Maintenance path reasoning method based on graph neural network

Abstract

The application discloses a kind of based on graph neural network's maintenance path inference method, belong to vehicle maintenance field, this method includes the following steps: using pre-training large language model to encode input vehicle maintenance record text, obtains the embedding representation of input text;The embedding representation of input text is combined with vehicle maintenance concept unique identifier and semantic relation to construct vehicle maintenance knowledge graph;Based on graph neural network, the node in vehicle maintenance knowledge graph is updated, and the updated node embedding representation is obtained;Based on the updated node embedding representation, the path embedding of the multi-hop inference path from the starting node is generated in a recursive manner, and the path embedding is scored and sorted by three-linear attention mechanism, and the most relevant path to the input text is selected.The application can more accurately capture the complex relationship between nodes by combining graph neural network and three-linear attention mechanism, and filter out the most relevant path to the input text.

Description

Technical Field

[0001] This invention belongs to the field of vehicle maintenance technology, and in particular relates to a maintenance path reasoning method based on graph neural networks. Background Technology

[0002] In the field of vehicle repair, knowledge graphs, as a structured knowledge representation tool, can effectively integrate complex information such as vehicle fault symptoms, repair items, and component relationships, assisting repair personnel in quickly locating problems and generating repair solutions. However, existing knowledge graph-based vehicle repair reasoning methods still have significant shortcomings, specifically in the following scenarios:

[0003] Insufficient multi-hop reasoning ability: Traditional graph neural networks (GNNs) struggle to effectively capture semantic associations between long-distance nodes when dealing with multi-hop relationships in knowledge graphs. For example, when a vehicle fault description involves multiple indirectly related components (such as "engine noise causing transmission shift delay"), existing models may only be able to identify single-hop repair items (such as checking the engine) while ignoring multi-hop paths (such as the associated path between the engine, drive shaft, and transmission), leading to the omission of crucial steps in the repair plan.

[0004] Inaccurate path relevance assessment: Existing attention mechanisms (such as multi-head attention) lack the ability to jointly model the complex relationships between input text, repair concepts, and paths when scoring candidate repair paths. For example, in response to a user description of "steering wheel vibration when braking," the model may incorrectly recommend a path directly related to "brake pad wear," while ignoring the implicit correlation path of "brake disc deformation - loose wheel bearing," resulting in a deviation between the repair plan and the actual needs.

[0005] Limited contextual semantic fusion: Traditional methods often rely on static word vectors to encode maintenance record text, which cannot fully capture the dynamic semantics of the context. For example, the term "oil leak" may refer to oil pan seal failure or oil filter damage in different scenarios. Existing models, lacking fine-grained semantic discrimination capabilities, are prone to confusing the maintenance paths of different faults.

[0006] Negative sample interference and low convergence efficiency: Existing contrastive learning strategies are not designed for the characteristics of the vehicle repair field, making it difficult to distinguish highly similar negative sample paths (such as "spark plug failure - ignition coil" and "spark plug failure - fuel injector"), resulting in slow model convergence and limited generalization ability.

[0007] Therefore, this invention proposes a maintenance path reasoning method based on graph neural networks. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a maintenance path reasoning method based on graph neural networks, thereby resolving the issues present in the prior art.

[0009] To achieve the above objectives, this invention provides a maintenance path reasoning method based on graph neural networks, comprising:

[0010] The input vehicle repair record text is encoded using a pre-trained large language model to obtain the embedded representation of the input text;

[0011] The embedded representation of the input text combines the unique identifier of the vehicle maintenance concept and semantic relationships to construct a vehicle maintenance knowledge graph;

[0012] The node embedding representation in the vehicle maintenance knowledge graph is updated based on the graph neural network to obtain the updated node embedding representation.

[0013] Based on the updated node embedding representation, a recursive approach is used to generate path embeddings for multi-hop inference paths starting from the initial node. The path embeddings are then scored and ranked using a trilinear attention mechanism to select the path most relevant to the input text.

