Road electromechanical equipment fault knowledge graph construction and dynamic updating method based on heterogeneous multi-source data

Abstract

The invention relates to a road electromechanical equipment fault knowledge graph construction and dynamic updating method based on heterogeneous multi-source data, and the method comprises the steps: collecting operation and maintenance data of road electromechanical equipment, inputting the operation and maintenance data into an entity recognition model for operation and maintenance entity recognition, the entity recognition model being constructed by a BERT layer, a CNN layer, a BiLSTM layer and a CRF layer; constructing a predefined rule, extracting a semantic relationship between the entities, carrying out entity alignment, and constructing a basic fault knowledge graph; the basic fault knowledge graph is subjected to inductive link prediction based on a dynamic updating model of feature enhancement, dynamic updating of the knowledge graph is completed, the dynamic updating model converts an entity graph into a relation graph, the similarity of the relation graph is obtained, information transmission is carried out, the relation graph is optimized, and the dynamic updating of the knowledge graph is completed. And predicting the relationship of unknown entities. Reliable knowledge support is provided for intelligent operation and maintenance of equipment.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and in particular to a method for constructing and dynamically updating a knowledge graph of highway electromechanical equipment faults based on heterogeneous multi-source data. Background Technology

[0002] With the continuous expansion of my country's highway network and the constant improvement of its intelligent level, the complexity and importance of the operation and maintenance management of electromechanical equipment, as the core infrastructure ensuring the stable operation of highways, are becoming increasingly prominent. Highway electromechanical equipment encompasses multiple subsystems such as toll collection, monitoring, communication, and power supply and distribution, generating operation and maintenance data characterized by wide sources, heterogeneous structure, and dynamic growth. Knowledge graph technology, with its powerful semantic processing and relational expression capabilities, provides an effective way to integrate multi-source operation and maintenance data and realize structured knowledge storage and intelligent applications, demonstrating great potential in areas such as equipment fault diagnosis and health management.

[0003] However, constructing high-quality knowledge graphs for the field of highway electromechanical equipment faults still faces many challenges. First, maintenance data largely originates from manually recorded work orders and reports, resulting in inconsistent text quality and a large amount of unstructured and semi-structured information, making it difficult to guarantee the accuracy of entity extraction and relationship recognition. Second, different data sources describe the same fault entity or phenomenon in various ways (e.g., "monitor not lit" vs. "monitor not displaying"), leading to serious semantic inconsistencies. Effective entity alignment methods are urgently needed to ensure the uniqueness and standardization of knowledge. More importantly, electromechanical equipment fault knowledge continuously evolves with equipment iteration and accumulated maintenance experience. Traditional static knowledge graph construction methods struggle to adapt to the dynamic injection of new fault modes and solutions, leading to "knowledge stagnation" in the graph, failing to reflect the latest maintenance status. Therefore, researching a dynamic update mechanism for fault knowledge graphs is crucial for maintaining their timeliness and practicality.

[0004] Existing knowledge graph completion techniques mostly focus on predicting missing links in static environments, making them difficult to directly apply to electromechanical maintenance scenarios where entity sets are open and constantly expanding. In dynamic (or inductive) settings, the model needs the ability to reason about relationships with new entities not seen during training; this is the core of achieving low-cost, automated knowledge graph updates. Current graph neural network-based methods often suffer from insufficient representation learning and weak generalization ability when dealing with sparsely connected new entities.

[0005] To address the aforementioned issues, this invention provides a method for constructing and dynamically updating a knowledge graph of highway electromechanical equipment faults based on heterogeneous multi-source data. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing and dynamically updating a knowledge graph of highway electromechanical equipment faults based on heterogeneous multi-source data, which effectively supports the continuous evolution of the knowledge graph of highway electromechanical equipment faults and provides reliable knowledge support for intelligent operation and maintenance of equipment.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for constructing and dynamically updating a fault knowledge graph for highway electromechanical equipment based on heterogeneous multi-source data, including: The operation and maintenance data of highway electromechanical equipment is collected, and the operation and maintenance data is input into an entity recognition model for operation and maintenance entity recognition. The entity recognition model is constructed by BERT layer, CNN layer, BiLSTM layer and CRF layer. Construct predefined rules to extract semantic relationships between entities and perform entity alignment to build a basic fault knowledge graph; A feature-enhanced dynamic update model is used to perform inductive link prediction on the basic fault knowledge graph, thereby completing the dynamic update of the knowledge graph. The dynamic update model converts the entity graph into a relationship graph, obtains the similarity of the relationship graph to optimize the relationship graph through information transmission, and predicts the relationships of unknown entities.

