A method for constructing an operation and maintenance knowledge graph, a rail operation and maintenance method, and related devices

Abstract

The present invention discloses a method for constructing an operation and maintenance knowledge graph, a track operation and maintenance method and related devices. The present invention adopts the parameter initialization of the BiRTE model when training the PRGC model, so that the PRGC model has the stability of the BiRTE model in relationship extraction, ensuring the accuracy of the extraction of operation and maintenance knowledge triples, and adopts a multi-task learning model to filter out noisy operation and maintenance knowledge triples, solving the noise interference of the operation and maintenance knowledge triples, thereby ensuring that the constructed operation and maintenance knowledge graph has better accuracy and robustness.

Description

Technical Field

[0001] The present invention relates to an operation and maintenance knowledge graph construction method, a track operation and maintenance method and related devices, and belongs to the field of intelligent track operation and maintenance. Background Art

[0002] Rail operation and maintenance refers to the management and maintenance of rail transit (subway) facilities. With the increase in rail projects, the content of equipment technical data (product manuals, maintenance manuals, fault handling manuals, fault analysis reports, etc.) has become complicated, which has caused certain obstacles for relevant personnel to access and store data, and further caused difficulties in rail operation and maintenance.

[0003] In recent years, machine learning technologies represented by deep learning have made breakthrough progress, and knowledge graphs have been widely used and developed in many fields. In order to improve the efficiency of rail operation and maintenance, knowledge graphs have been preliminarily applied to rail operation and maintenance. The process is mainly to first build an operation and maintenance knowledge graph based on the data, and then perform operation and maintenance based on the operation and maintenance knowledge graph. However, the accuracy and robustness of the operation and maintenance knowledge graph constructed by the existing methods are average, which affects the rail operation and maintenance effect. Summary of the invention

[0004] The present invention provides a method for constructing an operation and maintenance knowledge graph, a track operation and maintenance method and related devices, which solve the problems disclosed in the background technology.

[0005] According to one aspect of the present disclosure, a method for constructing an operation and maintenance knowledge graph is provided, comprising:

[0006] The operation and maintenance data are preprocessed, and the pre-trained PRGC model is used to extract the operation and maintenance knowledge triples from the pre-processed operation and maintenance data; the PRGC model is trained using the pre-processed training data, and during the training, parameters are extracted from the pre-trained BiRTE model, and the extracted parameters are used as the parameters for initializing the PRGC model; the BiRTE model is trained using the unpreprocessed training data;

[0007] A multi-task learning model is used to filter out noisy operation and maintenance knowledge triplets from the extracted operation and maintenance knowledge triplets;

[0008] The filtered operation and maintenance knowledge triples are used to construct the operation and maintenance knowledge graph.

[0009] In some embodiments of the present disclosure, parameters are extracted from a pre-trained BiRTE model and the extracted parameters are used as parameters for initializing the PRGC model, including extracting embedding layer weights, context coding layer weights and classifier weights corresponding to the task from the pre-trained BiRTE model, and using the extracted weights as parameters for initializing the PRGC model.

[0010] In some embodiments of the present disclosure, a multi-task learning model includes a DistMult model, a TransE model and an MTRL model; the multi-task learning model is used to filter out noisy operation and maintenance knowledge triples from the extracted operation and maintenance knowledge triples, including: using the DistMult model to obtain semantic scores of operation and maintenance knowledge triples; using the TransE model to obtain logical scores of operation and maintenance knowledge triples; and using the MTRL model to filter out noisy operation and maintenance knowledge triples based on the semantic scores and logical scores.

[0011] In some embodiments of the present disclosure, in the DistMult model, entities and relations in the operation and maintenance knowledge triples are embedded into a predefined low-dimensional vector space, and semantic scores of the operation and maintenance knowledge triples are calculated based on the entities and relations in the low-dimensional vector space;

[0012] The semantic score calculation formula is:

[0013] ;

[0014] Where f2 is the semantic score, They are the head entity, relation and tail entity in the low-dimensional vector space respectively, and T represents transpose.