[0014] Optionally, the process of obtaining the embedded representation of the input text includes:

[0015] Match the repair item concepts and unique identifiers of vehicle repair concepts in the input vehicle repair record text to identify all source PCD sets with unique identifiers;

[0016] The concept name and its associated text for each PCD are encoded using a bidirectional converter model to generate an initial concept representation;

[0017] The initial concept representation is weighted based on word frequency statistical features to obtain a weighted concept representation.

[0018] The weighted concept representation is iteratively updated by using a multi-layer graph isomorphic network, which aggregates neighborhood information and outputs an embedded representation of the input text.

[0019] Optionally, after generating the embedded representation of the input text, a weighting scheme based on term frequency-inverse document frequency is introduced to assign weights to high-frequency and semantically relevant PCDs.

[0020] Optionally, the expression for calculating the updated node embedding representation is:

[0021]

[0022] In the formula, h i (l) Let N(i) represent the embedding representation of node i at layer l, and let N(i) represent the set of neighboring nodes of node i. j,i (l) Let h represent the features of the edge from node j to node i, where ∈ is a learnable parameter.i (l+1) h represents the updated node embedding representation. j (l) Let represent the embedding representation of node j at layer l, and ReLU represent the ReLU activation function.

[0023] Optionally, the computational expression for the path embedding of the multi-hop inference path originating from the starting node is generated recursively as follows:

[0024]

[0025] In the formula, FFN() represents a feedforward network, and n is the subgraph G. src The number of jumps in p i p represents the path embedding of node i. i-1 h represents the path embedding of node i-1. i This represents the embedding representation of node i.

[0026] Optionally, the expression for scoring the path embedding using a trilinear attention mechanism is:

[0027]

[0028] S Tri =φ(Relu(σ(α) i )));

[0029] In the formula, h x h is an embedded representation of the input text. v To input the PCD list, It is an MLP layer, where αi is the original relevance score calculated by trilinear attention, and S Tri For the final path score after MLP normalization, a is the dimension index of the embedding representation of the input text, b is the dimension index of the input PCD list, and c is the path embedding p. i Dimension index, W abc Let σ be the trilinear attention weight tensor, and let σ denote the sigmoid activation function. x ) a h x scalar value in the a-th dimension, (h v ) b h v scalar value in the b-th dimension, (p i ) c Indicates the embedded path p i In the c-th dimension, ReLU represents the ReLU activation function.

[0030] Optionally, the method introduces a contrastive learning strategy to optimize the maintenance path reasoning process;

[0031] The optimization steps based on the contrastive learning strategy include:

[0032] Construct a set of positive sample paths and a set of negative sample paths;

[0033] Calculate the cosine similarity between the node embedding and the paths of positive and negative samples;

[0034] The model parameters are optimized by using a contrastive loss function that maximizes the cosine similarity of positive samples and minimizes the cosine similarity of negative samples.

[0035] Optionally, the total loss function of the maintenance path inference method includes: PCD prediction loss and contrastive learning loss.

[0036] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a maintenance path reasoning method based on a graph neural network.

[0037] The present invention also provides a storage medium on which a computer program is stored, which, when executed by a processor, implements a maintenance path reasoning method based on a graph neural network.

[0038] Compared with the prior art, the present invention has the following advantages and technical effects:

[0039] This invention presents a maintenance path reasoning method based on graph neural networks. By encoding vehicle maintenance record text using a pre-trained large language model, it efficiently extracts semantic information from the text and transforms it into embedded representations. A knowledge graph is constructed by combining unique identifiers for vehicle maintenance concepts and semantic relationships, further integrating professional knowledge in the vehicle maintenance field and providing a rich semantic background for reasoning. Representation updates of nodes in the knowledge graph are performed using graph neural networks, dynamically capturing complex relationships between nodes and generating multi-hop reasoning paths recursively. A trilinear attention mechanism scores and ranks the path embeddings, effectively selecting the most relevant paths to the input text, improving the accuracy and efficiency of reasoning. This method can quickly and accurately provide reasoning paths for vehicle maintenance, significantly improving the scientific rigor and reliability of maintenance decisions, and has significant practical application value. Attached Figure Description

[0040] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0041] Figure 1 This is a flowchart of the maintenance path reasoning method based on graph neural networks according to an embodiment of the present invention. Detailed Implementation

