Equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning

Abstract

The invention discloses an equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning, which comprises the following steps of: firstly, extracting a multi-granularity sub-graph of a subject entity in an equipment operation and maintenance knowledge graph, and integrating multilevel information such as an equipment state, a fault mode and an environmental factor to form a relation path set; secondly, performing joint coding on a question text and a relation path through an attention enhanced word embedding module based on a graph convolutional network, forming a word embedding vector, calculating a similarity score between a relation in the path and the question, pruning a sub-graph, and reducing the calculation amount; and finally, fusing a plurality of particle size sub-graphs after pruning processing, inputting the fused particle size sub-graphs into a reasoning decision module, extracting feature information from graph structure data by the reasoning decision module through graph convolution operation, dynamically adjusting a GCN learning process in combination with LSTM, finally realizing question reasoning and answer generation, and ensuring the accuracy of an equipment operation and maintenance decision.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment operation and maintenance decision-making, and specifically relates to an equipment operation and maintenance decision-making method based on multi-granularity learning and reasoning of a large-scale knowledge graph. Background Art

[0002] As an important pillar of the global economy, the manufacturing industry plays a key role in promoting technological innovation and social development. Under the background of the deep integration of industrialization and informatization, the manufacturing industry faces an increasingly complex production environment and market demands, which requires enterprises to not only ensure production efficiency and quality but also pay attention to the reliability and operating conditions of equipment. Therefore, for the manufacturing industry, the level of an enterprise's equipment maintenance management ability directly determines the strength of its core competitiveness. At the production manufacturing site, the equipment operating state directly affects product quality, production efficiency, and safety. Through effective equipment operation and maintenance, enterprises can maximize the service life of equipment, reduce production downtime caused by equipment failures, and improve the stability and reliability of the production line. However, during the process of equipment operation and maintenance, even if a large number of sensors are deployed to apply data-driven models for process monitoring, fault diagnosis, and predictive maintenance, operation and maintenance experts still need to query from the accumulated text data and domain knowledge to make reasonable decisions. And these data are often discrete and discontinuous, and the causes of failures are diverse and not in a one-to-one correspondence. The complexity of equipment failures often stems from the interaction of multi-level factors - from direct component abnormalities (such as a sudden increase in bearing temperature), to cross-system parameter coupling (such as the dynamic balance of lubrication pressure and motor speed), and then to the long-term influence of macro factors such as the external environment (such as humidity changes accelerating component corrosion). These different levels of associated information are presented as topological structures of different granularities in the knowledge graph: fine-grained represents the direct causal relationship between equipment entities, medium-grained reflects the composite effects formed by cross-domain interactions, and coarse-grained captures the potential association patterns between the system and the environment. In this case, making decisions only through a single level or granularity may often ignore other more important failure causes, posing a huge challenge to operation and maintenance decision-making.

[0003] At present, the intelligent decision-making methods in the field of equipment operation and maintenance can be mainly divided into data-driven, knowledge-driven, etc. Although data-driven decision-making methods can identify equipment failures in advance from signals, realize automatic monitoring of equipment, and can assist maintenance personnel in discovering fault problems in a timely manner and performing predictive maintenance on equipment, the setting of thresholds or parameters for equipment status evaluation still cannot get rid of manual formulation, and this method can only perform data analysis, and the operation and maintenance decision-making plan still needs to rely on expert experience; knowledge-driven decision-making methods currently rely on expert experience or predefined rules to diagnose and make decisions on equipment status through logical reasoning. Although the decision-making efficiency is high and the flexibility is strong, and it is applicable to fixed-mode faults, in the face of complex faults, experience may not be sufficient to support effective decision-making, and the applicability to unknown faults is relatively poor.

[0004] When dealing with complex and changeable equipment operation and maintenance data, the above intelligent decision-making methods often show limitations. Therefore, how to effectively utilize and process these complex and changeable equipment operation and maintenance knowledge is crucial for realizing intelligent operation and maintenance decision-making of equipment. With the continuous development of artificial intelligence, due to its unique advantages in knowledge expression, organization, and semantic reasoning, the knowledge graph can integrate and utilize complex domain knowledge more efficiently than other technologies, and has become a key tool for realizing intelligent decision-making in manufacturing, medicine, and other fields. In equipment operation and maintenance, the decision-making method based on the knowledge graph also has significant advantages. It can organize and reason data such as equipment, fault modes, operating states, and maintenance measures in a structured form, and through the mining and explicitization of implicit knowledge, closely associate the originally isolated knowledge nodes.

[0005] At present, most decision-making methods based on the knowledge graph mainly rely on single-granularity information for reasoning, only focus on knowledge at a single level and granularity, and ignore important information hidden in other levels, resulting in one-sided decision-making results and being difficult to comprehensively reflect the actual situation. In complex equipment operation and maintenance scenarios, the causes of faults are often not caused by a single factor, but the result of the interaction of multiple factors. For example, equipment failures may be affected by multiple factors such as improper operation, environmental changes, and equipment aging at the same time. There are complex and dynamic mutual relationships among these factors. Single-granularity decision-making methods are difficult to capture this multi-level and multi-granularity information interaction, thus affecting the accurate identification of fault modes. In addition, the scale of current equipment operation and maintenance data is expanding rapidly, and the scale of the constructed knowledge graph is also growing larger and more complex in structure. When facing large-scale knowledge graphs, the relationships between nodes often show highly complex and hierarchical characteristics. Traditional reasoning models often struggle to handle such complex associations and are difficult to efficiently capture and utilize these subtle and key information for accurate decision-making and reasoning. How to achieve high accuracy and high response speed of decision-making in the environment of large-scale equipment operation and maintenance knowledge graphs has become an urgent problem to be solved. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a device operation and maintenance decision-making method based on multi-granularity learning and reasoning of a large-scale knowledge graph, which can realize the reasoning of problems and the generation of answers, and ensure the accuracy and effectiveness of device operation and maintenance decisions.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A device operation and maintenance decision-making method based on multi-granularity learning and reasoning of a large-scale knowledge graph. Given an input problem and a corresponding topic entity, a multi-granularity subgraph generation module extracts information among device failure phenomena, failure causes, and device operation and maintenance methods around the topic entity in the device operation and maintenance knowledge graph, and converts it into multi-granularity subgraph information with multiple granularity levels. Each multi-granularity subgraph is composed of multiple relationship paths;