[0015] In some embodiments of the present disclosure, in the TransE model, entities and relations in the operation and maintenance knowledge triples are embedded into a predefined low-dimensional vector space, and relations are regarded as translations from the head entity to the tail entity. Based on the entities and relations in the low-dimensional vector space, the logical scores of the operation and maintenance knowledge triples are calculated;

[0016] The calculation formula for the logic score is:

[0017] ;

[0018] Where f1 is the logical score, || || is the norm, and L n is the L1 or L2 norm, They are the head entity, relation and tail entity in the low-dimensional vector space respectively.

[0019] According to another aspect of the present disclosure, a device for constructing an operation and maintenance knowledge graph is provided, comprising:

[0020] The extraction module preprocesses the operation and maintenance data and uses the pre-trained PRGC model to extract the operation and maintenance knowledge triples from the pre-processed operation and maintenance data. The PRGC model is trained using the pre-processed training data. During the training, parameters are extracted from the pre-trained BiRTE model and the extracted parameters are used as the parameters for initializing the PRGC model. The BiRTE model is trained using the unpre-processed training data.

[0021] The filtering module uses a multi-task learning model to filter out noisy operation and maintenance knowledge triplets from the extracted operation and maintenance knowledge triplets;

[0022] The construction module uses the filtered operation and maintenance knowledge triples to build the operation and maintenance knowledge graph.

[0023] According to another aspect of the present disclosure, a track operation and maintenance method is provided, comprising:

[0024] Adopt the operation and maintenance knowledge graph construction method to build the operation and maintenance knowledge graph;

[0025] In response to rail transit related equipment failures, the maintenance plan for the faulty equipment is obtained from the operation and maintenance knowledge graph to repair the faulty equipment.

[0026] According to another aspect of the present disclosure, there is provided a track operation and maintenance device, comprising:

[0027] The knowledge graph module uses the operation and maintenance knowledge graph construction method to build the operation and maintenance knowledge graph;

[0028] The operation and maintenance module, in response to the failure of rail transit related equipment, obtains the maintenance plan of the faulty equipment from the operation and maintenance knowledge graph to repair the faulty equipment.

[0029] According to another aspect of the present disclosure, a computer-readable storage medium is provided, which stores one or more programs, and the one or more programs include instructions. When the instructions are executed by a computing device, the computing device executes an operation and maintenance knowledge graph construction method or a track operation and maintenance method.

[0030] According to another aspect of the present disclosure, a computer device is provided, including one or more processors and one or more memories, one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing an operation and maintenance knowledge graph construction method or a track operation and maintenance method.

[0031] The beneficial effects achieved by the present invention are as follows: the present invention adopts the parameter initialization of the BiRTE model when training the PRGC model, so that the PRGC model has the stability of the BiRTE model in relationship extraction, ensuring the accuracy of the extraction of operation and maintenance knowledge triples, and adopts a multi-task learning model to filter out noisy operation and maintenance knowledge triples, solving the noise interference of the operation and maintenance knowledge triples, thereby ensuring that the constructed operation and maintenance knowledge graph has better accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flowchart of the method for building the operation and maintenance knowledge graph;

[0033] Figure 2 Data processing flow chart for multi-task learning model;

[0034] Figure 3 A block diagram of the device for constructing the operation and maintenance knowledge graph;

[0035] Figure 4 It is a flow chart of the track operation and maintenance method;

[0036] Figure 5 This is a block diagram of the track operation and maintenance device. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. It is obvious that the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present disclosure and its application or use. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0038] Unless otherwise specified, the relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure.

[0039] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0040] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0041] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0042] It should be noted that similar symbols and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it does not require further discussion in subsequent figures.

[0043] In order to solve the general problems of accuracy and robustness of operation and maintenance knowledge graphs constructed by existing methods, the present disclosure proposes an operation and maintenance knowledge graph construction method, a track operation and maintenance method and related devices, specifically, a knowledge graph is produced based on the knowledge of reinforcement learning and graph neural network.

[0044] Figure 1This is a schematic diagram of an embodiment of the method for constructing an operation and maintenance knowledge graph disclosed in the present invention. Figure 1 The embodiments may be executed by a terminal (such as a computer).