[0042] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0043] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0044] like Figure 1 As shown, this embodiment provides a maintenance path reasoning method based on graph neural networks. This method includes an innovative vehicle maintenance knowledge graph reasoning model that integrates graph neural networks and a trilinear attention mechanism. It focuses on generating maintenance item paths closely related to the input text through multi-hop reasoning to assist in tasks such as vehicle maintenance. Specifically, the model first encodes the input text and maintenance item concepts using a pre-trained large language model to obtain their embedding representations. These embedding representations contain semantic information of the text and concepts, providing rich context for the model. Then, the model updates the representations of nodes in the vehicle maintenance knowledge graph using a graph neural network, capturing the complex relationships between nodes and better understanding the structured information in the vehicle maintenance knowledge graph. Next, the model scores and ranks candidate paths using a trilinear attention mechanism, comprehensively considering the relationships between the input text, maintenance item concepts, and paths, and selecting the path most relevant to the input text. Finally, the model optimizes parameters through contrastive learning and other strategies to further improve model performance, making it more accurate and efficient in reasoning tasks. The collaborative work of these modules achieves efficient reasoning from input text to relevant maintenance item concept paths. This model not only provides strong support for tasks such as vehicle maintenance, but also demonstrates significant advantages in handling complex vehicle maintenance knowledge graphs. By combining graph neural networks and trilinear attention mechanisms, the model can more accurately capture the complex relationships between nodes and select the paths most relevant to the input text, thus providing new ideas and methods for the intelligent development of the vehicle maintenance field.

[0045] Specifically, this includes the following steps: encoding the input vehicle repair record text using a pre-trained large language model to obtain the embedded representation of the input text; constructing a vehicle repair knowledge graph by combining the embedded representation of the input text with the unique identifier of the vehicle repair concept and semantic relationships; updating the representation of nodes in the vehicle repair knowledge graph based on a graph neural network to obtain the updated node embedded representation; and generating path embeddings of multi-hop inference paths starting from the initial node using a recursive approach based on the updated node embedded representations, and scoring and ranking the path embeddings using a trilinear attention mechanism to select the path most relevant to the input text.

[0046] The overall model architecture mainly includes the following key modules: input encoding module, graph neural network module, and path encoding and scoring module. The components and technical details of the input encoding module are shown in Table 1, the graph neural network module in Table 2, and the path encoding and scoring module in Table 3.

[0047] Table 1

[0048]

[0049] Table 2

[0050]

[0051] Table 3

[0052]

[0053] like Figure 1As shown, the overall model architecture mainly includes the following key modules: input encoding module, graph neural network module, and path encoding and scoring module. Specifically, all Processed (PCD) repair project concepts with assigned unique identifiers (PCDs) identified from the input vehicle repair record text are used to retrieve one-hop subgraphs from the constructed large vehicle repair knowledge graph. Each node in the graph represents a PCD. These one-hop subgraphs are encoded by a Stack Graph Isomorphism Network (SGIN), whose generated node embeddings capture neighborhood concept information and pre-trained concept embeddings. SGIN maximizes the ability of graph neural networks to capture meaningful representations. The generated node embeddings serve as the basis for path embeddings and are further processed by the path encoder. The path encoder module then assigns a relevance score to each path by checking the semantic and logical consistency of these one-hop paths with the input text and concept representations. The first N scores of these paths, aggregated with the neighborhood path scores of each node, guide the selection of the next hop node. If no suitable diagnostic node is found, path exploration is terminated by assigning a self-loop to the current node. In this invention, the expressive power of graph structures is utilized to enhance the retrieval process, and the path-concept relationship is encoded through an attention network, thereby ensuring the accuracy of path selection and contextual relevance.