[0009] The attention-enhanced word embedding module jointly encodes the problem text and the relationship paths through the multi-head attention mechanism to generate a context-aware distributed vector representation, providing an accurate semantic embedding vector for the multi-granularity subgraph generation module;

[0010] The multi-granularity subgraph generation module calculates the similarity score between the relationships in the path and the problem based on the semantic embedding vector, and prunes the paths with lower scores from the multi-granularity subgraph; the graph structure data obtained after fusing the pruned multi-granularity subgraphs is input into the long short-term temporal graph convolutional network inference module;

[0011] The long short-term temporal graph convolutional network inference module includes a Transformer attention module, a graph convolutional network module, and a long short-term memory network module. The long short-term temporal graph convolutional network inference module processes the graph structure data through the combination of the Transformer attention module and the graph convolutional network module, extracts the feature information in the graph, and forms a feature vector; the long short-term memory network module is used to dynamically adjust the learning process of the graph convolutional network module, realize the reasoning of problems and the generation of answers, and obtain device operation and maintenance decisions.

[0012] Further, the method for the multi-granularity subgraph generation module to generate a multi-granularity subgraph is as follows: Assume that the topic entity is e, the problem is q, and the partial relationships retrieved from the topic entity e at time t are r(t) = (r1, ···, r t ), and the relationship path formed by filling the intermediate entities along the path is p k = {(h i , r i , t i )}; where: h i and t i are the head and tail entities in the path; r iis a relationship; traverse the graph structure of the device operation and maintenance knowledge graph to obtain subgraphs of k granularities, where each subgraph of a granularity contains relationship paths of different depths.

[0013] Further, the method for the multi-granularity subgraph generation module to calculate the similarity score between the relationship in the path and the problem based on the semantic embedding vector is as follows: Let the path set be where: a single path is composed of n k triples, and the relationship set of the problem and each relationship path is r k ={r1, r2,..., r m}; generate the semantic embedding vector based on the attention-enhanced word embedding module, and then calculate the similarity score between the path and the problem through the similarity function:

[0014]

[0015] where: E q represents the embedding vector of the problem q; is the weighted average of the embedding vectors of the relationships r k in the relationship set r i , and:

[0016]

[0017] where: represents the embedding vector of the relationship r i .

[0018] Further, the method for the multi-granularity subgraph generation module to fuse the multi-granularity subgraphs is as follows: Extract the nodes and edges of each subgraph in the multi-granularity subgraph, and perform fusion through the union operation according to their hierarchical relationships to ensure that the information at all granularity levels is effectively aggregated and avoid duplication of entities and relationships.

[0019] Further, the attention-enhanced word embedding module uses the RoBERTa encoder to process the input problem text q = {q1, q2,..., q n} and the relationship set r k ={r1, r2,..., r m} to generate their corresponding word embedding vectors and where: E q is the embedding vector of the problem q; E R is the embedding vector of the relationship r i ;

[0020] The attention-enhanced word embedding module introduces the kernel function k(h v , hu ) to calculate the similarity between nodes:

[0021]

[0022] where: h v and h u are the embedding vectors of nodes v and u; σ is a hyperparameter that controls the kernel smoothness;

[0023] When performing graph convolution operations, the attention-enhanced word embedding module updates the representation of node v by introducing an attention mechanism in the graph convolutional network and performing weighted aggregation in combination with node features:

[0024]

[0025] where: α vu is the attention weight between node v and adjacent node u; W (k) represents the weight matrix; and are the embedding vectors of nodes v and u before update respectively; N(v) represents the set of neighbor nodes of node v;

[0026] The attention-enhanced word embedding module adds the representation of the input node to the updated node representation to form a residual connection. After being processed by the normalization layer, the representation of the node maintains a stable distribution in each layer; then, a non-linear transformation is performed through a feed-forward neural network to increase the non-linear expression ability; the output result of the feed-forward neural network is standardized again through layer normalization to ensure that the final node representation has good numerical stability, and a semantic embedding vector is obtained.

[0027] Furthermore, in the long short-term temporal graph convolutional network inference module, the graph convolutional network module adopts a spatio-temporal joint convolution kernel to combine the spatial topology of the graph structure and the temporal information of the nodes to capture the dynamic changes of the nodes in the spatio-temporal dimension and the long-range dependencies between nodes; the principle of the spatio-temporal joint convolution kernel is: for nodes v and adjacent node u of a given graph structure G, the spatio-temporal joint convolution kernel updates the representation of the node by calculating the features and temporal information between nodes:

[0028]

[0029] where: N(v) represents the set of neighbor nodes of node v; α vu is the weight calculated based on the similarity between nodes; W (k) represents the weight matrix; and are the embedding vectors of nodes v and u before update respectively; k(h v ,h u ,t v ,tu ) is a spatio-temporal joint convolution kernel; h v and h u are the embedding vectors of nodes v and u; t v and t u are the times of nodes v and u respectively; the spatio-temporal joint convolution kernel is expressed as:

[0030]

[0031] where: ‖h v ―h u ‖ 2 is the Euclidean distance of the feature vectors between node v and node u; σ is a hyperparameter that controls the influence of feature similarity; |t v ―t u | is the time difference between node v and node u; τ is a hyperparameter that controls the influence of time difference.

[0032] Furthermore, in the long short-term time series graph convolution network inference module, the Transformer attention module is used to dynamically adjust the information transmission between nodes, focusing on the nodes most relevant to the query to optimize the accuracy of information aggregation; the principle of the Transformer attention module is:

[0033]

[0034] where: Q, K, and V are the query vector, key vector, and value respectively; d k is the dimension of the key vector.

[0035] Furthermore, the long short-term memory network module includes the LSTMInstruction module and the LSTM network; the LSTMInstruction module is used to encode the query text input, generate the instruction vector of the query, and convert the query text into word embeddings; the LSTM network processes the input word embeddings through a recursive formula to generate a time series hidden state, and this hidden state is used as the instruction vector to guide the inference process of the graph convolution operation.