[0045] like Figure 1 As shown, step 1 of the embodiment pre-processes the operation and maintenance data, and uses a pre-trained PRGC (Pipeline Relation Graph Convolution) model to extract operation and maintenance knowledge triples from the pre-processed operation and maintenance data; wherein the PRGC model is trained using the pre-processed training data, and during training, parameters are extracted from the pre-trained BiRTE (Bidirectional Framework for Relational Triple Extraction) model, and the extracted parameters are used as parameters for initializing the PRGC model; the BiRTE model is trained using unpreprocessed training data.

[0046] It should be noted that operation and maintenance materials and training materials are all equipment technical materials, such as product manuals, maintenance manuals, fault handling manuals, fault analysis reports, etc. These documents cover the equipment's technical specifications, operating guides, and detailed troubleshooting information.

[0047] The preprocessing of operation and maintenance data mainly uses the Transformers library and the Hugging Face Transformers library for data processing, which can be done as follows:

[0048] First, a large amount of text information about the equipment is collected from various documents. Then, the Transformers library is used to clean the text information. Special characters and irrelevant formats can be removed through cleaning to ensure the cleanliness and consistency of the text information. The Hugging Face Transformers library is used for word segmentation and part-of-speech tagging. The Hugging Face Transformers library provides accurate and efficient word segmentation and part-of-speech tagging functions, which provides strong support for subsequent entity recognition and relationship extraction tasks.

[0049] The preprocessing process aims to provide the PRGC model with optimized and structured input data so that the model can more effectively learn and understand the relational information in the device technical documents. By integrating natural language processing tools (i.e., Transformers library and Hugging Face Transformers library), the overall process can improve the accuracy and efficiency of text processing in the preprocessing stage.

[0050] It should be noted that the PRGC model uses the extraction mode of the Pipeline model, which leads to the error transmission problem in the PRGC model. This is because the entity extraction and relationship classification models are independent of each other and do not depend on each other. The recognition errors and omissions in the entity extraction stage cannot be corrected and changed, which will directly affect the effect of the relationship classification stage. To solve this problem, the performance of the BiRTE model in the relationship extraction task is used here to conduct targeted transfer learning on the PRGC model and optimize the extraction of operation and maintenance knowledge triples. Specifically, when training the PRGC model, parameters are extracted from the pre-trained BiRTE model, and the extracted parameters are used as the parameters for initializing the PRGC model.

[0051] It should be noted that the BiRTE model is trained using unpreprocessed training data, specifically unlabeled corpus obtained from equipment technical data, i.e., unlabeled text information, to ensure that the BiRTE model can fully capture the high-level features of the data during the training phase.

[0052] The BiRTE model is based on the Transformer encoder and is designed specifically for relation extraction tasks. For unstructured text, the extracted relation triples goal is to identify all possible triples in a given sentence.

[0053] Assume that the input sentence has L tokens (basic units), which can be expressed as:

[0054] S={w1,w2,…,w L};

[0055] In the formula, S is the input sentence, w i is the i-th basic unit, 1≤i≤L.

[0056] The output triple can be expressed as:

[0057] T = {(h,r,t)|h,t∈ε,r k ∈R};

[0058] Always, T is the output triple, R={r1,r2,…,r K} represents a set of K predefined relations, r k represents the kth predefined relationship, 1≤k≤K, ε represents the entity set, h and t are the head entity and tail entity respectively.

[0059] All consecutive token sequences in a sentence whose length is less than C (C < L) are taken as candidate entities. For a sentence with L tokens, the number of candidate entities |ε| can be expressed as: |ε|=L×C+C / 2-C 2 / 2.

[0060] In order to solve the problems that the training process of the BiRTE model will be biased towards negative triples, which will damage the model's ability to recognize positive triples, the large number of training sentences, and too many candidate entities will reduce the training efficiency, n negative entities are randomly selected from ε to train the model together with all ground truth entities. The entities obtained after training are recorded as a new subset as .

[0061] Given a sentence and its , use BERT (Bidirectional Encoder Representationsfrom Transformers) as the sentence encoder to obtain a 128-dimensional context representation for each token:

[0062] [h1,h2,…,h L ] = BERT([x1,x2,…,x L ]);

[0063] In the formula, h i is the 128-dimensional context representation of the i-th token, x i is the input representation of the i-th token, which is the sum of the corresponding tag embedding and position embedding.

