A method for intelligent planning and auxiliary design of ship equipment maintenance and support tasks

Abstract

To address the shortcomings of existing technologies, this invention provides an intelligent planning and auxiliary design method for ship equipment maintenance and support tasks. The method includes: constructing a maintenance and support task element system for ship equipment, comprising maintenance and support operation sequences and attributes; collecting and annotating ship equipment maintenance technical documents based on this system, and constructing a maintenance and support task element graph; using the node data of this graph as learning samples for a graph convolutional neural network (Graph Convolutional Neural Network) for model training; and intelligently planning and auxiliary designing ship equipment maintenance and support tasks based on the trained model. This invention, based on a decision-making paradigm that integrates knowledge-driven and data-driven approaches, effectively ensures the temporal compliance and resource adaptability of operational processes, providing technical support for the full life-cycle support of ship equipment.

Description

Technical Field

[0001] This invention belongs to the field of ship maintenance technology, specifically relating to an intelligent planning and auxiliary design method for ship equipment maintenance and support tasks. Background Technology

[0002] Current ship equipment systems exhibit a high degree of integration, with various ship systems forming a composite architecture composed of multi-layered heterogeneous components. This complex equipment system exposes significant process coordination challenges in maintenance operations, namely, the overhaul of a single piece of equipment often requires disassembly and reassembly operations across multiple subsystems. Due to the strict logical dependencies between ship maintenance procedures, conventional manual scheduling methods are prone to resulting in logical breaks in procedures or misalignments of work nodes, severely impacting the efficiency of maintenance procedure development and the reliability of its execution.

[0003] From the perspective of maintenance engineering implementation, ship equipment support activities exhibit multi-dimensional coupling characteristics: at the temporal constraint level, each operational unit must follow a strict topological execution sequence. At the resource coordination level, it involves the dynamic matching of elements such as multi-disciplinary technical teams, specialized testing instruments, and heterogeneous support materials. This combination optimization problem under dual constraints of time, space, and resources exceeds the capability boundaries of traditional human experience-based decision-making, and generally suffers from systemic risks such as improper resource allocation and frequent process conflicts, severely restricting the improvement of the comprehensive support efficiency of marine equipment. Summary of the Invention

[0004] This invention addresses the problems existing in the prior art by providing an intelligent planning and auxiliary design method for ship equipment maintenance and support tasks, comprising:

[0005] Construct a system of maintenance and support task elements for ship equipment, including two aspects: maintenance and support operation sequence and maintenance and support operation attributes.

[0006] Based on the maintenance and support task element system for ship equipment, we collect and annotate ship equipment maintenance technical documents and construct a maintenance and support task element diagram.

[0007] Based on the maintenance and support task element graph, the nodes of the maintenance and support task element graph are preprocessed and then used as learning samples for graph convolutional neural network (GCNN) to train the model.

[0008] The trained model enables intelligent planning and assisted design of ship equipment maintenance and support tasks.

[0009] Furthermore, the maintenance and support operation sequence refers to the maintenance and support operation steps included in the maintenance task that have an irreversible sequential relationship, and the maintenance and support operation attributes are the attribute information possessed by the maintenance and support operation steps.

[0010] Furthermore, the maintenance and support operation attributes include:

[0011] The professional attributes of maintenance personnel refer to the professional skills required of maintenance personnel in the maintenance and support operation steps.

[0012] The skill level attribute of maintenance personnel refers to the requirements for the professional skill title or level of maintenance personnel in this maintenance and support operation step.

[0013] Support equipment / tool ​​attributes refer to the support equipment or tools required in this maintenance support operation step.

[0014] Spare parts and consumables refer to the spare parts and consumables required in this maintenance and support operation.

[0015] Furthermore, the method for collecting and annotating ship equipment maintenance technical documents based on the maintenance and support task element system for ship equipment includes:

[0016] Collect technical documents on ship equipment maintenance.

[0017] Based on the maintenance and support task element system for ship equipment, information on ship equipment is labeled from six dimensions: "maintenance and support operations", "maintenance and support operation steps", "maintenance personnel specialization", "maintenance personnel skill level", "support equipment / tools" and "spare parts and consumable materials".

[0018] Arranged in the order of the original technical documents from top to bottom.

[0019] Furthermore, the method for constructing the maintenance support task element graph includes: constructing a directed graph consisting of nodes and edges based on the labeled ship equipment maintenance technical documents, and connecting the directed graph based on the association relationship to obtain the maintenance support task element graph.

[0020] Furthermore, the nodes of the directed graph represent maintenance information.

[0021] The edges of the directed graph represent the relationships between maintenance information, that is, the sequence relationship between two adjacent maintenance support operations in a maintenance task.

[0022] Furthermore, the method for preprocessing the nodes of the maintenance and support task element map includes:

[0023] By embedding word vector features into the text information of each node in the maintenance and support task element graph, the text of each node in the maintenance and support task element graph is converted into word vectors.

[0024] Based on the transformed word vector data, index mapping information for each node is constructed and a numeric identifier is assigned. Index mapping information for the relationships between nodes is constructed and a mapping from source node to target node is created.

[0025] A feature dataset is created based on the processed information: nodes, node index mapping information, relationships between nodes, and relationship index mapping information.

[0026] The feature dataset is divided into training, validation, and test sets in a 7:2:1 ratio, which will serve as input information for subsequent models.

[0027] Furthermore, the graph convolutional neural network includes:

[0028] Input layer: Receives node features and relationship information between nodes.

[0029] The first convolutional layer performs the first dimensionality transformation on the information from the input layer and then connects it to the activation function.

[0030] The second convolutional layer further processes the data output from the first convolutional layer while maintaining the same data dimensions.

[0031] The output layer includes: link prediction branch and attribute prediction branch.