[0036] The beneficial effects of the present invention are as follows:

[0037] The device operation and maintenance decision-making method based on multi-granularity learning and reasoning of large-scale knowledge graphs first extracts multi-granularity subgraphs of topic entities in the device operation and maintenance knowledge graph, integrates multi-level information such as device status, failure modes, and environmental factors to form a set of relationship paths. Secondly, through an attention-enhanced word embedding module based on the graph convolutional network, the problem text and relationship paths are jointly encoded to form word embedding vectors, and the similarity scores between the relationships in the paths and the problem are calculated to prune the subgraphs and reduce the computational amount. Finally, multiple granularity subgraphs after pruning are fused and input into the inference decision-making module. This module extracts feature information from the graph-structured data through graph convolutional operations and dynamically adjusts the learning process of GCN in combination with LSTM, ultimately realizing the inference of problems and the generation of answers to ensure the accuracy of device operation and maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the following drawings are provided for illustration:

[0039] Figure 1 It is the overall architecture diagram of the device operation and maintenance decision-making model;

[0040] Figure 2 It is the schematic diagram of the multi-granularity subgraph generation and fusion module;

[0041] Figure 3 It is the schematic diagram of the attention-enhanced word embedding module;

[0042] Figure 4 It is the schematic diagram of the long short-term time series graph convolutional network inference module;

[0043] Figure 5 It is the partial knowledge graph of the device operation and maintenance in a certain automobile workshop;

[0044] Figure 6 It is the training and test loss of the MulKGR model;

[0045] Figure 7 It is the trend graph of F1 and Hit@1;

[0046] Figure 8 It is the index trend graph. DETAILED IMPLEMENTATION MANNER

[0047] The present invention will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited are not intended to limit the present invention.

[0048] 1. Model Architecture

[0049] In this embodiment, the device operation and maintenance decision-making method based on multi-granularity learning inference of a large-scale knowledge graph is implemented using a device operation and maintenance decision-making model (MulKGR). Specifically, as Figure 1 shown, the device operation and maintenance decision-making model (MulKGR) of this embodiment includes a multi-granularity subgraph generation and fusion module (MGSG), an attention-enhanced word embedding module (ATGCN), and a long short-term temporal graph convolutional network inference module (LSTGCN). Specifically, the operating mechanism of MulKGR can be described as follows: Given an input question q and its corresponding topic entity e, first, MGSG extracts information among device failure phenomena, failure causes, and device operation and maintenance methods around the topic entity in the device operation and maintenance knowledge graph and converts it into subgraph information at multiple granularity levels. Each subgraph consists of multiple relation paths (h, r, t). Then, through the pre-trained ATGCN model, the question text and relation paths are jointly encoded through the multi-head attention mechanism to generate a context-aware distributed vector representation. The similarity scores between the relations in the path and the question are calculated. By comparing these scores, MulKGR can effectively filter out the paths with lower scores and prune them from the multi-granularity subgraphs. Finally, the multiple granularity subgraphs after pruning are fused and input into the inference decision module, which processes the graph-structured data through the combination of Transformer attention and graph convolutional network (GCN), extracts the feature information in the graph, and then forms a feature vector. Combining with the long short-term memory network (LSTM), this module can dynamically adjust the learning process of GCN and finally realize the inference of the question and the generation of the answer, ensuring the accuracy and effectiveness of the device operation and maintenance decision-making.

[0050] That is, in this embodiment, the device operation and maintenance decision-making method based on multi-granularity learning and reasoning of a large-scale knowledge graph is as follows: Given an input problem and the corresponding topic entity, the multi-granularity subgraph generation module extracts the information among the device failure phenomena, failure causes, and device operation and maintenance methods around the topic entity from the device operation and maintenance knowledge graph and converts it into multi-granularity subgraph information with multiple granularity levels. Each multi-granularity subgraph consists of multiple relationship paths. The attention-enhanced word embedding module jointly encodes the problem text and the relationship paths through the multi-head attention mechanism to generate a context-aware distributed vector representation, providing an accurate semantic embedding vector for the multi-granularity subgraph generation module. The multi-granularity subgraph generation module calculates the similarity score between the relationships in the path and the problem based on the semantic embedding vector, and prunes the paths with lower scores from the multi-granularity subgraph; the graph structure data obtained after fusing the pruned multi-granularity subgraphs is input into the long short-term temporal graph convolutional network inference module. The long short-term temporal graph convolutional network inference module includes a Transformer attention module, a graph convolutional network module, and a long short-term memory network module. The long short-term temporal graph convolutional network inference module processes the graph structure data through the combination of the Transformer attention module and the graph convolutional network module to extract the feature information in the graph and form a feature vector; the long short-term memory network module is used to dynamically adjust the learning process of the graph convolutional network module to realize the inference of the problem and the generation of the answer, and obtain the device operation and maintenance decision.

[0051] 2. Multi-Granularity SubGraph Generation AndFusion Module (MGSG)

[0052] The main objective of the multi-granularity subgraph generation and fusion module is to obtain the knowledge and information related to the topic entity e in the device operation and maintenance knowledge graph through the generation and fusion of multi-granularity subgraphs, generate multi-granularity subgraphs, and provide a refined subgraph set for the subsequent device operation and maintenance inference and decision-making module. As Figure 2 shown, the multi-granularity subgraph generation and fusion module MGSG module includes three key steps: subgraph search, subgraph pruning and scoring, and subgraph fusion.

[0053] Given the input problem and the topic entity, multiple granularity subgraphs are constructed through multi-hop relationships in the graph structure of the device operation and maintenance knowledge graph. Its purpose is to start from the topic entity and find multiple relationship paths starting from the topic entity in the graph. Here, the path is defined as a series of relationships (r1, ···, r |p| ), where: r iRepresents the i-th relationship in the path. Each path represents a potential device failure phenomenon, cause of failure, or operation and maintenance method. Specifically, the method for the multi-granularity subgraph generation module to generate a multi-granularity subgraph is as follows: Assume the subject entity is e, the question is q, and the partial relationships retrieved from the subject entity e at time t are r(t) = (r1, ···, r t ), and the relationship path formed by filling the intermediate entities along the path is p k = {(h i , r i , t i )}; where: h i and t i are the head and tail entities in the path; r i is the relationship; traverse the graph structure of the device operation and maintenance knowledge graph to obtain subgraphs of k granularities, where each subgraph of a granularity contains relationship paths of different depths.