[0054] As a specific implementation of this embodiment, the input encoding construction process includes: defining a deterministic vehicle maintenance knowledge graph G = (V, E) based on a custom vehicle maintenance concept unique identifier (PCD) and semantic relations, where V is the set of concept unique identifiers and E is the set of semantic relations. Given input text x, which contains a set of source PCDs... and its 1-hop relationship It can be done for each source node Construct a relation path P = {p1, p2, ..., pj} such that pj = {v1, e1, v2, ..., ej-1, vj}. Where J is the maximum reachable length of the source node vsrc and is nondeterministic. Relation e is encoded as a one-hot embedding representation. Then, all concept names of vi are concatenated with special tags (such as [SEP] for "delimiter"), such that l i = [name1[SEP]name2[SEP]...], and use the Bidirectional Encoder Representations from Transformers (BERT) model to represent l i Encode to obtain h as a conceptual representation iThis representation allows the PCD representation to serve as a contextualized representation of its corresponding concept name. BERT was chosen because its contrastive learning-based training can distinguish similar concepts and their synonyms, and it has shown state-of-the-art performance on entity linking tasks. Neighborhood information is iteratively aggregated through graph convolution and nonlinear functions, and a Graph Isomorphism Network (GIN) is used to process h. i A topological representation update is performed to model the interactions within the node set N(vi). Furthermore, a stacking mechanism is introduced to combine multiple GIN layers. The final node representation vi of the last layer (the Kth layer) is computed by stacking the GIN layers.

[0055] Furthermore, the process of obtaining the embedded representation of the input text includes: matching the repair item concepts and unique identifiers of vehicle repair concepts in the input vehicle repair record text to identify all source PCD sets with unique identifiers; using a bidirectional converter model to encode the concept name and its associated text for each PCD to generate an initial concept representation; weighting the initial concept representation based on word frequency statistical features to obtain a weighted concept representation; and iteratively updating the weighted concept representation through a multi-layer graph isomorphic network to aggregate neighborhood information and output the embedded representation of the input text.

[0056] However, certain types of PCDs are unlikely to generate useful paths for vehicle repair. Therefore, this invention designs a weighted scheme based on word frequency-inverse document frequency to assign higher weights to more relevant PCDs and semantic types.

[0057] W PCD =tf-idf(PCD) (1)

[0058] Then W PCD Multiply by the corresponding h i TF-IDF is used to assign weighted representations to the concept representation. TF-IDF is used to weight the initial representation of the vehicle maintenance concept (PCD).

[0059] As a specific implementation of this embodiment, the graph neural network update process includes: using a graph neural network (GNN) to update the nodes in the vehicle maintenance knowledge graph to better capture neighborhood information between nodes. Taking GNN as an example, its update formula is:

[0060]

[0061] Where h i (l) Let N(i) represent the embedding representation of node i at layer l, and let N(i) represent the set of neighboring nodes of node i. j,i(l) h represents the characteristics of the edge from node j to node i. i (l+1) The updated node embedding representation is represented by ∈, which is a learnable parameter used to control the degree to which the node's own information is preserved. By stacking multiple layers of GNNs, multi-hop graph structure information propagation can be achieved, enabling the node embedding representation to incorporate information from multiple levels of neighborhood, thus providing a more accurate node representation for subsequent path reasoning. j (l) This represents the embedding representation of node j at layer l.

[0062] As a specific implementation of this embodiment, the path encoding and scoring process includes: for each node representing h... i The corresponding path embedding is generated using the nodes and edges in its n-hop neighborhood, where t is the index of the node and its associated neighborhood and relation:

[0063]

[0064] Where FFN() represents a feedforward network, and n is the subgraph G src Number of hops in the path. Path embedding p i In the first hop, the node embedding is itself, and as the path extends to the next hop, it recursively aggregates with new nodes and edges, p i-1 h represents the path embedding of node i-1. i This represents the embedding representation of node i.

[0065] In the multi-head attention mechanism, the context relevance matrix H is first defined. i and concept relevance matrix Z i as follows:

[0066] H i =[h x ;p i h x -p i h x ⊙p i (4)

[0067] Z i =[h v ;p i h v -p i h v ⊙p i (5)

[0068] α i =MultiAttn(H i ⊙Z i (6)

[0069] S Multi =φ(Relu(σ(α) i ))) (7)

[0070] These relevance matrices are inspired by previous work on natural language reasoning, and represent logical relationships such as neutrality, contradiction, and implication through matrix concatenation, difference, and product, respectively.

[0071] To determine the relevance of each path to a patient's specific symptoms, this invention uses trilinear attention (TriAttn) to compute a score S for each path. This mechanism uses an embedded representation h of the input text. x and input PCD list h v Encoded by BERT to capture explicit and complex relationships in the input data. TriAttn automatically learns the complex relationships between the input text, the concept list, and the current path through the inner product of three matrices.