[0032] Furthermore, the link prediction branch is as follows: the 3D features of two nodes are concatenated and mapped to a 1D output through a linear layer, representing the probability of the existence of a relationship between adjacent nodes.

[0033] The attribute prediction branch directly maps the node data output from the convolutional layer of the second graph to a 4-dimensional output through a linear layer, representing the predicted probabilities of four maintenance attributes. These four maintenance attributes are: maintenance personnel professional attributes, maintenance personnel skill level attributes, support equipment / tool ​​attributes, and spare parts and consumable material attributes.

[0034] Furthermore, given a maintenance and support task element graph G=(V, E, A, R), where V is the set of nodes, E is the set of edges, A is the set of node types, and R is the set of edge types. For any node v∈V, its type is denoted as τ(v)∈A. For any edge e=(s, t)∈E, its type is denoted as... ϕ(e) ∈R. At this point, the graph convolutional neural network includes:

[0035] Input Layer: For each node type in the maintenance and support task element graph, a set of type-specific linear transformation layers is constructed, using Equation 1 to define an independent linear mapping function for each input node type τ.

[0036] Formula 1

[0037] In formula 1: The initial text embedding features for the node. A type-specific weight matrix. A type-specific bias vector. This represents the unified dimension latent feature after mapping.

[0038] A graph convolutional layer consists of two layers, and its computation process can be formally represented as follows:

[0039] Formula 2

[0040] In Formula 2: l Indicates the number of network layers. The type of the target node v is The set of neighboring nodes, For edge type Calculated attention weights, It is a ternary parameter matrix that depends on the source node type, target node type, and edge type. This is the ReLU activation function.

[0041] The output layer's attribute prediction branches include:

[0042] For any attribute node Its eigenvectors The output dimension is obtained by encoding the node text description by the bidirectional encoder module, and its processing method is shown in Equation 3:

[0043] Formula 3

[0044] In Equation 3, the embedding is based on a given maintenance step node. Then, the multi-class decoder predicts the category of each attribute as shown in Equation 4:

[0045] Formula 4

[0046] Formula 4 Given a maintenance step node, this is the predicted probability distribution vector of four attributes. Given the weight matrix of the maintenance step nodes, Given the bias vector of a maintenance step node, It is a four-dimensional vector space.

[0047] The link prediction branch of the output layer includes:

[0048] First, given the final embedding representations of two nodes u and v. ,in u For source node type, vFor the target node type, It is a 128-dimensional vector space.

[0049] Then, based on Equation 5, the two are concatenated to form a joint representation:

[0050] Formula 5

[0051] Formula 5 It is a 256-dimensional vector space.

[0052] Finally, the activation function layer maps the data to scalar scores, and the autoregressive decoder performs Equation 6 for decoding, thus predicting and generating the next task sequence text for this maintenance task.

[0053] Formula Six

[0054] in The text for predicting the next task sequence. This is the link weight matrix. This is the link bias vector.

[0055] Furthermore, the method for intelligent planning and assisted design of ship equipment maintenance and support tasks based on the trained model includes: completion of maintenance and support task elements and prediction of unknown maintenance and support task elements.

[0056] The completion of maintenance and support task elements includes:

[0057] The model is trained by using the task sheet as the seed node input.

[0058] The trained model compares and analyzes the relevant information in the task sheet from two aspects: maintenance and support operation sequence and maintenance and support operation attributes. It also fills in the missing elements and predicts the probability distribution of the supplemented element information.

[0059] The Softmax function is used to normalize the supplementary feature information, and the feature information with the highest probability is selected as the prediction result to complete the task sheet.

[0060] The prediction of unknown maintenance and support task elements includes:

[0061] The model is trained by using the currently known maintenance and support operation steps as seed nodes.

[0062] The trained model obtains possible steps for the next maintenance and support operation as prediction steps, and predicts the probability that each prediction step is the next maintenance and support operation step.

[0063] When there is a prediction step with a prediction probability greater than a set threshold, it is determined that there is a maintenance sequence association, and the prediction step is output.

[0064] The obtained prediction steps are used as seed nodes to input into the trained model. The above process is repeated to generate a maintenance task sequence and calculate the joint probability of the entire path until the link prediction probability is lower than the set threshold and / or the number of inference steps reaches the preset upper limit, thereby completing the prediction of unknown maintenance support task elements.

[0065] The advantages of this invention are as follows: Addressing the multidimensional complexity of ship maintenance engineering, this invention proposes a novel artificial intelligence method that integrates domain knowledge representation and temporal reasoning. By constructing a ship maintenance ontology knowledge base, core elements such as equipment disassembly and assembly logic and process constraints can be structurally represented. Combining graph embedding technology to deeply model the spatiotemporal relationships between maintenance elements enables the establishment of an intelligent derivation mechanism for maintenance process chains. This decision-making paradigm, based on the fusion of knowledge-driven and data-driven approaches, provides an end-to-end optimized path for maintenance solution generation, effectively ensuring the temporal compliance and resource adaptability of the work process, and providing technical support for the full life-cycle maintenance of modern ship equipment. Attached Figure Description

[0066] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation thereof. Obviously, those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0067] Figure 1 A schematic diagram of the basic process of the model constructed for an embodiment of the present invention;

[0068] Figure 2 The chart shows the loss variation (a) and accuracy variation (b) during the training process of the HG-BART model constructed for this embodiment of the invention.

[0069] Figure 3 This is a schematic diagram illustrating the steps involved in the repair manual information for this invention – replacing circuit boards for repair and maintenance.

[0070] Figure 4 This is a schematic diagram of a specific structure of a model according to an embodiment of the present invention. Detailed Implementation

[0071] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0072] This invention provides, by way of example, an intelligent planning and auxiliary design method for ship equipment maintenance and support tasks, comprising:

[0073] Construct a system of maintenance and support task elements for ship equipment, including two aspects: maintenance and support operation sequence and maintenance and support operation attributes.