[0054] After generating multiple granularity subgraphs, it is necessary to score the relationship paths and prune the irrelevant or low-scoring relationship paths to reduce the size of the multi-granularity subgraph. For this purpose, a similarity measurement method based on relationship embedding is adopted to evaluate the similarity between the relationships in the path and the problem semantics. Specifically, the method for the multi-granularity subgraph generation module to calculate the similarity score between the relationships in the path and the problem based on the semantic embedding vector is as follows: Let the path set be where: A single path is composed of n k triples, and the relationship set of the problem and each relationship path is r k = {r1, r2,..., r m}; Generate a semantic embedding vector based on the attention-enhanced word embedding module, and then calculate the similarity score between the path and the problem through a similarity function:

[0055]

[0056] where: E q represents the embedding vector of the question q; is the weighted average of the embedding vectors of the relationships r k in the relationship set r i , and:

[0057]

[0058] where: represents the embedding vector of the relationship r i .

[0059] According to the above similarity scores, select the top K paths with the highest scores, and prune the paths with lower scores from the multi-granularity subgraph, thereby reducing the computational complexity and improving the inference efficiency.

[0060] After pruning is completed, the multi-granularity subgraphs will be fused into a set of refined subgraphs with rich semantic information, and the fused subgraph set will be passed as input to the LSTGCN module. In the multi-granularity subgraph fusion step, entities and relationships of each subgraph are first extracted from subgraphs at multiple granularities. Each subgraph represents different depths of interaction between entities and relationships, containing information at different granularities and potential inference paths. Therefore, to ensure that the global information of the graph is not lost, all entities and relationships from subgraphs at different granularities will be pooled together to form a high-dimensional composite graph. Specifically, the method for the multi-granularity subgraph generation module to fuse multi-granularity subgraphs is as follows: extract the nodes (entities) and edges (relationships) of each subgraph in the multi-granularity subgraphs, and fuse them through the union operation according to their hierarchical relationships, ensuring that information at all granularity levels is effectively aggregated and avoiding duplication of entities and relationships.

[0061] 3. Attention-enhanced Text-GCN (ATGCN)

[0062] In equipment operation and maintenance tasks, how to effectively understand and reason about complex relationships in text is the key to improving the accuracy of decision-making. To address this challenge, this embodiment proposes an attention-enhanced text-GCN (ATGCN) that combines a graph convolutional network (GCN) with a multi-head attention mechanism using continuous kernels. ATGCN uses sentence embeddings obtained from the RoBERTa encoder as input to better capture relevant syntactic and semantic relationships in equipment operation and maintenance problem texts. As Figure 3 shown, the core framework of the ATGCN model includes two main modules: a graph convolutional layer based on continuous convolutional kernels and a multi-head attention mechanism.

[0063] The attention-enhanced text-GCN processes the input problem text and relationships through the RoBERTa model to capture deep semantic relationships in the text and generate context-aware embeddings for each word. Specifically, the attention-enhanced text-GCN uses the RoBERTa encoder to process the input problem text q = {q1, q2, …, q n} and the relationship set r k = {r1, r2,..., r m} of each relationship path to generate their corresponding word embedding vectors and where: E q is the embedding vector of the problem q; E R is the embedding vector of the relationship r i .

[0064] After the word embeddings are generated, the word vectors are converted into a graph structure, where each word has a corresponding node. Then, ATGCN can capture the semantic dependencies between words more comprehensively through graph convolution operations. Based on the traditional Graph Convolutional Network (GCN), ATGCN introduces a continuous kernel and an attention mechanism, enabling the model to handle complex semantic relationships more flexibly. Traditional GCN usually relies on the adjacency matrix and degree matrix for information aggregation. Although this method can capture the local structural information of nodes, it performs relatively limited when dealing with high-order dependencies or long-range node relationships between nodes. To solve this problem, the ATGCN in this embodiment introduces a continuous kernel function to measure the similarity between nodes and combines it with the attention mechanism to further enhance the model's ability. Specifically, the similarity between nodes no longer depends solely on simple adjacency relationships, but is calculated through a kernel function k(h v ,h u ), where h v and h u are the embedding vectors of nodes v and u. The continuous kernel function adopts the form of a Gaussian kernel. The attention-enhanced word embedding module in this embodiment calculates the similarity between nodes by introducing the kernel function k(h v ,h u ) in the graph convolutional network:

[0065]

[0066] where: h v and h u are the embedding vectors of nodes v and u; σ is a hyperparameter that controls the smoothness of the kernel.

[0067] This kernel function calculates the similarity between nodes by measuring the Euclidean distance between word nodes in the problem description of equipment failures, thereby capturing the local semantic structure of nodes while also being able to identify the indirect relationships between nodes.

[0068] When performing graph convolution operations, the attention-enhanced word embedding module introduces an attention mechanism in the graph convolutional network, combines node features for weighted aggregation, and updates the representation of node v:

[0069]

[0070] where: α vu is the attention weight between node v and its adjacent node u; W (k) represents the weight matrix; and are the embedding vectors of nodes v and u before update respectively; N(v) represents the set of neighbor nodes of node v.

[0071] By calculating the attention weights between nodes, ATGCN can dynamically assign different influences to different neighbor nodes. Specifically, ATGCN obtains the attention weights between nodes by calculating the dot product between Query, Key, and Value (V). Then, the outputs of each attention head are concatenated and passed through a linear transformation to generate the final node embedding. Finally, the representation of the input node is added to the transformed node representation to form a residual connection. After passing through a normalization layer, the distribution of the node representation is kept stable in each layer. Then, a non-linear transformation is performed through a Feed-Forward Network (FFN). The feed-forward network consists of two fully connected layers, with a ReLU activation function in the middle to increase the non-linear expression ability. After the feed-forward neural network processing, the output result is normalized again through layer normalization to ensure good numerical stability of the final node representation. Through the above series of steps, ATGCN can better understand the complex text semantics of device operation and maintenance and provide accurate semantic embedding vectors for the MGSG module.