[0064] For the problem of inconsistent dimensions of different entities, take the entity The average vector between the start and end tags is represented as:

[0065] e s =(h start +h end ) / 2;

[0066] In the formula, h start is the start mark, h end is the end tag.

[0067] After obtaining the above data, a bipartite graph link operation is performed on it, and the entity projection formula can be expressed as:

[0068] ;

[0069] ;

[0070] Where E is e s 128-dimensional representation, superscript T indicates transposition, W h and W tThey are two projection matrices that reduce the dimension from the token feature space to the head entity space and the tail entity space, which are used to assist in identifying the head and tail of each entity. h and b t is the bias, E head is the starting entity projection, E tail It is the end entity projection.

[0071] For each relation r k , predict the links between each entity pair to determine whether they can form a valid triple:

[0072] ;

[0073] In the formula, σ represents the sigmoid activation function, U k is a specific relation link matrix, which models the correlation between two entities with respect to the kth relation, P k is the probability of selecting this triplet.

[0074] If the selection triple (e s ,r k ,e j ) Exceeding the threshold θ = 1.0, the triple (e s ,r k ,e j ) will be considered correct, otherwise it is wrong, that is, for each relation r k , if the corresponding probability is greater than θ, then the predicted triple is (e s .span, r k ,e j .span); where e s .span represents entity e s The text range of e j .span represents entity e j The text range.

[0075] In the selection of training tasks, masked language modeling is used here, and Perplexity is defined as:

[0076] ;

[0077] Where M is the number of selected triplet samples, P(T m ) is the predicted probability of the BiRTE model for the mth sample. This task enables the BiRTE model to have a deeper understanding of the relational information in the text by learning how to predict masked words.

[0078] The model parameters are randomly initialized, and the objective function is optimized on the training data through repeated iterations. During the entire training process, the indicators of concern include accuracy, which can be expressed as:

[0079] ;

[0080] In the formula, Number of Correct Predictions is the number of correctly predicted samples, that is, the number of samples correctly classified by the model in the prediction, and Total Number of Predictions is the total number of samples, including the number of all samples in the prediction of the model.

[0081] It should be noted that the PRGC model includes a graph convolution layer, a path modeling component and a relationship classification layer. The selection of the graph convolution layer, the path modeling component and the relationship classification layer enables the PRGC model to fully explore the complex contextual relationships between entities in the relationship extraction task. The introduction and mutual collaboration of these components endow the PRGC model with powerful representation learning capabilities, thereby improving the performance of the model in the relationship extraction task.

[0082] The graph convolution layer plays a vital role in the model. The relationship between entities is modeled through the graph convolution layer to capture the complex dependencies in the context. This process can be expressed by the following formula:

[0083] ;

[0084] In the formula, H (l) is the node representation of the lth layer of the graph convolutional layer, H (l+1) is the node representation of the l+1th layer of the graph convolutional layer, is the adjacency matrix, is the degree matrix, W (l) is the weight matrix of the graph convolution layer. This graph convolution operation helps the model understand the complex contextual relationships between entities.

[0085] The introduction of the path modeling component can better capture the association path information between entities. By modeling the path, the model can have a deeper understanding of the relationship between entities.

[0086] The design of the path modeling component can be expressed as:

[0087] ;

[0088] Where P (l+1) The path representation of the l+1th layer of the path modeling component, P (l) The path representation of the lth layer of the path modeling component, * represents the convolution operation, The weight matrix that models the path, The activation function of the path modeling component, which enables the model to better capture long-range dependencies between entities.

[0089] The relationship classification layer is used to classify the extracted relationships. This layer can be used to accurately classify the relationships between entities to obtain the final extraction results.

[0090] The calculation of the relation classification layer can be expressed as:

[0091] ;

[0092] In the formula, is the output of the model, are the node representation and weight matrix of the final layer of the relationship classification layer, and b is the bias vector. Through the relationship classification layer, the specific relationship between entities can be accurately judged.