[0074] Based on the maintenance and support task element system for ship equipment, we collect and annotate ship equipment maintenance technical documents and construct a maintenance and support task element diagram.

[0075] Based on the maintenance and support task element graph, the nodes of the maintenance and support task element graph are preprocessed and then used as learning samples for graph convolutional neural network (GCNN) to train the model.

[0076] The trained model enables intelligent planning and assisted design of ship equipment maintenance and support tasks.

[0077] This invention addresses the multidimensional complexity of ship maintenance engineering by proposing a novel artificial intelligence method that integrates domain knowledge representation and temporal reasoning. By constructing a ship maintenance ontology knowledge base, core elements such as equipment disassembly and assembly logic and process constraints can be structurally represented. Combined with graph embedding technology, deep modeling of the spatiotemporal relationships between maintenance elements enables the establishment of an intelligent derivation mechanism for maintenance process chains. This decision-making paradigm, based on the fusion of knowledge-driven and data-driven approaches, provides an end-to-end optimized path for maintenance solution generation, effectively ensuring the temporal compliance and resource adaptability of operational processes, and providing technical support for the full life-cycle maintenance of modern ship equipment.

[0078] This invention provides an exemplary maintenance support operation sequence, which consists of maintenance support operation steps with an irreversible sequential relationship included in a maintenance task.

[0079] This invention provides an example of a maintenance and support operation attribute, which is: attribute information possessed in the maintenance and support operation steps.

[0080] This invention provides, by way of example, specific content of maintenance and support operation attributes, including:

[0081] The professional attributes of maintenance personnel refer to the professional skills required of maintenance personnel in the specific maintenance and support procedures. Examples include mechanical engineering and electronics specialties.

[0082] The skill level attribute of maintenance personnel refers to the professional skill title or level required for maintenance personnel in this maintenance and support operation procedure. For example: junior, intermediate, and senior.

[0083] Support equipment / tool ​​attributes refer to the support equipment or tools required in this maintenance and support operation. Examples include: Phillips screwdrivers, Allen wrenches, and explosion-proof open-end wrenches.

[0084] Spare parts and consumables refer to the spare parts and consumables required in this maintenance and support operation. Examples include spare circuit boards, anti-reverse gaskets, and anhydrous alcohol swabs.

[0085] Based on relevant national (military) standards and expert experience, this invention proposes a maintenance support task analysis system for ship equipment maintenance activities. This system models the maintenance support task analysis system for ship equipment maintenance activities from two perspectives: maintenance support operation sequence and maintenance support operation attributes.

[0086] This invention provides, by way of example, a method for collecting and annotating ship equipment maintenance technical documents based on a maintenance and support task element system for ship equipment, including:

[0087] Collect technical documents on ship equipment maintenance.

[0088] Based on the task element system for ship equipment maintenance, the ship equipment maintenance technical documents are labeled with information from six dimensions: "maintenance and support operations", "maintenance and support operation steps", "maintenance personnel specialties", "maintenance personnel skill levels", "support equipment / tools" and "spare parts and consumable materials".

[0089] Arranged in the order of the original technical documents from top to bottom.

[0090] The present invention provides an exemplary method for constructing a maintenance support task element graph, comprising: constructing a directed graph consisting of nodes and edges based on annotated ship equipment maintenance technical documents; and connecting the directed graph based on association relationships to obtain the maintenance support task element graph.

[0091] The present invention provides, by way of example, a node in the directed graph that represents maintenance information.

[0092] The present invention provides an example of an edge in a directed graph that represents the association between maintenance information, namely the sequential relationship between two adjacent maintenance support operations in a maintenance task.

[0093] Compared with existing technologies, by constructing a maintenance support task element graph that covers two dimensions—maintenance support operation sequence and attributes—the temporal constraints and resource matching requirements in maintenance activities are formally expressed. Based on graph data training, the model realizes end-to-end association modeling of maintenance links and attribute elements, laying the knowledge and model foundation for maintenance task planning.

[0094] This invention provides, by way of example, a method for preprocessing data of nodes in a maintenance and support task element diagram, comprising:

[0095] By embedding word vector features into the text information of each node in the maintenance and support task element graph, the text of each node in the maintenance and support task element graph is converted into word vectors.

[0096] Based on the transformed word vector data, index mapping information for each node is constructed and a numeric identifier is assigned. Index mapping information for the relationships between nodes is constructed and a mapping from source node to target node is created.

[0097] A feature dataset is created based on the processed information: nodes, node index mapping information, relationships between nodes, and relationship index mapping information.

[0098] The feature dataset is divided into training, validation, and test sets in a 7:2:1 ratio, which will serve as input information for subsequent models.

[0099] This invention provides an exemplary graph convolutional neural network, comprising:

[0100] Input layer: Receives node features and relationship information between nodes.

[0101] The first convolutional layer performs the first dimensionality transformation on the information from the input layer and then connects it to the activation function.

[0102] The second convolutional layer further processes the data output from the first convolutional layer while maintaining the same data dimensions.

[0103] The output layer includes: link prediction branch and attribute prediction branch.

[0104] The present invention provides an exemplary link prediction branch, which involves concatenating the dimensional features of two nodes and mapping them to a 1-dimensional output through a linear layer, representing the probability of the existence of a relationship between adjacent nodes.

[0105] This invention provides an exemplary attribute prediction branch, which involves directly mapping the node data output from the second graph's convolutional layer to a 4-dimensional output via a linear layer, representing the predicted probabilities of four maintenance attributes. The four maintenance attributes are: maintenance personnel professional attributes, maintenance personnel skill level attributes, support equipment / tool ​​attributes, and spare parts and consumable material attributes.