[0072]

[0073] S Tri =φ(Relu(σ(α) i (9)

[0074] Among them, h x h v and p i Having the same dimension D, h x h is an embedded representation of the input text. v To input the PCD list, It is an MLP layer, α i S is the original relevance score calculated for trilinear attention. Tri For the final path score after MLP normalization, a is the dimension index of the embedding representation of the input text, b is the dimension index of the input PCD list, and c is the path embedding p. i Dimension index, W abc Let σ be the trilinear attention weight tensor, and let σ denote the sigmoid activation function. x ) a h x scalar value in the a-th dimension, (h v ) b h v scalar value in the b-th dimension, (p i ) c Indicates path embedding p i The scalar value in the c-th dimension. Finally, aggregate the MultiAttn or TriAttn scores across all candidate nodes, and select the top N nodes (concepts) V based on the aggregated attention scores. N For use in the next iteration:

[0075]

[0076] β represents the aggregated value of the path attention scores. By comparing the attention scores of candidate paths, the path ranking system selects the top N nodes most relevant to each patient's symptoms, maximizing contextual relevance.

[0077] As a specific implementation method of this embodiment, the comparative learning process includes:

[0078] To further optimize the model's inference performance, a contrastive learning strategy is introduced. Positive and negative sample pairs are constructed, where positive samples represent vehicle repair project concept paths relevant to the input text, and negative samples represent irrelevant paths. A contrastive learning loss function is constructed by calculating the distance between the anchor embedding and the positive and negative sample embeddings, with the following formula:

[0079]

[0080] Among them, A i It is an anchor point embedding, defined as h x Θh v Σ represents the input text and concept representation. i This represents the summation over a set of indices i, typically representing distinct training samples or pairs. and Calculate anchor points and positive features separately. or negative features The cosine similarity between the anchor points is used. Positive features represent paths that correctly lead to the target concept, while negative features represent paths that, although starting from the source, ultimately lead to the wrong concept. This equation measures the loss when the similarity between the anchor point and positive features is not significantly greater than the similarity with negative features, taking into account the expected separation.

[0081] Furthermore, the steps for optimization based on the contrastive learning strategy include: constructing a set of positive sample paths and a set of negative sample paths; calculating the cosine similarity between node embeddings and positive and negative sample paths; and optimizing model parameters by using a contrastive loss function that maximizes the cosine similarity of positive samples and minimizes the cosine similarity of negative samples.

[0082] As a specific implementation method in this embodiment, the loss function consists of two parts: PCD prediction loss L pred And contrastive learning loss L CL :

[0083] L = L pred +L CL (12)

[0084] The PCD prediction loss uses binary cross-entropy loss to calculate the prediction node V. N Is it in the true label Y?

[0085]

[0086] Where M is the number of real label sets, y i This represents the true label of the i-th PCD. This represents the probability of the i-th PCD predicted by the model.

[0087] Furthermore, as a specific implementation of this embodiment, the following experiment was conducted.

[0088] Dataset: A real vehicle repair record dataset is used, containing a large amount of vehicle repair record text, for training and testing the model's inference performance. Model parameters: The learning rate is set to 1e-5, the training epochs are 10, the maximum number of hops for multi-hop inference is 2, the number of candidate paths selected per hop is 8, the number of heads in the graph neural network is 3, the path encoder type is Transformer, the path scorer type is Flat, the distance metric for contrastive learning is cosine distance, and the pruning threshold is 0.8, etc.

[0089] During training, stochastic gradient descent or the Adam optimizer is used to update the model parameters. Simultaneously, a learning rate scheduler is introduced to dynamically adjust the learning rate based on training progress, accelerating model convergence and avoiding local optima. After each training cycle, the model's performance is evaluated using a validation set, and the optimal model parameters are selected and saved based on metrics such as the loss function value and accuracy on the validation set.

[0090] To evaluate the effectiveness of the model in predicting and diagnosing PCDs, this invention selects the set of PCDs connected to at least one path as the ground truth labels for model training and evaluation. The path length between the target and input PCDs is at most 2 hops. Since the model predicts the first N PCDs, this invention measures Recall@N and Precision@N:

[0091]

[0092] In addition, the F1 score, which is the harmonic mean of recall and precision, will be used.