[0093] In order to make full use of the prior knowledge of the BiRTE model, parameters are extracted from the trained BiRTE model and used as the parameters for initializing the PRGC model. Specifically, the embedding layer weights, context encoding layer weights, classifier weights corresponding to the task, and possible attention mechanism weights are extracted from the pre-trained BiRTE model, and the extracted weights are used as the parameters for initializing the PRGC model. These extracted weights can help the PRGC model quickly acquire basic semantic and relational modeling capabilities during initialization, improving the efficiency and effectiveness of subsequent training.

[0094] Good initialization can provide a good starting point for the model, which is conducive to subsequent fine-tuning. The method of initializing parameters can be expressed by the following formula:

[0095] PRGC_params_initialized=Load_BiRTE_Pretrained_Parameters();

[0096] In the formula, PRGC_params_initialized is the initialization parameter of PRGC, and Load_BiRTE_Pretrained_Parameters() is the function for loading the PRGC model parameters. This function is responsible for extracting parameters from the pre-trained BiRTE model and using these parameters as the initialization of the PRGC model. This ensures that the PRGC model can fully utilize the advantages of BiRTE in relational representation learning.

[0097] When training the PRGC model (i.e., fine-tuning the PRGC model), the cross entropy loss function is used to guide the model to learn on the target task, and its expression is:

[0098] ;

[0099] In the formula, is the cross entropy loss function, c is the number of categories, For the model The predicted probability of each category, is the true label.

[0100] In terms of learning rate adjustment, a gradient descent strategy is adopted to ensure that the model is more robust during fine-tuning. The formula for learning rate adjustment can be expressed as:

[0101] ;

[0102] In the formula, Learning_Rate adjusted is the adjusted learning rate, Learning_Rate initial is the initial learning rate, decay_rate is the decay rate of the learning rate, and epoch is the number of rounds of current training.

[0103] During the training process, the model parameters can be updated through the back-propagation algorithm and the optimizer to minimize the loss function.

[0104] The whole process can be expressed by the following formula:

[0105] PRGC_FineTuned=

[0106] FineTune_ PRGC(PRGC_params_initialized,Target_Task_Data);

[0107] Where PRGC_FineTuned is the fine-tuned PRGC model, FineTune_PRGC is the function of fine-tuning the PRGC model, PRGC_params_initialized is the initialization parameter, and Target_Task_Data is the data of the target task, which is used to update the parameters of the model in fine-tuning to adapt it to the target task.

[0108] The trained PRGC model uses its relation-aware graph convolutional network architecture to capture the semantic features between entities and relations by encoding the context of the text. In the entity extraction stage, the PRGC model identifies the key entities in the text; in the relation classification stage, it infers the relationship between entity pairs based on the contextual information between them, and finally generates the knowledge triple of "entity-relationship-entity".

[0109] return Figure 1,Step 2 of the embodiment uses a multi-task learning model to filter out noisy operation and maintenance knowledge triples from the extracted operation and maintenance knowledge triples.

[0110] It should be noted that the operation and maintenance knowledge triples extracted in step 1 contain noisy knowledge triples, which will lead to low quality and unreliability of the knowledge graph (that is, general robustness). Therefore, reinforcement learning can be used to select high-quality operation and maintenance knowledge triples and filter out noisy operation and maintenance knowledge triples. In addition, the correlation between semantically similar relations can be fully utilized. Through multi-task learning, the triple selection process of similar relations can be trained in a collective manner, making the representation of the knowledge graph more robust.

[0111] In multi-task learning, multiple tasks share a model, which can better utilize data, improve generalization performance and reduce computational complexity.

[0112] The multi-task learning model mainly includes three models, DistMult (embedding model for knowledge graph completion) model, TransE (knowledge graph embedding method) model and MTRL (Multi-Task Deep Reinforcement Learning) model, so the multi-task learning model can be named DistMult-TransE-MTRL model.

[0113] The model defines semantically similar relations in a triple as r1 and r2. Since these two relations are semantically related, the two r1 and r2 share some common information, which allows the two r1 and r2 to be trained in a collective manner; wherein, a golden triplet is selected for each relation through reinforcement learning, and a cleaned training set is provided for the DistMult-TransE-MTRL model; in each state, whether to select the current triplet is decided according to a random strategy, and then the DistMult-TransE-MTRL model is trained based on the cleaned training dataset, and rewards are provided based on the model performance, and the model's selection strategy is further updated.