[0106] This invention provides an exemplary model structure based on a given maintenance and support task element graph G=(V, E, A, R), where V is the set of nodes, E is the set of edges, A is the set of node types, and R is the set of edge types; for any node v∈V, its type is denoted as τ(v)∈A; for any edge e=(s, t)∈E, its type is denoted as... ϕ(e) ∈R. At this point, the graph convolutional neural network is as follows: Figure 4 As shown, it includes:

[0107] Input Layer: For each node type in the maintenance and support task element graph, a set of type-specific linear transformation layers is constructed, using Equation 1 to define an independent linear mapping function for each input node type τ.

[0108] Formula 1

[0109] In formula 1: The initial text embedding features for the nodes; A type-specific weight matrix; A type-specific bias vector; This represents the unified dimension latent feature after mapping.

[0110] A graph convolutional layer consists of two layers, and its computation process can be formally represented as follows:

[0111] Formula 2

[0112] In Formula 2: l Indicates the number of network layers; The type of the target node v is The set of neighboring nodes, For edge type Calculated attention weights, A ternary parameter matrix that depends on the source node type, target node type, and edge type; This is the ReLU activation function.

[0113] The output layer's attribute prediction branches include:

[0114] For any attribute node Its eigenvectors The output dimension is obtained by encoding the node text description by the bidirectional encoder module, and its processing method is shown in Equation 3:

[0115] Formula 3

[0116] In Equation 3, the embedding is based on a given maintenance step node. Then, the multi-class decoder predicts the category of each attribute as shown in Equation 4:

[0117] Formula 4

[0118] Formula 4 Given a maintenance step node, this is the predicted probability distribution vector of four attributes. Given the weight matrix of the maintenance step nodes, Given the bias vector of a maintenance step node, It is a four-dimensional vector space.

[0119] The link prediction branch of the output layer includes:

[0120] First, given the final embedding representations of two nodes u and v. ,in u For source node type, v For the target node type, It is a 128-dimensional vector space.

[0121] Then, based on Equation 5, the two are concatenated to form a joint representation:

[0122] Formula 5

[0123] Formula 5 It is a 256-dimensional vector space.

[0124] Finally, the activation function layer maps the data to scalar scores, and the autoregressive decoder performs Equation 6 for decoding, thus predicting and generating the next task sequence text for this maintenance task.

[0125] Formula Six

[0126] in The text for predicting the next task sequence. This is the link weight matrix. This is the link bias vector.

[0127] The model constructed in this invention has high link prediction accuracy on the test set and also shows good inference performance in maintenance attribute matching. This indicates that the model can learn the structural constraints and semantic association information contained in the element graph of maintenance support task well, and can provide a more accurate reference scheme for actual maintenance task planning.

[0128] This invention provides an exemplary method for intelligent planning and assisted design of ship equipment maintenance and support tasks based on the trained model, including: completing maintenance and support task elements and predicting unknown maintenance and support task elements.

[0129] This invention provides an exemplary method for completing maintenance and support task elements, including...

[0130] The model is trained by using the task sheet as the seed node input.

[0131] The trained model compares and analyzes the relevant information in the task sheet from two aspects: maintenance and support operation sequence and maintenance and support operation attributes. It also fills in the missing elements and predicts the probability distribution of the supplemented element information.

[0132] The supplementary element information is normalized, and the element information with the highest probability is selected as the prediction result to complete the task list.

[0133] This invention provides, by way of example, a method for predicting unknown maintenance support task elements, including:

[0134] The model is trained by using the currently known maintenance and support operation steps as seed nodes.

[0135] The trained model obtains possible steps for the next maintenance and support operation as prediction steps, and predicts the probability that each prediction step is the next maintenance and support operation step.

[0136] When there is a prediction step with a prediction probability greater than a set threshold, it is determined that there is a maintenance sequence association, and the prediction step is output.

[0137] The obtained prediction steps are used as seed nodes to input into the trained model. The above process is repeated to generate a maintenance task sequence and calculate the joint probability of the entire path until the link prediction probability is lower than the set threshold and / or the number of inference steps reaches the preset upper limit, thereby completing the prediction of unknown maintenance support task elements.

[0138] Given the initial task, the model can automatically complete the subsequent task sequence and match the corresponding attribute elements. The generated maintenance plan is highly consistent with the results of manual review, indicating that this method can significantly improve the task planning efficiency of maintenance managers and realize intelligent maintenance task reasoning and planning based on similar product models.

[0139] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0140] Step 1: Construct a system of maintenance and support task elements for ship equipment

[0141] Ensure the standardization and normalization of maintenance support analysis processes. Based on relevant national (military) standards and expert experience, a maintenance support task element system for ship equipment is proposed. This system models the maintenance support task analysis system for ship equipment maintenance activities from two perspectives: maintenance support operation sequence and maintenance support operation attributes. Specific details are as follows:

[0142] The maintenance support operation sequence refers to the maintenance support operation steps included in a maintenance task, and these maintenance steps have an irreversible sequential relationship. The maintenance support operation attribute refers to the attribute information possessed in a maintenance support operation step, including four attributes: maintenance personnel specialty, maintenance personnel skill level, support equipment / tools, spare parts and consumable materials. The specific descriptions of these four attributes are shown in Table 1.