[0072] 4. Long Short-Term Graph Convolutional Network Inference Module (LSTGCN)

[0073] The Long Short-Term Graph Convolutional Network Inference Module (LSTGCN) combines the Graph Neural Network (GCN), Long Short-Term Memory Network (LSTM), and the attention mechanism of Transformer to improve the accuracy and effectiveness of device operation and maintenance decisions. The LSTGNN module extracts feature information from graph-structured data through graph convolution operations and combines LSTM to dynamically adjust the learning process of GCN to process query inputs and generate inference instructions, thus achieving precise aggregation of feature information. The attention mechanism of Transformer further optimizes the relationship modeling between nodes by calculating the similarity between queries and keys and dynamically weighting the transmission of node information. Finally, the LSTGCN module combines graph convolution, temporal information learning, and the Transformer attention mechanism to provide a powerful inference framework that can effectively make device operation and maintenance decisions.

[0074] Such as Figure 4As shown in the figure, LSTGCN first processes the input query text and graph structure data separately. For the graph structure data, it is processed through the Transformer attention mechanism and GCN. GCN uses an innovative spatio-temporal joint convolution kernel to replace the traditional static neighborhood aggregation operation. The spatio-temporal joint convolution kernel combines the spatial topology of the graph and the temporal information of the nodes, enabling it to capture the dynamic changes of nodes in the spatio-temporal dimension and the long-range dependencies between nodes. This allows LSTGCN to not only handle the local relationships between nodes but also capture the global information that changes over time in the graph, enhancing the model's temporal learning ability. The key to the spatio-temporal joint convolution kernel is that it performs joint convolution processing by considering the features and time information of the nodes.

[0075] Specifically, the principle of the spatio-temporal joint convolution kernel is as follows: For a node v and its adjacent node u in a given graph structure G, the spatio-temporal joint convolution kernel updates the representation of the node by calculating the features and temporal information between the nodes:

[0076]

[0077] where: N(v) represents the set of neighbor nodes of node v; α vu is the weight calculated based on the similarity between nodes; W (k) represents the weight matrix; and are the embedding vectors of nodes v and u before update respectively; k(h v , h u , t v , t u ) is the spatio-temporal joint convolution kernel; h v and h u are the embedding vectors of nodes v and u; t v and t u are the times of nodes v and u respectively.

[0078] The spatio-temporal joint convolution kernel is expressed as:

[0079]

[0080] where: ‖h v ― h u ‖ 2 is the Euclidean distance of the feature vectors between node v and node u; σ is the hyperparameter that controls the influence of feature similarity; |t v ― t u | is the difference in time between node v and node u; τ is the hyperparameter that controls the influence of time difference.

[0081] After the graph convolution processes the features of the nodes, the attention mechanism of the Transformer further optimizes the relationship modeling between the nodes. The core of the Transformer's attention mechanism is self-attention. By calculating the similarity between the query vector (Q) and the key vector (K), it determines how to weight the value (V) of the node. The relationship between them is measured by the dot product, and after the dot product, the softmax function is used for normalization so that the sum of each attention weight is 1. That is, in the long short-term temporal graph convolution network inference module of this embodiment, the Transformer attention module is used to dynamically adjust the information transmission between nodes, focusing on the nodes most relevant to the query to optimize the accuracy of information aggregation. The principle of the Transformer attention module is:

[0082]

[0083] where: Q, K, and V are the query vector, key vector, and value respectively; d k is the dimension of the key vector.

[0084] Through this self-attention mechanism, LSTGCN can dynamically adjust the information transmission between nodes, focus on the nodes most relevant to the query, and thus optimize the accuracy of information aggregation.

[0085] The long short-term memory network module includes the LSTMInstruction module and the LSTM network. The LSTMInstruction module is used to encode the query text input and generate the instruction vector of the query. The LSTMInstruction module converts the query text into word embeddings and performs temporal learning through the LSTM network to obtain the hidden layer representation h query of the query. Specifically, the LSTM network processes the input word embeddings through a recursive formula to generate a temporal hidden state, and this hidden state is used as the instruction vector to guide the inference process of the graph convolution operation. In this embodiment, the principle of the LSTM network is:

[0086] i t = σ(W i x t + U i h t―1 + b i )

[0087] f t = σ(W f x t + U f h t―1 + b f )

[0088] o t = σ(W ox t +U o h t―1 +b o )

[0089]

[0090]

[0091] Among them: x t is the query word embedding; h t is the hidden state of the LSTM at time t; h t―1 is the hidden state of the LSTM at time t-1; i t , f t , o t respectively represent the activation values of the input gate, forget gate, and output gate, controlling the flow of information; c t is the cell state at time t; c t―1 is the cell state at time t-1; σ is the sigmoid activation function; represents element-wise multiplication; W i , W f , W o and W c are the weight matrices of the forget gate, input gate, output gate, and candidate state respectively, used for linear transformation of the input x t ; U i , U f , U o and U c are the weight matrices of the forget gate, input gate, output gate, and candidate state respectively, used for the hidden state h t-1 at the previous moment; b i , b f , b o and b c are the bias terms of the forget gate, input gate, output gate, and candidate state respectively.

[0092] The instruction vector generated by the LSTM. The LSTGCN can dynamically adjust the inference process of the graph convolutional network according to the semantic information of the query, ensuring that the inference process can focus on the nodes and graph structures related to the query.

[0093] 5. Experiments

[0094] This embodiment details the data source, experimental design, and evaluation method to verify the effectiveness of the MulKGR model in the intelligent decision-making task of equipment operation and maintenance from different perspectives.

[0095] 5.1. Dataset

[0096] The dataset used in this embodiment is sourced from the equipment maintenance records of a new energy vehicle enterprise's production workshop. This dataset includes equipment names, equipment numbers, model specifications, equipment locations, operating conditions, fault descriptions, fault causes, and relevant operation and maintenance measures, as shown in Table 1.