[0008] Optionally, inputting the operation and maintenance data into the entity recognition model for operation and maintenance entity recognition includes: The BERT layer is used to process the operation and maintenance data to generate a word vector sequence of operation and maintenance text; The CNN layer extracts features from the word vector sequence through one-dimensional convolution operations to obtain character-level features of the operation and maintenance text; A BiLSTM layer is used to capture the long-distance dependencies of the character-level features from both the forward and backward directions, and the feature vector after concatenating the forward and backward hidden states is output. The feature vector is input into the CRF layer to perform global optimization on the entity label sequence. The legal transitions between labels are constrained by the label transition matrix, and the entity label sequence with the highest probability is output to obtain the entity recognition result.

[0009] Optionally, constructing the predefined rules includes: constructing a head entity-relationship-tail entity electromechanical equipment failure relationship extraction rule based on the causal relationship of highway electromechanical equipment failures.

[0010] Optionally, entity alignment can be performed by calculating the cosine similarity of similar entities to determine their similarity and thus achieve entity alignment.

[0011] Optionally, converting an entity diagram into a relationship diagram includes: The relationship between entities is used as nodes in the relationship graph, and the threshold for determining whether there is a connection between two relationships is used as an edge in the relationship graph. The relationship graph is constructed by using the relationship between entities as nodes in the relationship graph, and the threshold is the average probability that there is a connection between two relationships.

[0012] Optionally, obtaining the similarity of the relationship graph to optimize the relationship graph for information transmission, and predicting the relationship of unknown entities includes: A random walk graph kernel is used to obtain the structural similarity between different relation graphs; A graph isomorphic network is used to learn node representations in a relation graph. Based on the structural similarity, relation knowledge is conveyed and transferred to similar neighboring graphs to obtain a relation graph with neighbor-related features. A multilayer perceptron is used to predict the probability of target relationships in the relationship graph of the neighbor-related features to obtain the relationships of unknown entities.

[0013] Optionally, after obtaining the relationships of unknown entities, the following steps are taken: based on the learnable relation domain embedding and range embedding, and the incoming and outgoing relation sets of the entities, calculate the implicit type features in the head and tail entities of each relationship.

[0014] Optionally, the loss function used during the training of the dynamically updated model. for: in, For proportional parameters, The global loss is calculated by constructing a similarity graph using binary cross-entropy. The local loss is derived by performing self-adversarial negative sampling on the positive sample set to obtain the negative sample set, and by constructing refined features from the positive and negative samples.