[0143] Table 1. Description of Maintenance and Support Operation Attributes

[0144]

[0145] Step 2: Construction of Task Element Diagram for Ship Equipment Maintenance and Support

[0146] Based on the maintenance and support task element system for ship equipment constructed in step 1, and driven by ship equipment maintenance technical documents, a maintenance and support task element graph for ship equipment is constructed, providing the dataset input for step 3 based on a graph model. Specific details are as follows:

[0147] Step 2.1: Collect and annotate ship equipment maintenance technical document data

[0148] Based on the maintenance and support task element system for ship equipment constructed in step 1, ship equipment maintenance technical documents are collected and labeled. The ship equipment maintenance technical documents are labeled with information from six dimensions: "maintenance and support operations", "maintenance and support operation steps", "maintenance personnel specialties", "maintenance personnel skill levels", "support equipment / tools" and "spare parts and consumable materials". They are arranged from top to bottom in the original order of the technical documents and saved in CSV file format.

[0149] Step 2.2: Construction of Maintenance Support Task Element Diagram

[0150] Based on collected and labeled technical documentation data on ship equipment maintenance, a directed graph was constructed using the Python programming language and the NetworkX library. The directed graph consists of two main parts: nodes and edges. The nodes represent maintenance information, specifically six key categories: "Maintenance Support Operations," "Maintenance Support Operation Steps," "Maintenance Personnel Specialty," "Maintenance Personnel Skill Level," "Support Equipment / Tools," and "Spare Parts and Consumables." These six categories are distinguished by six different colors. The edges represent the relationships between maintenance information. They can represent the sequential relationship between two adjacent maintenance support operations within a maintenance task. For example, in the task of "replacing the magnetic starter junction box transformer," the relationship between maintenance support operation steps ① and ② can be represented as ①→②. They can also represent the dependency relationship between a maintenance support operation step and its maintenance attribute. For instance, the maintenance support operation "replacing the magnetic starter junction box transformer" is related to the maintenance personnel specialty "Electrical Specialty," which can be represented as "Replacing the magnetic starter junction box transformer" → "Electrical Specialty."

[0151] Specifically, firstly, based on a pre-established maintenance support task analysis system for ship equipment maintenance activities, and driven by ship equipment maintenance technical documents, the Python programming language and NetworkX library are used to traverse each row of data in a CSV file, adding maintenance support operations as main nodes in the graph, and adding corresponding maintenance support operation steps as child nodes for each maintenance support operation. Simultaneously, directed edges are established from maintenance support operations to operation steps to represent the dependency relationship between them. For each maintenance support operation step, maintenance support operation attribute information (maintenance personnel specialty, skill level, support equipment, spare parts, and consumable materials) is added as a new node and connected to the corresponding maintenance support operation step node. Directed edges are established between adjacent maintenance support operation steps within the same maintenance support operation to represent the execution order of the steps. Index mapping information is constructed for the nodes and edges of the directed graph, assigning a numeric identifier to each node and creating a source node to target node mapping for each edge. Subsequently, PyG data objects are created using the processed node features, edge features, and their index information. The PyG data object is a data structure specifically designed for graph neural networks. It can encapsulate the topological structure and feature information of a graph, integrating text word vectors and their node and edge information into a 768-dimensional tensor data format, thus forming a maintenance and support task element graph dataset. Finally, the constructed feature dataset is split into training, validation, and test sets in a 7:2:1 ratio, providing feature dataset information as input for subsequent GNN model training.

[0152] Step 3: Training the Graph Convolutional Neural Network Model

[0153] Since the maintenance support task element graph constructed by this invention is essentially a heterogeneous graph structure, containing various types of nodes and semantic edge relationships of different types, traditional graph neural networks such as GNN and GAT have inherent defects when processing such heterogeneous graphs. That is, they cannot effectively distinguish the semantic differences between different types of nodes and edges and adopt a unified weighting mechanism in the information aggregation process. Therefore, it is easy to weaken or lose the heterogeneous features contained in maintenance domain knowledge.

[0154] This invention addresses the heterogeneous nature of knowledge in ship equipment maintenance and support tasks. By introducing bidirectional encoders and autoregressive decoders in the processing of heterogeneous graphs and node text, an improved Transformer model for heterogeneous graphs is obtained. Independent parameterized representation spaces are designed for different types of nodes and edges. By dynamically learning the importance weights of heterogeneous neighbors using an attention mechanism, the model can accurately capture complex cross-type and cross-level dependencies in maintenance tasks while preserving node type semantics. The core objective is to explicitly model the heterogeneity of node and edge types during graph convolution.

[0155] Step 3.1: HG-BART Model Architecture Design for Maintenance and Support Tasks

[0156] Given a maintenance and support task element graph G=(V,E,A,R), where V is the set of nodes, E is the set of edges, A is the set of node types, and R is the set of edge types (relationships). For any node v∈V, its type is denoted as τ(v)∈A; for any edge e=(s,t)∈E, its type is denoted as ϕ(e)∈R. This invention constructs an HG-BART model for maintenance and support task element completion based on the PyTorch deep learning framework and the PyTorchGeometric graph neural network library. The specific construction process is as follows:

[0157] The model architecture includes the following core components:

[0158] (1) Type-aware feature encoding layer: For the six node types in the maintenance and support task graph, a set of type-specific linear transformation layers is first constructed. Using the PyTorch nn.ModuleDict data structure, an independent linear mapping function is defined for each input node type τ:

[0159] Formula 1

[0160] in, The initial BERT text embedding features for the nodes; A type-specific weight matrix; A type-specific bias vector; This represents the unified dimension latent feature after mapping. In this invention, ∈R 768 R 768 It is a 768-dimensional vector space; ∈R 128×768 R 128×768 It is a cross vector space of 768 dimensions and 128 dimensions; ∈R 128 R 128 It is a 128-dimensional vector space; ∈R 128 R 128 It is a 128-dimensional vector space.