[0114] See Figure 2 , filter out noisy operation and maintenance knowledge triples from the extracted operation and maintenance knowledge triples. The specific process can be as follows:

[0115] 21) The DistMult model is used to obtain the semantic score of the operation and maintenance knowledge triples, and the TransE model is used to obtain the logical score of the operation and maintenance knowledge triples.

[0116] It should be noted that in the DistMult model, the entities and relations in the operation and maintenance knowledge triples are embedded into a predefined low-dimensional vector space, and the semantic scores of the operation and maintenance knowledge triples are calculated based on the entities and relations in the low-dimensional vector space.

[0117] The DistMult model is a representative model in the category of tensor decomposition, which uses a bilinear function to calculate the semantic score of a triple.

[0118] The semantic score calculation formula is:

[0119] ;

[0120] Where f2 is the semantic score, are the head entity, relation and tail entity in the low-dimensional vector space respectively. T represents transposition. The score of positive triples should be higher than that of negative triples.

[0121] For the DistMult model, the ReLU loss function is used, which can be expressed as:

[0122] ;

[0123] In the formula, is the true label, is the prediction of the DisMult model, that is The corresponding semantic score, is the ReLU loss function; It is a nonlinear function commonly used in mathematics and computing. When the value represented by · is greater than 0, the value is substituted into the calculation. When the value represented by · is less than or equal to 0, 0 is substituted into the calculation. To represent a negative triple, it is generated by replacing the head entity or tail entity of a positive triple with a random entity in the KG (Knowledge Graph).

[0124] It should be noted that in the TransE model, the entities and relations in the operation and maintenance knowledge triples are embedded into a pre-defined low-dimensional vector space, and the relations are regarded as the translation from the head entity to the tail entity. The logical scores of the operation and maintenance knowledge triples are calculated based on the entities and relations in the low-dimensional vector space.

[0125] The TransE model is one of the most widely used KGE (Knowledge Graph Embedding) models. It considers a relationship as a translation from the head entity to the tail entity on the same low-dimensional hyperplane, i.e. Established ;in, are the head entity, relation and tail entity in the low-dimensional vector space, respectively.

[0126] The calculation formula for the logic score is:

[0127] ;

[0128] Where f1 is the logical score, || || is the norm, and L n is the L1 or L2 norm, which is determined by the performance of the model on the validation set. Positive triplets should have higher scores than negative triplets.

[0129] The TransE model uses a marginal loss function, which can be defined as:

[0130] ;

[0131] In the formula, Based on the marginal loss function, is the set of triplets selected in the training set, and To separate the positive and negative examples, , They are The corresponding logical score and the corresponding logical score of (h,r,t).

[0132] 22) Based on the semantic score and logical score, the MTRL model is used to filter out noisy operation and maintenance knowledge triples.

[0133] It should be noted that the MTRL model is based on a multi-task deep reinforcement learning method, which dynamically weighs the impact of different subtasks (such as relationship extraction tasks, noise filtering tasks, and background modeling tasks), and combines contextual features to filter out inaccurate triplets caused by noise or fuzzy semantics, that is, to filter out noisy operation and maintenance knowledge triplets.

[0134] The MTRL model jointly optimizes the objectives of multiple subtasks through a multi-task loss function:

[0135] ;

[0136] In the formula, is the multi-task loss function, Q is the number of tasks, is the weight corresponding to the qth task, is the loss function corresponding to the qth task.

[0137] In the selection of multi-task learning models, the DistMult model and the TransE model are responsible for evaluating the validity of triples from the perspectives of semantic consistency and logical rationality, respectively. Combined with MTRL's ability to dynamically optimize task weights, a synergistic improvement in model performance is achieved. At the same time, a data processing method that combines preprocessing and enhancement is adopted to eliminate noise samples, balance data, and perform hierarchical training to ensure the model's adaptability to complex data scenarios and its ability to recognize long-tail relationships. In addition, the MTRL model reduces interference between tasks and improves the ability to extract valid triples by integrating feature sharing of different tasks. Overall, the model selection and data processing strategy significantly improved the accuracy and robustness of the construction of the operation and maintenance knowledge graph.