[0015] The beneficial effects of this invention are as follows: Addressing the issues of diverse and unevenly sourced maintenance data, this invention employs a bottom-up knowledge graph construction method. Through an improved sequence labeling model and rule-constrained relation extraction, it achieves the extraction of structured fault knowledge from multi-source heterogeneous data. To address the challenge of inconsistent representations of similar entities in the maintenance field, a cosine similarity-based entity alignment strategy is introduced to improve the data consistency of the knowledge graph. More importantly, to adapt to the rapid evolution of electromechanical equipment fault knowledge, a feature-enhanced dynamic update model is proposed. Through relation graph reconstruction and implicit feature refinement, it achieves inductive knowledge completion for new fault phenomena. Experimental results show that the proposed dynamic update mechanism performs excellently in link prediction tasks, effectively supporting the continuous evolution of highway electromechanical fault knowledge graphs and providing reliable knowledge support for intelligent equipment maintenance. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the construction of a knowledge graph for the operation and maintenance of electromechanical equipment, as described in an embodiment of the present invention. Figure 2 This is a structural diagram of the BERT-CNN-BiLSTM-CRF entity recognition model according to an embodiment of the present invention; Figure 3 This is a basic structural framework diagram of the BERT model according to an embodiment of the present invention; Figure 4 This is a knowledge link diagram of electromechanical equipment operation and maintenance faults in an embodiment of the present invention; Figure 5 This is a visualization result of the Neo4j knowledge graph for electromechanical operation and maintenance in an embodiment of the present invention. Figure 6 This is a feature-enhanced knowledge graph update model according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] This embodiment proposes a method for constructing and dynamically updating a knowledge graph of highway electromechanical equipment faults based on heterogeneous multi-source data, including: The operation and maintenance data of highway electromechanical equipment is collected and input into the entity recognition model for operation and maintenance entity recognition. The entity recognition model is constructed by BERT layer, CNN layer, BiLSTM layer and CRF layer. Construct predefined rules to extract semantic relationships between entities and perform entity alignment to build a basic fault knowledge graph; A feature-enhanced dynamic update model is used to perform inductive link prediction on a basic fault knowledge graph, thereby completing the dynamic update of the knowledge graph. The dynamic update model converts the entity graph into a relationship graph, obtains the similarity of the relationship graph to optimize the relationship graph through information transmission, and predicts the relationships of unknown entities.

[0021] Furthermore, inputting operation and maintenance data into the entity recognition model for operation and maintenance entity recognition includes: The BERT layer is used to process the operation and maintenance data to generate a word vector sequence of operation and maintenance text; The CNN layer extracts features from the word vector sequence through one-dimensional convolution operations to obtain character-level features of the operation and maintenance text; A BiLSTM layer is used to capture the long-distance dependencies of the character-level features from both the forward and backward directions, and the feature vector after concatenating the forward and backward hidden states is output. The feature vector is input into the CRF layer to perform global optimization on the entity label sequence. The legal transitions between labels are constrained by the label transition matrix, and the entity label sequence with the highest probability is output to obtain the entity recognition result.

[0022] Specifically, knowledge graph construction techniques are mainly divided into two approaches: top-down and bottom-up. The top-down approach constructs an ontology from a schema layer, then extracts schema information from high-quality data and matches it with the ontology to form a data layer. This approach is suitable for processing structured data such as encyclopedic data. The bottom-up approach utilizes data extraction techniques to extract entities and relationships from all data, and selects entities and relationships with high confidence to summarize into a schema layer. This method is suitable for constructing knowledge graphs from semi-structured and unstructured data. Since the data in this embodiment is basic maintenance information for highway electromechanical systems, including unstructured data, this embodiment adopts a bottom-up approach to construct the electromechanical maintenance knowledge graph. The construction process is as follows: Figure 1 As shown.

[0023] Highway electromechanical equipment comprises multiple subsystems, such as toll collection facilities, monitoring facilities, tunnel electromechanical facilities, power supply and distribution facilities, communication facilities, network security facilities, and lighting facilities. These devices not only ensure the daily operation of highways but also form a crucial foundation for realizing highway informatization and intelligentization. However, due to the large variety and quantity of equipment involved, it is difficult to establish unified maintenance standards, leading to challenges in fault analysis during operation and maintenance. In one embodiment, the operation and maintenance information of electromechanical equipment on a highway operated and managed by a certain group in Gansu Province is used as the basis for analysis, as detailed in Table 1.

[0024] Table 1 The maintenance data includes information such as work order number, affiliated organization, affiliated system, type, fault level, fault phenomenon, work order status, occurrence time, end time, whether it has timed out, repair location, repair content, and fault cause. By collecting and organizing this maintenance data, this embodiment identifies the key attributes that need to be extracted for knowledge extraction, providing the necessary underlying data for building a knowledge graph for the field of transportation electromechanical maintenance and for question-and-answer applications.