[0161] (2) Multi-layer heterogeneous graph convolution module: The core uses two heterogeneous graph convolutional layers (HGConv) for information propagation and feature aggregation. The calculation process of the HGConv layer can be formally represented as:

[0162] Formula 2

[0163] Where l represents the number of network layers, which is set to 2 layers in this model; The type of the target node v is The set of neighboring nodes, For edge type Calculated attention weights, σ is a ternary parameter matrix that depends on the source node type, target node type, and edge type; σ is the activation function, and this model uses the ReLU activation function.

[0164] (3) Bidirectional encoder-autoregressive decoder hybrid architecture for multi-task learning: In order to simultaneously achieve the two major goals of link prediction and attribute completion of maintenance task elements, this method designs an attribute prediction branch based on the bidirectional encoder module and a link prediction branch based on the autoregressive decoder module, which are described in detail below:

[0165] 1) Attribute Prediction Branch: Used to complete the four types of attribute information for maintenance steps (personnel specialty, skill level, support equipment, spare parts). For any attribute node... Its eigenvectors ∈R 128 The node text description is obtained by encoding the node text description using a bidirectional encoder module, where Given a 128-dimensional vector space acceptable to the bidirectional encoder module, the output dimension is processed as follows:

[0166] Formula 3

[0167] Embedded based on given maintenance step nodes The model predicts the category of each attribute using a multi-class decoder in the manner shown in Equation 4:

[0168] Formula 4

[0169] in Given a maintenance step node, this is the predicted probability distribution vector of four attributes. Given the weight matrix of the maintenance step nodes, Given the bias vector of a maintenance step node, It is a four-dimensional vector space.

[0170] 2) Link Prediction Branch: Used to predict and generate potential relationships in a maintenance task sequence. Given the final embedding representation of two nodes u (source node type, such as step) and v (target node type, such as support equipment). This model first concatenates the two to form a joint representation:

[0171] Formula 5

[0172] Formula 5 It is a 256-dimensional vector space.

[0173] Subsequently, the data is mapped to scalar scores through an activation function layer, and then decoded using Equation 6 via an autoregressive decoder to predict and generate the next task sequence text for this maintenance task.

[0174] Formula Six

[0175] in The text for predicting the next task sequence. This is the link weight matrix. This is the link bias vector.

[0176] The final HG-BART model constructed for maintenance and support task elements is as follows: Figure 4 As shown, the basic operation flow of this model is as follows: Figure 1 As shown, this design allows the model to learn both the structural information of the graph (for link prediction) and the semantic information of the nodes (for attribute prediction), thus providing comprehensive predictive capabilities in maintenance and support tasks.

[0177] Step 3.2: HG-BART Model Training Evaluation

[0178] Based on the HG-BART model constructed in step 3.1, the Adam optimizer of the nn.Module class in the PyTorch library was used for training, with the learning rate set to 0.01. After training the model for 200 epochs, the final model training test results showed a training loss of 0.4403 and a training accuracy of 0.7925. The validation loss was 0.5147 and the validation accuracy was 0.7453. The training process is as follows: Figure 2 As shown. The test set divided in step 3.1 was used for testing, and the test results are as follows: the test loss is 0.4951 and the test precision is 0.7740.

[0179] To further explore the impact of each module of the HG-BART model on performance, an ablation study was conducted using a heterogeneous graph convolutional neural network (HGCNN) as the baseline model. This study included a heterogeneous transformer model (decoder only) and a heterogeneous BERT model (encoder only). The final experimental results are shown in Table 2.

[0180] Table 2 Comparison of ablation experiments

[0181]

[0182] Experimental results show that the complete HG-BART model significantly outperforms the HGCNN baseline and the two ablation models across all four evaluation metrics. Specifically, the decoder-only HG-Transformer model has an F1 score that is approximately 7.76% lower and an accuracy that is approximately 7.53% lower than the complete HG-BART model, indicating that the absence of the encoder module severely limits the model's ability to extract and understand complex structural information from heterogeneous graphs. On the other hand, while the encoder-only HG-BERT model outperforms the baseline, its F1 score is still approximately 3.34% lower and its accuracy that is approximately 2.75% lower than HG-BART, reflecting the indispensable role of the decoder in inference and sequence generation tasks based on encoded information.

[0183] The encoder's role is to deeply fuse and learn representations of multi-level features of nodes and relationships in heterogeneous graphs, providing high-quality information condensation for downstream tasks. The decoder, on the other hand, uses this representation for autoregressive generation or task inference, forming a collaborative "understanding-inference" mechanism. Ablation experiments show that removing any part disrupts this information processing loop, leading to a decrease in the model's expressive power, and further verifying the necessity of the encoder-decoder joint architecture in heterogeneous graph representation learning.

[0184] Step 4: Intelligent planning and auxiliary design for ship equipment maintenance and support tasks

[0185] Complete existing maintenance and support task elements:

[0186] Based on the HG-BART model constructed in step 3, and within the framework established in step 1, predictions are made using existing maintenance and support tasks as the target, thereby supplementing the missing elements of existing maintenance and support tasks (maintenance personnel expertise, maintenance personnel skill levels, support equipment / tools, and spare parts and consumables). This step first inputs the currently known maintenance operation steps as seed nodes into the HG-BART model. Using the encoder function defined in step 3.1, the 768-dimensional input features are mapped to a 64-dimensional node embedding representation through two layers of graph convolution. Secondly, this step outputs a 4-dimensional vector through a 64-dimensional node embedding linear mapping layer, corresponding to the predicted probability distributions of the four attributes: maintenance personnel expertise, maintenance personnel skill levels, support equipment / tools, spare parts, and consumables. The Softmax function is used to normalize the outputs of each attribute category, selecting the category with the highest probability as the prediction result, thus ultimately completing the element supplementation for existing maintenance and support tasks.