[0097] Table 1 Partial Records of Equipment Maintenance

[0098]

[0099] By organizing and extracting the above valid data, an equipment operation and maintenance knowledge graph for the automotive workshop is constructed, as Figure 5 shown. This knowledge graph stores a large amount of equipment operation and maintenance knowledge for the production workshop in the form of triples (h, r, t), including 102,910 entities, 200 relationships, and 105,600 triples are formed based on this. To further utilize this knowledge graph for intelligent decision-making and question-answering tasks, this embodiment generates questions, the topic entities in the questions, and the corresponding answer entities based on the triples. Specifically, using the information of entities and relationships in the triples, a large number of question pairs are constructed, and the questions, topic entities, and answer entities are saved in jsonl format. The format of the dataset is {"question": "What is the reason for the weld offset during the welding process of the welding robot", "entities": ["welding robot"], "answers": ["Laser tracking system calibration deviation"]}, which is convenient for subsequent model training and inference. The dataset is divided into a training set, a validation set, and a test set according to a ratio of 3:1:1.

[0100] 5.2. Evaluation Metrics and Parameter Settings

[0101] 5.2.1. Evaluation Metrics

[0102] The evaluation metric adopted in this embodiment is Hits@K. Hits@K refers to the proportion of questions in which at least one answer is included in the candidate answer list generated by the model. The formula is as follows:

[0103]

[0104] where: γ is the set of triples, |γ| is the number of triples in the set, rank iIt refers to the candidate ranking of the i-th triple. In addition, Π() represents the indicator function, which outputs 1 when the input is True and 0 when the input is False. Generally speaking, the larger the value of Hits@K, the more effectively the decision model can capture the correct answer among fewer returned results. Due to some problems in the dataset where there are multiple correct answers, relying solely on ranking metrics often cannot comprehensively reflect the quality of the model's predicted answer set. The F1 score combines Precision and Recall, can measure both the accuracy and integrity of the model's predicted answers simultaneously, and provides a more balanced evaluation in the multi-answer scenario. Therefore, it is also necessary to evaluate the F1 score of the model. The specific formula is as follows:

[0105]

[0106] Where: TP represents the number of samples predicted as the positive class and actually being the positive class, FP represents the number of samples predicted as the positive class but actually being the negative class, and FN represents the number of samples misjudged as correct by the model.

[0107] 5.2.2 Parameter Settings

[0108] The decision model proposed in this embodiment is written and implemented based on Python 3.10 and the Pytorch framework. During the training process, we use the KL divergence as the loss function and the Adam optimizer to optimize the model, which improves the training stability of the model and can save hardware resources. The entity embedding dimension of the LSTGCN decision module is set to 128, the word embedding dimension is set to 300, the number of attention heads is set to 4, and the ε-greedy strategy is adopted, where ε is set to 0.95, the kernel smoothing hyperparameter σ is set to 0.5, and the time τ parameter is set to 50. The specific core parameters of this experiment are shown in Table 2.

[0109] Table 2 Parameter Settings of the MulKGR Model

[0110]

[0111] At the same time, in this embodiment, the maximum path width K in the subgraph is set to 5, that is, the relationship paths with a score threshold in the top 5 (including 5) are considered. In order to quickly obtain the fault causes or operation and maintenance measures closely related to the topic entity, capture complex association information, and mine potential multi-level causal chains, in this embodiment, the maximum number of hops N of the path is set to 3 (specifically explained in 4.3.3), and subgraphs with three granularities of 1-hop subgraph, 3-hop subgraph, and 5-hop subgraph are generated to comprehensively aggregate and utilize device operation and maintenance knowledge from multiple granularity levels.

[0112] 5.3 Experimental Analysis

[0113] In this embodiment, the MulKGR model is tested from two perspectives: comparative experiments and ablation experiments, and F1 and Hits@1 are selected as evaluation metrics.

[0114] 5.3.1 Comparative Experiments

[0115] To verify the decision-making level of the MulKGR model in the field of equipment operation and maintenance, this embodiment selects an embedding-based knowledge graph question-answering decision-making model for comparison. Among them, EmbedKGQA is an embedding model based on link prediction, which directly infers multi-hop question answers using the semantic relevance in the embedding space and does not rely on the restricted local subgraph traversal in traditional methods. GraIL, NSM, and NSM+SR all belong to subgraph-based embedding models, which first extract local subgraphs related to the question, and then perform embedding representation learning and multi-step reasoning on the topological structure of the subgraph. By comparing with the embedding-based question-answering decision-making model on the equipment operation and maintenance dataset, the effectiveness of the proposed model is verified. The specific results are shown in Table 3.

[0116] Table 3 Results of Comparative Experiments

[0117]

[0118] The comparison results show that all indicators of this model in the equipment operation and maintenance intelligent decision-making task are better than other question-answering decision-making models. In the test evaluation, the F1 score of the model on the dataset reached 65.88%, which is 14.17%, 12%, 10.59%, and 7.34% higher than the baseline model. The Hits@1 index reached 62.50%, which is 13.9%, 13.4%, 10.8%, and 4.6% higher than the baseline model. These indicators fully prove that the MulKGR model can effectively extract and fuse information from subgraphs of different granularities, so as to achieve high-precision reasoning in the complex interaction of multi-level fault causes and equipment states, providing strong decision-making support for actual production. In addition, Figure 6 shows the change trend of the loss value during the training process of the MulKGR model. As the training progresses, the loss value decreases continuously with the number of training times, and the average loss value remains at a low level with small fluctuations, indicating that the performance of this model on the test set is very stable. Generally speaking, by constructing multi-granularity subgraphs and fusing multi-level information, the MulKGR model not only realizes the deep mining of equipment states and fault modes, but also achieves significantly better performance than existing models in the experiment, providing strong technical support for complex equipment operation and maintenance decisions.

[0119] 5.3.2 Ablation Experiments

[0120] To verify whether each module plays a role in the MulKGR model, we conducted ablation experiments to verify the effectiveness of three modules: the attention-enhanced word embedding module based on graph convolutional network (ATGCN), the multi-granularity subgraph generation and fusion module (MGSG), and the long short-term temporal graph convolutional network inference module (LSTGCN). We designed three groups of experiments, namely: (1) Delete the attention-enhanced word embedding module based on graph convolutional network (ATGCN) in the MulKGR model. (2) Delete the multi-granularity subgraph generation and fusion module (MGSG) in the MulKGR model. (3) Delete the long short-term temporal graph convolutional network inference module (LSTGCN) in the MulKGR model. The specific results of the experiments are shown in Table 4.