[0138] return Figure 1 , step 3 of the embodiment, uses the filtered operation and maintenance knowledge triples to construct an operation and maintenance knowledge graph.

[0139] It should be noted that the construction of the operation and maintenance knowledge graph is mainly to complete the knowledge fusion and storage of operation and maintenance knowledge triples. Specifically, two key technologies, OWL (Web Ontology Language) and Neo4j graph database, can be used for fusion and storage.

[0140] First, OWL is used for knowledge fusion. Through the ontology description language of OWL, the concepts, entities, relationships and attributes in the field of equipment technology are defined. This formal ontology modeling ensures the consistency and semantic clarity of knowledge. By clearly defining various terms and the relationships between them, a structured ontology is created, providing a consistent interpretation framework for equipment technology knowledge. OWL's ontology engineering ensures the formal consistency of knowledge, making the semantic relationships between different knowledge fragments clearly visible.

[0141] Secondly, we chose Neo4j as the graph database engine for knowledge storage. Neo4j stores data in the form of graphs, where entities are represented as nodes and relationships exist in the form of edges. This storage method of graph databases is very suitable for expressing complex associations between entities. It provides efficient query and association analysis capabilities, allowing knowledge graph operators to easily perform complex query operations, thereby better understanding the knowledge structure in the field of equipment technology. Neo4j graph databases provide a highly flexible storage structure, allowing operators to easily expand and update knowledge graphs. At the same time, graph databases perform well in processing entity relationships and supporting complex queries and reasoning.

[0142] The above method uses the parameters of the BiRTE model to initialize the PRGC model when training the PRGC model, so that the PRGC model has the stability of the BiRTE model in relationship extraction, ensuring the accuracy of the extraction of operation and maintenance knowledge triples, and using a multi-task learning model to filter out noisy operation and maintenance knowledge triples, solving the noise interference of operation and maintenance knowledge triples, thereby ensuring that the constructed operation and maintenance knowledge graph has better accuracy and robustness.

[0143] Figure 3 This is a schematic diagram of an embodiment of the operation and maintenance knowledge graph construction device disclosed in the present invention. Figure 3 The embodiment of the invention is a virtual device which can be loaded and executed by a terminal (such as a computer), and includes an extraction module, a filtering module and a construction module.

[0144] The extraction module of the embodiment is configured to preprocess the operation and maintenance data, and use the pre-trained PRGC model to extract operation and maintenance knowledge triples from the pre-processed operation and maintenance data; wherein the PRGC model is trained using the pre-processed training data, and during training, parameters are extracted from the pre-trained BiRTE model, and the extracted parameters are used as parameters for initializing the PRGC model; the BiRTE model is trained using unpreprocessed training data.

[0145] The filtering module of the embodiment is configured to filter out noisy operation and maintenance knowledge triples from the extracted operation and maintenance knowledge triples using a multi-task learning model.

[0146] The construction module of the embodiment is configured to construct an operation and maintenance knowledge graph using filtered operation and maintenance knowledge triples.

[0147] Similar to the above method, this patent uses the parameter initialization of the BiRTE model when training the PRGC model, so that the PRGC model has the stability of the BiRTE model in relationship extraction, ensuring the accuracy of the extraction of operation and maintenance knowledge triples, and using a multi-task learning model to filter out noisy operation and maintenance knowledge triples, solving the noise interference of operation and maintenance knowledge triples, thereby ensuring that the constructed operation and maintenance knowledge graph has better accuracy and robustness.

[0148] Figure 4 is a schematic diagram of an embodiment of the track operation and maintenance method disclosed in the present invention, Figure 4 The embodiments may also be executed by a terminal (such as a computer).

[0149] like Figure 4 As shown, step S1 of the embodiment adopts the above-mentioned operation and maintenance knowledge graph construction method to construct an operation and maintenance knowledge graph.

[0150] Through the above-mentioned operation and maintenance knowledge graph construction method, an operation and maintenance knowledge graph with better accuracy and robustness can be constructed.