[0025] An entity is an instantiation of an ontology concept. Entity recognition requires identifying the start and end points of the entity and determining its type. In the field of electromechanical equipment operation and maintenance, there are many entity types and texts of varying lengths, posing a significant challenge to accurately predicting the adjacency relationships between entities. To dynamically represent text semantic features and address sentence ambiguity, this embodiment proposes a BERT-CNN-BiLSTM-CRF entity recognition model, consisting of four parts: a BERT layer, a CNN layer, a BiLSTM layer, and a CRF layer. The functions of each layer are described below. The model structure is as follows: Figure 2 As shown, it includes: 1. BERT layer: Input: Original sequence of electromechanical equipment maintenance text; Function: Obtains deep bidirectional semantic representation of text through a pre-trained BERT model, generating word vector sequences containing contextual information; Output: The context vector representation for each word; Output flow: The output of the BERT layer is used as the input of the CNN layer.

[0026] 2. CNN layer: Input: The sequence of word vectors output by the BERT layer; Function: Extracts local character-level features from text through one-dimensional convolution operations, enhancing sensitivity to entity boundaries; Output: Feature map after convolutional feature extraction; Output flow: The output of the CNN layer is used as the input of the BiLSTM layer.

[0027] 3. BiLSTM layer: Input: Feature sequences output by CNN layers; Function: By using a bidirectional long short-term memory network to capture long-distance dependencies in text sequences from both forward and backward directions, the ability to model word order and semantic structure in entity recognition is further optimized. Output: The feature vector obtained by concatenating the forward and backward hidden states at each time step; Output flow: The output of the BiLSTM layer is used as the input of the CRF layer.

[0028] 4. CRF layer: Input: Sequence features output from the BiLSTM layer; Function: Utilizes a conditional random field model to globally optimize entity label sequences, constrains the legal transitions between labels through a label transition matrix, and outputs the entity label sequence with the highest probability; Output: The final entity recognition result, i.e., the entity type label sequence corresponding to each word.

[0029] To improve the accuracy of text semantic representation in the field of electromechanical equipment faults, this embodiment employs a BERT layer with a BERT pre-trained module to process maintenance data. The BERT module is characterized by its ability to generate deep bidirectional language representations. By combining knowledge of electromechanical equipment faults with the learning experience of the pre-trained model on general data, the semantic representation capability of maintenance text vectors is improved, thereby reducing the interference of sentence semantic sparsity on model recognition. Its basic structure is as follows: Figure 3 As shown.

[0030] In this embodiment, a one-dimensional convolutional layer is added to the second layer. The relation extraction capability of the CNN layer is used to learn character-level features of the maintenance text, thereby achieving accurate entity word prediction. Furthermore, considering the limited semantic representation of a single character, a BiLSTM layer is constructed by arranging LSTM layers in both the forward and backward directions. This fully captures the text arrangement patterns contained in the electromechanical equipment fault data, enabling the model to learn long-distance dependency information and thus improving the entity recognition performance of the electromechanical equipment fault data.

[0031] After connecting the CRF layer to the network's output layer, its main function is to calculate the probability of an entity belonging to a category, thereby obtaining the optimal sequence for the objective function—that is, acquiring a sequence of entity labels that is probabilistically optimal and structurally compliant. Objective function It includes the input state feature function and the label transition feature function, and their calculation formulas are as follows: (1); in, This is the score matrix output by the network layer. To transform the fractional matrix, The corresponding statement The word Each marked score, The input is the label to be predicted. The transition feature matrix is ​​used to store the probability of transitions between all labels.

[0032] Furthermore, the construction of predefined rules includes: constructing mechanical and electrical equipment failure relationship extraction rules based on the causal relationship of highway mechanical and electrical equipment failures, using a head entity-relationship-tail entity structure.

[0033] Specifically, relation extraction can identify semantic relationships between entities, providing a foundation for generating triples for electromechanical equipment faults. In electromechanical equipment fault data, most relationships are causal. Therefore, this study first extracts all causal relationships and corrects erroneous data through manual verification. Considering that the relationships between different types of entities are relatively fixed, a rule-based method is used to process relation extraction from unstructured text. For example, the relationship between "fault phenomenon" and "fault cause" is a causal relationship. Therefore, this study constructs a rule for "[fault phenomenon] cause [fault cause]", and the specific electromechanical equipment fault relation extraction rules are shown in Table 2.