[0187] Prediction of unknown maintenance and support task elements:

[0188] Based on the HG-BART model constructed in step 3, and within the architecture established in step 1, the next unknown maintenance and support task and its corresponding elements are predicted using existing maintenance and support tasks as the object. This step first inputs the currently known maintenance operation steps as seed nodes into the HG-BART model. Using the encoder function defined in step 3.1, the 768-dimensional input features are mapped to a 64-dimensional node embedding representation through two layers of graph convolution. Next, the link prediction branch is used to infer the next maintenance operation step. The 64-dimensional embedding of the current node is concatenated with the 64-dimensional embedding of the candidate node to form a 128-dimensional edge feature vector. This vector is compressed into a 1-dimensional output through a linear mapping layer, and after passing through a Sigmoid activation function, the link existence probability value in the [0,1] interval is obtained. A probability threshold of 0.5 is set; when the predicted probability is greater than this threshold, a maintenance sequence association is determined, thus achieving the prediction of unknown maintenance and support tasks. Subsequently, for newly added maintenance operation steps, the attribute prediction branch completes their four types of maintenance attribute information. Based on the predicted probability distribution, the most matching specific value is selected for each attribute, such as "Electrical Specialty," "Junior," and "Phillips Screwdriver." Finally, the newly generated maintenance task steps and their attributes are used as known information and re-inputted into the HG-BART model for the next round of inference. For the generated maintenance task sequence, the joint probability of the entire path is calculated. Let the predicted probability of the link from step i to step i+1 be... The confidence score of the entire maintenance sequence path containing n steps is:

[0189]

[0190] Repeat the link prediction and attribute prediction process until the termination conditions are met. The termination conditions include: 1. The link prediction probability falls below a set threshold, indicating that the end of the maintenance sequence has been reached. 2. The number of inference steps reaches a preset upper limit to prevent infinite loops.

[0191] Furthermore, this invention provides an application process example based on a specific embodiment:

[0192] Based on a maintenance support operation "replacing a circuit board" in the test set, information from the maintenance manual was consulted, and maintenance technical document data was obtained, such as... Figure 3 As shown.

[0193] Then, based on the maintenance and support task element system for ship equipment constructed in step 1, the steps in step 2, such as collecting and annotating ship equipment maintenance technical document data and constructing the maintenance and support task element diagram for ship equipment, are executed to annotate the associated maintenance tasks and corresponding elements and construct the maintenance and support task element diagram.

[0194] Execute step 3, input the maintenance support task element map established above as a learning sample into the HG-BART model for training and debugging. Then, based on the debugging results, input the first maintenance support operation step of the task, "open the frame with a screwdriver". After converting the information of this step into word vector features and loading it into the HG-BART model, the attribute reasoning information of the "replace board" step is completed, as shown in Table 3.

[0195] Table 3. Complete Existing Maintenance Support Task Elements Based on the HG-BART Model

[0196]

[0197] Secondly, the unknown maintenance and support task elements were predicted based on the HG-BART model: the first maintenance operation step of the task was input as "open the frame with a screwdriver". After converting the step information into word vector features and loading it into the HG-BART model, the unknown maintenance and support task was inferred as "press the power off button to ensure that the system is in a normal shutdown state", and the inferred maintenance and support task element was "electrical profession, primary, explosion-proof Phillips screwdriver". The results are shown in Table 4.

[0198] Table 4. Prediction of unknown maintenance support task elements based on the HG-BART model

[0199]

[0200] Based on the predicted next operation step and its attribute information, the data is loaded into the HG-BART model for prediction and reasoning of subsequent steps. This process continues iteratively, and after repeated inferences, the reasoning results for the entire maintenance operation "replacing the circuit board" are shown in Table 5.

[0201] Table 5. Reasoning Results for “Replacing the Board”

[0202]

[0203] After manual review, the final results are shown in Table 6. All the predictions for maintenance and support operation steps were correct, with only slight deviations in steps ② and ⑤ regarding the reasoning for support equipment and spare parts. This indicates that the model's reasoning effect is good and can assist maintenance designers in quickly formulating large-scale equipment maintenance and support tasks.

[0204] Table 6 Results after manual review

[0205]

[0206] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for intelligent planning and auxiliary design of ship equipment maintenance and support tasks, characterized in that, include: Construct a system of maintenance and support task elements for ship equipment, including two aspects: maintenance and support operation sequence and maintenance and support operation attributes; Based on the task element system for maintenance and support of ship equipment, ship equipment maintenance technical documents are collected and labeled, and a maintenance and support task element diagram is constructed; among which: The method for collecting and annotating ship equipment maintenance technical documents based on the ship equipment maintenance and support task element system includes: Collect technical documents on ship equipment maintenance; Based on the task element system for ship equipment maintenance support, the ship equipment maintenance technical documents are labeled with information from six dimensions: "maintenance operation", "maintenance operation steps", "maintenance personnel specialty", "maintenance personnel skill level", "support equipment / tools" and "spare parts and consumables". Arranged from top to bottom according to the original technical document order; The method for constructing the maintenance support task element graph includes: constructing a directed graph consisting of nodes and edges based on the labeled ship equipment maintenance technical documents; and connecting the directed graph based on the association relationship to obtain the maintenance support task element graph. The nodes of the directed graph represent maintenance information; The edges of the directed graph represent the sequence relationship between two adjacent maintenance operations in a maintenance task; Based on the maintenance and support task element graph, the nodes of the maintenance and support task element graph are preprocessed and then used as learning samples for graph convolutional neural network (GCNN) to train the model. Based on the trained model, the system completes the maintenance and support task elements and predicts unknown maintenance and support task elements in the compilation of ship equipment maintenance and support tasks. The completion of maintenance and support task elements includes: The model is trained by inputting the task sheet as a seed node. The trained model compares and analyzes the relevant information in the task sheet from two aspects: maintenance and support operation sequence and maintenance and support operation attributes. It also fills in the missing elements and predicts the probability distribution of the supplemented element information. The supplementary element information is normalized, and the element information with the highest probability is selected as the prediction result to complete the task list. The prediction of unknown maintenance and support task elements includes: The model is trained by inputting the currently known maintenance operation steps as seed nodes; The trained model obtains possible steps for the next maintenance operation as prediction steps, and predicts the probability that each prediction step is the next maintenance operation step. When there is a prediction step with a prediction probability greater than a set threshold, it is determined that there is a maintenance sequence association, and the prediction step is output. The obtained prediction steps are used as seed nodes to input into the trained model. The above process is repeated to generate a maintenance task sequence and calculate the joint probability of the entire path until the link prediction probability is lower than the set threshold and / or the number of inference steps reaches the preset upper limit, thereby completing the prediction of unknown maintenance support task elements.

