping edge weights to form the final matching representation of the process-knowledge node; After the iteration is completed, the output node matching results and task allocation sequence are generated. The matching nodes are indexed and bound with the corresponding work instructions, 3D models, operation videos, tool instructions and risk warnings to form a task execution table.

7. The diesel engine maintenance and support method based on deep learning and data analysis according to claim 6, characterized in that, The process of generating feedback records includes the following steps: The maintenance task execution path generates a task execution table on the maintenance terminal, marking the part number, operation stage and operation step number corresponding to each task node; Collect the completion status of each task node, and record whether the operation is completed, the completion time, and the time consumed by each step; Count the number of times each task node is viewed and mark whether the work instructions, 3D models, operation videos and tool descriptions have been accessed; Mark rework indicators, set rework marks for failed or abnormal operation steps, and record the number of reworks and the time; Record maintenance results, including changes in component status, troubleshooting details, and task completion indicators; The completion status, step time, number of views, rework identifier, and repair results are generated into a structured feedback record table according to the task node index; Verify the consistency between the verification feedback records and the recommended maintenance tasks and priority component sequence; Structured feedback records are written into the database to provide input data for updating the associated edge weights of components and the associated weights of process-knowledge nodes.

8. The diesel engine maintenance and support method based on deep learning and data analysis according to claim 7, characterized in that, The specific steps of dynamically adjusting the component association weights and process-knowledge node association weights based on feedback records include: The completion status, step time, number of views, rework identifier and maintenance result of each task node in the structured feedback record table are indexed and matched with the nodes in the component association graph. The health decay correction value of the corresponding edge of the component node is calculated. The edge weight is adjusted according to the cumulative running time, number of failures and rework identifier. The corrected edge weight is written back to the edge weight matrix of the component association graph. Map the task node execution status in the feedback record to the process-knowledge node association matrix. Calculate the weight correction coefficient for each mapping edge according to the task completion status, step time consumption, and rework identifier. Update the process-knowledge node association weight and write it into the dual-graph neural network mapping matrix. Perform consistency checks on the updated component association diagram and process-knowledge node association matrix, check the integrity of node indexes and the validity of edge weight ranges, and complete the dynamic adjustment operation.

Images