[0034] Table 2 After extracting relations according to the rules defined in Table 2, all relations were manually verified. The resulting distribution of relation categories and quantities is shown in Table 3.

[0035] Table 3 Furthermore, this embodiment further demonstrates the relationship between fault phenomena (hardware and software) and fault causes in the electromechanical equipment fault triplet data obtained from the entity recognition and relation extraction steps, and visualizes the association between entities and different relation types, such as... Figure 4 As shown.

[0036] Furthermore, entity alignment includes determining similarity by calculating the cosine similarity of similar entities to achieve entity alignment.

[0037] Specifically, the alignment of similar operation and maintenance entities includes: In the raw data of electromechanical equipment operation and maintenance, the same entity may have multiple representations. Therefore, it is necessary to unify these representations to ensure the uniqueness of the same entity. For example, the entities "monitor not displaying", "monitor not displaying signal", and "monitor not lit" can all be summarized as "monitor not displaying". The goal of entity alignment is to ensure that the same entity has the same representation. To this end, the similarity of similar entities is determined by calculating their cosine similarity, thereby achieving entity alignment. The calculation process is as follows: (2); First, dictionary features are constructed based on entity names. Second, a cosine similarity algorithm is used to match similar entities in the dictionary. When the cosine similarity value is close to 1, it indicates that the two entities have a high degree of similarity and can be considered as the same entity. This study sets an alignment threshold of 80% for the calculated result and performs manual review of the results. Some entity alignment results are shown in Table 4.

[0038] Table 4 Finally, through the above process of constructing the transportation electromechanical operation and maintenance knowledge graph, a total of 2665 triplet data points of electromechanical equipment faults were obtained. The study used Neo4j to visualize these data, and the results are as follows: Figure 5 As shown.

[0039] Furthermore, converting an entity graph into a relationship graph includes: using the relationships between entities as nodes in the relationship graph, using a threshold for determining whether there is a connection between two relationships as an edge in the relationship graph, and constructing the relationship graph, wherein the threshold is the average probability that there is a connection between two relationships.

[0040] Furthermore, obtaining the similarity of the relationship graph to optimize the relationship graph for information transmission, and predicting the relationships of unknown entities includes: A random walk graph kernel is used to obtain the structural similarity between different relation graphs; A graph isomorphic network is used to learn node representations in a relation graph. Based on the structural similarity, relation knowledge is conveyed and transferred to similar neighboring graphs to obtain a relation graph with neighbor-related features. A multilayer perceptron is used to predict the probability of target relationships in the relationship graph of the neighbor-related features to obtain the relationships of unknown entities.

[0041] Furthermore, obtaining the relationships of unknown entities includes: Based on the learnable relation domain embedding and range embedding, and the incoming and outgoing relation sets of entities, calculate the implicit type features in the head and tail entities of each relation.

[0042] Specifically, knowledge graph updates include: This study collected statistical tables of electromechanical equipment failures on a highway belonging to a group in Gansu Province, and referenced relevant literature on highway electromechanical operation and maintenance. Based on the aforementioned knowledge graph construction process, the electromechanical equipment failure data was represented as basic triplet data. Since the "failure phenomena" and "failure causes" of electromechanical equipment are constantly updated in reality, there is a real need to expand the existing knowledge graph with new entities. Therefore, an inductive prediction task is introduced to achieve dynamic updates of the electromechanical equipment failure knowledge graph. The inductive prediction task can generate independent embeddings for invisible entities, corresponding to data that has not yet appeared in the actual electromechanical equipment knowledge. By constructing an inductive prediction model, the accuracy and completeness of the electromechanical equipment failure knowledge graph can be guaranteed.

[0043] 1. Inductive link prediction task: The task of inductive relational reasoning is to predict relations between unseen entities (unknown entities). A target triple in KGs. It can be defined as a training set In , , and These are the head entity, the relationship to be predicted, and the tail entity, respectively. It is the training set The set of relations in For a set of nodes, It is a set of triples. The purpose of relation prediction is to predict the relationship on the training set. Learn a scoring function and predict the test set. Invisible entities and The relationship. Test set and training set Should meet and Relationship. Furthermore, The number of edges in the subgraph being represented can be expressed as: ,in Let be the number of nodes in the subgraph. Relational embedding is represented as... ,in For the number of relations, It represents the shortest topological distance between two entities. The relation embedding is parameterized as a learnable matrix that can be shared between the training and test graphs.