2. The intelligent planning and auxiliary design method for ship equipment maintenance and support tasks according to claim 1, characterized in that, The maintenance support operation sequence is a maintenance operation step that has an irreversible sequential relationship within a maintenance task, and the maintenance support operation attribute is the attribute information possessed by the maintenance operation step.

3. The intelligent planning and auxiliary design method for ship equipment maintenance and support tasks according to claim 2, characterized in that, The maintenance and support operation attributes include: The professional attributes of maintenance personnel refer to the professional skills required of maintenance personnel in this maintenance operation step. The skill level attribute of maintenance personnel refers to the requirements of the professional skill title or level of maintenance personnel in this maintenance operation step. Support equipment / tool ​​attributes refer to the support equipment or tools required for this maintenance operation step. Spare parts and consumables refer to the spare parts and consumables required in this maintenance operation.

4. The intelligent planning and auxiliary design method for ship equipment maintenance and support tasks according to claim 1, characterized in that, The method for preprocessing the nodes of the maintenance and support task element map includes: Embed word vector features into the text information of each node in the maintenance and support task element diagram, thereby converting the text of each node in the maintenance and support task element diagram into word vectors. Based on the word vector data obtained from the transformation, index mapping information of each node is constructed and a numerical identifier is assigned to each node. Index mapping information of the relationship between nodes is constructed and a mapping from source node to target node is created. A feature dataset is created based on the processed information: nodes, node index mapping information, relationships between nodes, and relationship index mapping information. The feature dataset is divided into training, validation, and test sets in a 7:2:1 ratio, which will serve as input information for subsequent models.

5. The intelligent planning and auxiliary design method for ship equipment maintenance and support tasks according to claim 1, characterized in that, The graph convolutional neural network includes: Input layer: Receives node features and information about relationships between nodes; The first convolutional layer: performs the first dimensionality transformation on the information from the input layer and then connects it to the activation function; The second convolutional layer further processes the data output from the first convolutional layer while maintaining the same data dimensions. The output layer includes: link prediction branch and attribute prediction branch.

6. The intelligent planning and auxiliary design method for ship equipment maintenance and support tasks according to claim 5, characterized in that, The link prediction branch is as follows: the 3D features of two nodes are concatenated and mapped to a 1D output through a linear layer, representing the probability of the existence of a relationship between adjacent nodes; The attribute prediction branch is as follows: the node data output by the convolutional layer of the second graph is directly mapped to the 4-dimensional output through a linear layer, representing the prediction probability of four maintenance attributes; the four maintenance attributes are: maintenance personnel professional attributes, maintenance personnel skill level attributes, support equipment / tool ​​attributes, and spare parts and consumable material attributes.

7. The intelligent planning and auxiliary design method for ship equipment maintenance and support tasks according to claim 6, characterized in that, Given a maintenance and support task element graph G=(V, E, A, R), where V is the set of nodes, E is the set of edges, A is the set of node types, and R is the set of edge types; for any node v∈V, its type is denoted as τ(v)∈A; for any edge e=(s,t)∈E, its type is denoted as... ϕ(e) ∈R; at this point, the graph convolutional neural network includes: Input Layer: For each node type in the maintenance and support task element graph, a set of type-specific linear transformation layers is constructed, using Equation 1 to define an independent linear mapping function for each input node type τ. Set 1 In formula 1: The initial text embedding features for the nodes; A type-specific weight matrix; A type-specific bias vector; The unified dimension latent features after mapping; A graph convolutional layer consists of two layers, and its computation process can be formally represented as follows: Formula 2 In Formula 2: l Indicates the number of network layers; The type of the target node v is The set of neighboring nodes, For edge type Calculated attention weights, A ternary parameter matrix that depends on the source node type, target node type, and edge type; It is the ReLU activation function; The output layer's attribute prediction branches include: For any attribute node Its eigenvectors The output dimension is obtained by encoding the node text description by the bidirectional encoder module, and its processing method is shown in Equation 3: Formula 3 In Equation 3, the embedding is based on a given maintenance step node. Then, the multi-class decoder predicts the category of each attribute as shown in Equation 4: Formula 4 Formula 4 Given a maintenance step node, this is the predicted probability distribution vector of four attributes. Given the weight matrix of the maintenance step nodes, Given the bias vector of a maintenance step node, It is a four-dimensional vector space; The link prediction branch of the output layer includes: First, given the final embedding representations of two nodes u and v. ,in u For source node type, v For the target node type, It is a 128-dimensional vector space; Then, based on Equation 5, the two are concatenated to form a joint representation: Formula 5 Formula 5 It is a 256-dimensional vector space; Finally, the activation function layer maps the data to scalar scores, and the autoregressive decoder performs Equation 6 for decoding, thus predicting and generating the next task sequence text for this maintenance task. Formula Six in The text for predicting the next task sequence. This is the link weight matrix. This is the link bias vector.

Images