[0121] Table 4 Ablation Experiments of Each Module of the Model

[0122]

[0123] The above results show that in the MulKGR model, removing the ATGCN module led to a 3.02% and 6% decrease in the F1 and Hits@1 evaluation metrics of the model respectively, indicating that the ATGCN module can capture the semantic features of device operation and maintenance problem texts and achieve the understanding of complex texts for device operation and maintenance; removing the MGSG module led to a 12.84% and 11.2% decrease in the F1 and Hits@1 evaluation metrics of the model respectively, indicating that the MGSG module helps the model effectively capture multi-level relationships in large-scale knowledge graphs through operations such as extraction, fusion, and dynamic pruning of multi-granularity subgraphs, thereby improving the accuracy of device operation and maintenance intelligent decision-making; removing the LSTGCN module led to a 3.7% and 6.2% decrease in the F1 and Hits@1 evaluation metrics of the model respectively, indicating that the LSTGCN module can well improve the model's ability to extract graph features by combining the Transformer attention mechanism and graph convolutional network, and by using LSTM to generate dynamic instructions, thereby effectively improving the accuracy of device operation and maintenance intelligent decision-making. In summary, the three modules proposed in this embodiment all play a significant role in the MulKGR model, especially the multi-granularity subgraph generation and fusion module contributes the most to device operation and maintenance intelligent decision-making.

[0124] 5.3.3. Multi-Granularity Subgraph Configuration Experiment

[0125] This experiment aims to explore the role of multi-granularity subgraph configuration in the MulKGR model and evaluate it mainly from two aspects: one is to verify whether the effect of using three subgraphs is better than that of using two or four subgraphs; the other is to evaluate the performance differences between using 1-hop, 3-hop, 5-hop subgraphs and other combinations (such as 1-hop, 5-hop, 7-hop), and explore the impact of different hop levels on fault information capture and model decision-making and reasoning effects.

[0126] (1) Sub - graph Quantity Selection Experiment

[0127] To ensure the control - variable method is followed for the sub - graph quantity configuration in this experiment, a candidate hop - count set, namely {1, 3, 5, 7, 9}, is predefined in this embodiment. Under different configurations, according to the selected sub - graph quantity k, the smallest k hop - counts are taken from this set as the hop - counts for generating sub - graphs. On this basis, we set up four groups of experiments: (1) sub - graph quantity k = 2; (2) sub - graph quantity k = 3 (benchmark group); (3) sub - graph quantity k = 4. Using F1 and Hit@1 as evaluation metrics, the specific results are shown in Table 5.

[0128] Table 5 Sub - graph Quantity Selection Results

[0129]

[0130] It can be seen from the experimental results that as the sub - graph quantity increases, the performance of the model on F1 and Hit@1 first increases to the maximum and then decreases, as Figure 7 shown. When the sub - graph quantity k = 2, the model's capture of local relationships is relatively limited, resulting in sub - optimal decision - making and reasoning effects; when the sub - graph quantity k = 3, all indicators reach the optimal level, with good information coverage and decision - making capabilities; when the sub - graph quantity increases to k = 4 or k = 5, although the sub - graph quantity coverage range is further expanded, the redundant noise also increases, thus reducing the model performance.

[0131] In summary, using three sub - graph configurations can achieve a better balance in capturing multi - level information of equipment failures, fully mining direct, indirect, and cross - level correlation information without introducing too much noise.

[0132] (2) Hop - count Combination Selection Experiment

[0133] This experiment aims to explore the impact of different hop - count combinations on the performance of the MulKGR model. To ensure variable control, the sub - graph quantity is set to a fixed value of 3 in this experiment, keeping other factors the same and only varying the hop - count combinations. On this basis, we set up 4 experiments: (1) using an arithmetic - progression - interval hop - count combination, namely [1, 2, 3]; (2) using an exponential - progression - interval hop - count combination (benchmark group), namely [1, 3, 5]; (3) using a larger - span hop - count combination, namely [1, 5, 7]; (4) only using higher hop - counts without including lower hop - counts, namely [3, 5, 7]. The specific experimental results are shown in Table 6.

[0134] Table 6 Hop - count Combination Selection Results

[0135]

[0136] The experimental results show that the configuration based on [1,3,5] hops performs best in terms of F1 and Hit@1 metrics; when using the [1,2,3] combination, the model is insufficient in capturing shallow-level local information, resulting in low overall performance; when using the [1,5,7] combination, although the coverage range is larger, more noise is introduced, thus reducing the inference efficiency; while using the [3,5,7] combination cannot fully mine and utilize the key information directly related to equipment operation and maintenance due to the lack of low-hop information, and the evaluation indicators are not as good as those of the benchmark group.

[0137] In summary, a reasonable hop configuration helps to extract key local structure information, while an excessive number of hops (such as 7 hops) may lead to an increase in noise and a decrease in inference efficiency, thereby affecting the decision-making performance of the model.

[0138] 5.3.4. Large-scale Knowledge Graph Decision-making Performance Experiment

[0139] This experiment aims to verify the advantages of the MulKGR model proposed in this embodiment in large-scale knowledge graph decision-making and compare it with another model NSM+SR that only generates a single subgraph. For this purpose, we designed three experiments: (1) a knowledge graph containing 1000 triples; (2) a knowledge graph containing 10,000 triples; (3) a knowledge graph containing 100,000 triples. In each scenario, the F1 score and Hit@1 are used as evaluation indicators, and the number of subgraphs generated by the two models is controlled to be the same. The comparison of the decision-making performance of MulKGR and NSM+SR is shown in the following Table 7.

[0140] Table 7 Experimental Results of Decision-making Performance

[0141]

[0142] As Figure 8 shown, in the three experiments where the knowledge graph scale contains 1000 triples, 10,000 triples, and 100,000 triples, the MulKGR model is superior to the comparison model NSM+SR in terms of F1, Hit@1, and decision-making inference time, demonstrating excellent capabilities. This result fully proves the advantage of the MulKGR model in achieving fast and accurate decision-making in a large-scale knowledge graph environment.

[0143] The above-described embodiments are only preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is subject to the claims.