[0044] 2. Knowledge Graph Update Model: Since many prediction targets in the knowledge graph of electromechanical equipment faults have only a few neighborhood nodes, it is difficult to generate independent embeddings for entities with only a few links using a subgraph-based update model. Therefore, this paper proposes a feature-enhanced knowledge graph update model. First, the study considers using a relation transformation algorithm to convert the sparse entity structure into a relation graph. The similarity between two sub-relation graphs is judged by the graph kernel to obtain prior knowledge of similar sub-relation graphs and add additional relevance features. Second, the relation features of the connecting nodes are further refined to obtain implicit type features in the head and tail entities of each relation. Finally, an improved performer mechanism is proposed to represent the relevance between the prediction and existing relations, thereby improving the accuracy of relation matching. This framework is as follows: Figure 6 As shown.

[0045] 2.1 Generate a relationship diagram: In inductive prediction tasks, since entity states are unknown, relations serve as effective supervisory information between training and testing. Therefore, the high-quality acquisition of relation embeddings is crucial for the prediction task. This study considers converting entity graphs into relation graphs that express the relationships between entities, and extending the relational semantics required for prediction tasks with only a few connected entities by acquiring prior knowledge of similar relational structures. The process of generating relation graphs is as follows: the original entity graph is converted into a relation graph, where the relations between entities are treated as nodes in the relation graph. It is considered an edge. It is a threshold for determining whether a relationship exists between two relations; its value is the probability that an edge exists between the two relations. The average value, The calculation is as follows: (3); (4); in Represents a statistical function. Represents a mapping function. The function `set` represents the summation function. This represents a set that is joined together. and They represent the first The and the first There are n relations, where n is the total number of entities. Indicates the first The head entity connected to the r-th relation. , Indicates the first The tail entity connected by the r-th relation. , The head entity connecting the j-th and r-th relations , The tail entity is the link between the j-th and r-th relations.

[0046] 2.2 Comparison of similar images: The constructed relational graph contains only local semantic information. However, existing methods treat graphs as independent during the training phase, which prevents the acquisition of additional supervision information from other structures. Therefore, a random walk graph kernel is employed to obtain relational information; the graph kernel measures the similarity between two graphs. The graph kernel is defined as: (5); in The inner product on Hilbert space, used for comparing nodes. and Centered substructure , This represents the feature vector used to calculate the frequency count of each substructure. This represents all nodes in the relationship graph. These are the first and second sub-relationship diagrams used for comparison.

[0047] 2.3 Conveying similar information: After determining the similarity of the relationship graphs, a neural network is used for information transfer. Specifically, a graph isomorphism network (GIN) is used to learn the node representations in the relationship graphs, as follows: (6); in Represents the training graph set, Indicates the initial features of each independent relationship graph, superscript l For the first l Convolutional layer.

[0048] Through multiple iterations, each independent graph transmits relational knowledge and transfers it to similar neighboring graphs. The relational graph representing the obtained neighbor-related features is as follows: This enriches the relational information. Furthermore, a multilayer perceptron (MLP) is applied to predict... Target relationship probability: (7); To better integrate association structure information into relation features, the binary cross-entropy is used as the loss function at the global level, as shown below: (8); in For the true target attributes, For the s-th predicted target value, For the s-th true target value, This is the overall target value.

[0049] 2.4 Refine implicit features: Based on the problem of insufficient semantics in the predicted neighborhood of the target in electromechanical equipment, in addition to basic relation embedding In addition, two learnable relational domain embeddings were also considered. and range embedding This aims to capture more type information about entities by refining their implicit features. For entities... In the The embedding features of a layer are derived from the relational range and relational domain of its ingress and egress relationships, and the calculation process is as follows: (9); express The set of incoming relations, express The set of outgoing relationships.