[0093] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a maintenance path reasoning method based on a graph neural network.

[0094] The present invention also provides a storage medium on which a computer program is stored, which, when executed by a processor, implements a maintenance path reasoning method based on a graph neural network.

[0095] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A maintenance path reasoning method based on graph neural networks, characterized in that, Includes the following steps: The input vehicle repair record text is encoded using a pre-trained large language model to obtain the embedded representation of the input text; The embedded representation of the input text combines the unique identifier of the vehicle maintenance concept and semantic relationships to construct a vehicle maintenance knowledge graph; The node embedding representation in the vehicle maintenance knowledge graph is updated based on the graph neural network to obtain the updated node embedding representation. Based on the updated node embedding representation, a recursive approach is used to generate path embeddings for multi-hop inference paths starting from the initial node. The path embeddings are then scored and ranked using a trilinear attention mechanism to select the path most relevant to the input text. The expression for calculating the updated node embedding representation is: ； In the formula, h i (l) Let N(i) represent the embedding representation of node i at layer l, and let N(i) represent the set of neighboring nodes of node i. j,i (l) The characteristics of the edge from node j to node i are represented. h is a learnable parameter. i (l+1) h represents the updated node embedding representation. j (l) Represents the embedding representation of node j at layer l, and ReLU represents the ReLU activation function; The computational expression for the path embedding of a multi-hop inference path originating from the starting node, generated recursively, is as follows: ； In the formula, FFN() represents a feedforward network, n is the number of hops in the subgraph Gsrc, pi represents the path embedding of node i, pi-1 represents the path embedding of node i-1, and h i Represents the embedding representation of node i; The expression for scoring the path embedding using the trilinear attention mechanism is: ； In the formula, hx is the embedded representation of the input text, and hv is the input PCD list. It is an MLP layer, α i STri is the raw relevance score calculated by trilinear attention, STri is the final path score after MLP normalization, a is the dimensional index of the embedding representation of the input text, b is the dimensional index of the input PCD list, and c is the path embedding. The dimension index, Wabc is the trilinear attention weight tensor. This represents the activation function sigmoid. This represents the scalar value of hx in the a-th dimension. This represents the scalar value of hv in the b-th dimension. Let represent the scalar value of path embedding pi in the c-th dimension, and ReLU represent the ReLU activation function.

2. The maintenance path reasoning method based on graph neural networks according to claim 1, characterized in that, The process of obtaining the embedded representation of the input text includes: Match the repair item concepts and unique identifiers of vehicle repair concepts in the input vehicle repair record text to identify all source PCD sets with unique identifiers; The concept name and its associated text for each PCD are encoded using a bidirectional converter model to generate an initial concept representation; The initial concept representation is weighted based on word frequency statistical features to obtain a weighted concept representation. The weighted concept representation is iteratively updated by using a multi-layer graph isomorphic network, which aggregates neighborhood information and outputs an embedded representation of the input text.

3. The maintenance path reasoning method based on graph neural networks according to claim 2, characterized in that, After generating the embedded representation of the input text, a weighting scheme based on term frequency-inverse document frequency is introduced to assign weights to high-frequency and semantically relevant PCDs.

4. The maintenance path reasoning method based on graph neural networks according to claim 1, characterized in that, The method introduces a contrastive learning strategy to optimize the maintenance path reasoning process; The optimization steps based on the contrastive learning strategy include: Construct a set of positive sample paths and a set of negative sample paths; Calculate the cosine similarity between the node embedding and the paths of positive and negative samples; The model parameters are optimized by using a contrastive loss function that maximizes the cosine similarity of positive samples and minimizes the cosine similarity of negative samples.

5. The maintenance path reasoning method based on graph neural networks according to claim 4, characterized in that, The total loss function of the maintenance path inference method includes: PCD prediction loss and contrastive learning loss.

6. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the maintenance path reasoning method based on graph neural networks as described in claim 1.

7. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the maintenance path reasoning method based on graph neural networks as described in claim 1.

Images