Claims

1. A device operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning, characterized by: Given an input question and a corresponding subject entity, the multi-granularity subgraph generation module is used to extract information about equipment failure phenomena, failure causes, and equipment operation and maintenance methods around the subject entity in the equipment operation and maintenance knowledge graph, and convert it into multi-granularity subgraph information with multiple granularity levels. Each multi-granularity subgraph consists of multiple relationship paths. The attention-enhanced word embedding module is used to jointly encode the question text and the relationship path through a multi-head attention mechanism to generate a context-aware distributed vector representation, providing accurate semantic embedding vectors for the multi-granularity subgraph generation module. The multi-granularity subgraph generation module calculates the similarity score between the relationship in the path and the question based on the semantic embedding vector, and cuts the path with a lower score from the multi-granularity subgraph based on the similarity score; the graph structure data obtained by fusing the cut multi-granularity subgraph is input into the long-short term temporal graph convolution network reasoning module; The long-short-term temporal graph convolutional network reasoning module includes a Transformer attention module, a graph convolutional network module and a long-short-term memory network module. The long-short-term temporal graph convolutional network reasoning module processes graph structure data by combining the Transformer attention module with the graph convolutional network module, extracts feature information in the graph, and forms a feature vector; the long-short-term memory network module is used to dynamically adjust the learning process of the graph convolutional network module, realize the reasoning of problems and the generation of answers, and obtain equipment operation and maintenance decisions.

2. The equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning according to claim 1 is characterized by: The method for generating a multi-granularity subgraph by the multi-granularity subgraph generation module is: assuming that the subject entity is e, the question is q, and the partial relationship retrieved from the subject entity e at time t is r(t)=(r1, ···, r t ), the relationship path formed by filling in the intermediate entities along the path is p k ={(h i ,r i ,t i )}; where: h i and t i are the head and tail entities in the path; r i is a relationship; traverse the graph structure of the equipment operation and maintenance knowledge graph to obtain subgraphs of k granularities, where each subgraph of k granularity contains relationship paths of different depths.

3. The equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning according to claim 1 is characterized by: The method for the multi-granularity subgraph generation module to calculate the similarity score between the relationship in the path and the question based on the semantic embedding vector is: suppose the path set is Among them: Single path By k The relationship between the question and each relationship path is r k ={r1,r2,...,r m }; Generate semantic embedding vectors based on the attention-enhanced word embedding module, and then calculate the similarity score between the path and the question through the similarity function: Where: E q Embedding vector representing question q; is the relation set r k Chinese Relations i The weighted average of the embedding vectors, and: in: Represents the relationship i The embedding vector of .

4. The equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning according to claim 1 is characterized by: The method for fusing the multi-granularity subgraphs by the multi-granularity subgraph generation module is: extracting the nodes and edges of each subgraph in the multi-level subgraph, and fusing them through a union operation according to their hierarchical relationship, ensuring that information at all granularity levels is effectively aggregated and avoiding duplication of entities and relationships.

5. The equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning according to claim 1 is characterized by: The attention-enhanced word embedding module uses the RoBERTa encoder to encode the input question text q={q1,q2,…,q n } and the relation set r k ={r1,r2,...,r m } to generate the corresponding word embedding vector and Where: E q is the embedding vector of question q; E R For the relationship i The embedding vector of The attention-enhanced word embedding module introduces the kernel function k(h v ,h u ) to calculate the similarity between nodes: Where: h v and h u is the embedding vector of nodes v and u; σ is a hyperparameter that controls the smoothness of the kernel; When performing graph convolution operations, the attention-enhanced word embedding module introduces an attention mechanism into the graph convolution network, combines node features for weighted aggregation, and updates the representation of node v: Where: α vu is the attention weight between node v and its adjacent node u; W (k) represents the weight matrix; )and are the embedding vectors of nodes v and u before updating respectively; N(v) represents the set of neighbor nodes of node v; The attention-enhanced word embedding module adds the representation of the input node to the updated node representation to form a residual connection, and processes it through a normalization layer so that the node representation maintains a stable distribution in each layer; then a nonlinear transformation is performed through a feedforward neural network to increase the nonlinear expression capability; the output result of the feedforward neural network is standardized again through layer normalization to ensure that the final node representation has good numerical stability, thereby obtaining a semantic embedding vector.

6. The equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning according to claim 1 is characterized by: In the long-term and short-term temporal graph convolutional network reasoning module, the graph convolutional network module uses a spatiotemporal joint convolution kernel to combine the spatial topology of the graph structure and the temporal information of the nodes to capture the dynamic changes of nodes in the spatiotemporal dimension and the long-range dependencies between nodes; the principle of the spatiotemporal joint convolution kernel is: for a node v and an adjacent node u of a given graph structure G, the spatiotemporal joint convolution kernel updates the representation of the node by calculating the features and temporal information between the nodes: Where: N(v) represents the set of neighbor nodes of node v; α vu is the weight calculated based on the similarity between nodes; W (k) represents the weight matrix; )and are the embedding vectors of nodes v and u before updating; k(h v ,h u ,t v ,t u ) is the spatiotemporal joint convolution kernel; h v and h u is the embedding vector of nodes v and u; t v and t u are the times of node v and node u respectively; the spatiotemporal joint convolution kernel is expressed as: Where: ‖h v ―h u ‖ 2 is the Euclidean distance between the feature vectors of node v and node u; σ is a hyperparameter that controls the influence of feature similarity; |t v ―t u | is the difference in time between node v and node u; τ is a hyperparameter that controls the effect of the time difference.

7. The equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning according to claim 1 is characterized by: In the long-term and short-term temporal graph convolutional network reasoning module, the Transformer attention module is used to dynamically adjust the information transmission between nodes, focusing on the nodes most relevant to the query to optimize the accuracy of information aggregation; the principle of the Transformer attention module is: Where: Q, K and V are query vector, key vector and value respectively; d k is the dimension of the key vector.

8. The equipment operation and maintenance decision-making method based on large-scale knowledge graph multi-granularity learning reasoning according to claim 1 is characterized by: The long short-term memory network module includes an LSTMInstruction module and an LSTM network; the LSTMInstruction module is used to encode the query text input, generate a query instruction vector, and convert the query text into a word embedding; the LSTM network processes the input word embedding through a recursive formula to generate a temporal hidden state, and the hidden state is used as an instruction vector to guide the reasoning process of the graph convolution operation.

Images