[0050] 3. Model Training: The feature-enhanced dynamic update model uses globally acquired prior knowledge and locally refined feature models as training objectives to achieve global optimization of relation features. This part of the loss value... The reconstruction loss for converting the entity graph into a relation graph is shown in Equation (8). The local design attention coefficient represents the correlation between the prediction and the current relationship. Feature updates are achieved using the optimized performer model, reducing the number of model calculations. Finally, the positive sample set is... Perform self-adversarial negative sampling to obtain a negative sample set. loss value The edge loss, which is the score for positive and negative samples, can be expressed as: (10); in For the Sigmoid function, It is a fixed margin. It is the number of negative samples in each triple. yes The score of the triplet The first one generated by the replaced head or tail entity One negative triplet; The self-adversarial weight is represented as follows: (11); in The sum is 1, which ensures that when negative sampling is summed in equation (10), the final effect is still equivalent to a normal sample. The sampling temperature.

[0051] By minimizing the loss function, the model can better distinguish between positive and negative samples. The entire loss function is shown in the following equation.

[0052] (12); in For proportional parameters, The global loss is calculated by constructing a similarity graph using binary cross-entropy. The local loss is derived by performing self-adversarial negative sampling on the positive sample set to obtain the negative sample set, and by constructing refined features from the positive and negative samples.

[0053] Experimental verification of the method in this embodiment includes: Obtain the electromechanical operation and maintenance summary dataset: Considering the continuous updating of fault knowledge such as "fault phenomena" and "fault causes" of electromechanical equipment in reality, the existing electromechanical operation and maintenance knowledge graph faces the challenge of incorporating new knowledge into the existing graph. To address this, a multi-task completion model is used, combined with multimodal operation and maintenance data, to construct the knowledge graph through inductive tasks, thereby dynamically supplementing newly added knowledge and simulating the scenario where unknown operation and maintenance knowledge is continuously integrated into the existing domain knowledge graph. The specific implementation process is as follows: First, the experiment collected basic fault data of electromechanical equipment on highways managed by a group in Gansu Province. Referring to relevant literature on highway electromechanical operation and maintenance, and following the aforementioned domain knowledge graph construction process, the electromechanical operation and maintenance data was represented as basic triples. Then, based on the differences in the distribution of the number of relation types, the electromechanical operation and maintenance data was divided into two inductive datasets: Traffic_v1 and Traffic_v2. The electromechanical operation and maintenance inductive datasets have the following characteristics: entities in the training set and the test set are disjoint, but relations are interoperable between the training set and the test set. Furthermore, detailed information about the electromechanical operation and maintenance inductive datasets is shown in Table 5.

[0054] Table 5 Performance analysis of completion: By analyzing the node degree of the training graphs in the Traffic_v1 and Traffic_v2 datasets, it was found that the proportions of nodes with a degree greater than 15 were 35.3% and 27.6% in the two datasets, respectively. Furthermore, AUC-PR and Hits@10 were used as metrics to measure completion performance. The experiments were conducted on Windows 10, using an NVIDIA GeForce RTX3090 GPU, Python 3.7, and a PyTorch-based deep learning framework. Hyperparameter settings are shown in Table 6. The following models were selected as baselines: RuleN, Nodepiece, BERTRL, RED-GNN, TACT, REPORT, CoMPILE, and LCILP. The AUC-PR (AP) and Hits@10 (H@10) results of our model and the baseline on the Traffic_v1 and Traffic_v2 datasets are shown in Table 7, with the best and second-best results marked in bold and underlined, respectively.

[0055] Table 6 Table 7 As shown in Table 7, our model achieves the best results in 75% of the AUC-PR and Hits@10 scores across the Traffic_v1 and Traffic_v2 datasets. Specifically, in Traffic_v1, our model improves the AUC-PR score by 17.8% compared to RuleN and by 2.9% compared to the second-best model, LCILP. In Traffic_v1, the Hits@10 score improves the performance by 23.5% compared to RuleN and by 0.8% compared to the second-best model, REPORT. In Traffic_v2, the AUC-PR score improves the performance by 17.9% compared to RuleN and by 2.5% compared to the second-best model, LCILP. Notably, in Traffic_v2, our model performs slightly worse than LCILP in the Hits@10 score. The LCILP model employs a PageRank-based local clustering method to sample relevant subgraphs, effectively reducing the scope of node embedding. In the overall AP and H@10 results for Traffic_v1 and Traffic_v2, this model achieves 75% of the optimal values ​​compared to LCILP. This indicates that this model has comprehensive advantages in the task of summarizing and completing traffic operation and maintenance knowledge, and can effectively complete the task of summarizing and completing electromechanical operation and maintenance knowledge graphs.

[0056] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for constructing and dynamically updating a knowledge graph of highway electromechanical equipment faults based on heterogeneous multi-source data, characterized in that, include: The operation and maintenance data of highway electromechanical equipment is collected, and the operation and maintenance data is input into an entity recognition model for operation and maintenance entity recognition. The entity recognition model is constructed by BERT layer, CNN layer, BiLSTM layer and CRF layer. Construct predefined rules to extract semantic relationships between entities and perform entity alignment to build a basic fault knowledge graph; A feature-enhanced dynamic update model is used to perform inductive link prediction on the basic fault knowledge graph, thereby completing the dynamic update of the knowledge graph. The dynamic update model converts the entity graph into a relationship graph, obtains the similarity of the relationship graph to optimize the relationship graph through information transmission, and predicts the relationships of unknown entities.

2. The method according to claim 1, characterized in that, Inputting the aforementioned operation and maintenance data into the entity recognition model for operation and maintenance entity recognition includes: The BERT layer is used to process the operation and maintenance data to generate a word vector sequence of operation and maintenance text; The CNN layer extracts features from the word vector sequence through one-dimensional convolution operations to obtain character-level features of the operation and maintenance text; A BiLSTM layer is used to capture the long-distance dependencies of the character-level features from both the forward and backward directions, and the feature vector after concatenating the forward and backward hidden states is output. The feature vector is input into the CRF layer to perform global optimization on the entity label sequence. The legal transitions between labels are constrained by the label transition matrix, and the entity label sequence with the highest probability is output to obtain the entity recognition result.

3. The method according to claim 1, characterized in that, The construction of the predefined rules includes: constructing a head entity-relationship-tail entity electromechanical equipment failure relationship extraction rule based on the causal relationship of highway electromechanical equipment failures.

4. The method according to claim 1, characterized in that, Entity alignment involves determining similarity by calculating the cosine similarity of similar entities, thereby achieving entity alignment.

5. The method according to claim 1, characterized in that, Converting an entity diagram into a relationship diagram includes: The relationship between entities is used as nodes in the relationship graph, and the threshold for determining whether there is a connection between two relationships is used as an edge in the relationship graph. The relationship graph is constructed by using the relationship between entities as nodes in the relationship graph, and the threshold is the average probability that there is a connection between two relationships.

6. The method according to claim 5, characterized in that, Optimizing the relationship graph by obtaining the similarity of the graph and performing information transmission based on the graph, and predicting the relationships of unknown entities, includes: A random walk graph kernel is used to obtain the structural similarity between different relation graphs; A graph isomorphic network is used to learn node representations in a relation graph. Based on the structural similarity, relation knowledge is conveyed and transferred to similar neighboring graphs to obtain a relation graph with neighbor-related features. A multilayer perceptron is used to predict the probability of target relationships in the relationship graph of the neighbor-related features to obtain the relationships of unknown entities.

7. The method according to claim 1, characterized in that, After obtaining the relationships of unknown entities, the process includes: calculating the implicit type features of the head and tail entities of each relationship based on the learnable relation domain embedding and range embedding, and the incoming and outgoing relation sets of the entities.

8. The method according to claim 1, characterized in that, The loss function used in training the dynamically updated model for: in, For proportional parameters, The global loss is calculated by constructing a similarity graph using binary cross-entropy. The local loss is derived by performing self-adversarial negative sampling on the positive sample set to obtain the negative sample set, and by constructing refined features from the positive and negative samples.

Images