[0151] return Figure 4 , step S2 of the embodiment, in response to a failure of rail transit related equipment, a maintenance plan for the faulty equipment is obtained from the operation and maintenance knowledge graph to repair the faulty equipment.

[0152] Rail operation and maintenance uses knowledge graphs to integrate multi-source data such as equipment information, historical faults, and sensor data to build an operation and maintenance knowledge system that includes entities such as equipment, faults, and maintenance methods and their relationships. During the operation and maintenance process, the knowledge graph supports intelligent diagnosis and problem location. It quickly identifies the cause of the fault and the scope of impact by analyzing real-time data, while recommending the best historical maintenance plan and operating steps to optimize the maintenance process. Combined with predictive analysis, the knowledge graph can also perform health assessments and fault predictions, and provide preventive maintenance recommendations. New knowledge accumulated during the operation and maintenance process will be dynamically updated to the graph, continuously improving the intelligence level of rail operation and maintenance, and significantly improving efficiency and reliability.

[0153] The above method performs operation and maintenance based on an operation and maintenance knowledge graph with better accuracy and robustness, which can achieve better operation and maintenance results, and the construction of the knowledge graph can improve the efficiency of operation and maintenance.

[0154] Figure 5 This is a schematic diagram of an embodiment of the track operation and maintenance device disclosed in the present invention. Figure 5 An embodiment of the invention is a virtual device that can be loaded and executed by a terminal (such as a computer), including a knowledge graph module and an operation and maintenance module.

[0155] The knowledge graph module of the embodiment is configured to construct an operation and maintenance knowledge graph using the above-mentioned operation and maintenance knowledge graph construction method.

[0156] The operation and maintenance module of the embodiment is configured to, in response to a failure of rail transit related equipment, obtain a maintenance plan for the faulty equipment from the operation and maintenance knowledge graph to repair the faulty equipment.

[0157] Similar to the above method, this device performs operation and maintenance based on an operation and maintenance knowledge graph with better accuracy and robustness, which can achieve better operation and maintenance results, and the construction of the knowledge graph can improve the efficiency of operation and maintenance.

[0158] Based on the same technical solution, the present disclosure also relates to a computer-readable storage medium, which stores one or more programs, and the one or more programs include instructions. When the instructions are executed by a computing device, the computing device executes an operation and maintenance knowledge graph construction method or a track operation and maintenance method.

[0159] Based on the same technical solution, the present disclosure also relates to a computer device, including one or more processors and one or more memories, one or more programs are stored in the one or more memories and are configured to be executed by one or more processors, and the one or more programs include instructions for executing an operation and maintenance knowledge graph construction method or a track operation and maintenance method.

[0160] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0162] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0163] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0164] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A method for constructing an operation and maintenance knowledge graph, characterized in that: include: The operation and maintenance data are preprocessed, and the pre-trained PRGC model is used to extract the operation and maintenance knowledge triples from the pre-processed operation and maintenance data; the PRGC model is trained using the pre-processed training data, and during the training, parameters are extracted from the pre-trained BiRTE model, and the extracted parameters are used as the parameters for initializing the PRGC model; the BiRTE model is trained using the unpreprocessed training data; A multi-task learning model is used to filter out noisy operation and maintenance knowledge triplets from the extracted operation and maintenance knowledge triplets; Use the filtered operation and maintenance knowledge triples to build an operation and maintenance knowledge graph; Among them, the PRGC model includes a graph convolution layer, a path modeling component, and a relationship classification layer; in the PRGC model, the relationship between entities is modeled through the graph convolution layer, and the formula is expressed as: ； In the formula, H (l) is the first convolutional layer of the graph l The node representation of the layer, H (l+1) is the first convolutional layer of the graph l The nodes at the +1 layer represent is the adjacency matrix, is the degree matrix, W (l) is the weight matrix of the graph convolutional layer, σ Represents the sigmoid activation function; The path modeling component is used to model the association path between entities. The formula is expressed as follows: ； In the formula, P (l+1) Modeling the path component l +1 layer path representation, P (l) Modeling the path component l The path of the layer indicates that * represents the convolution operation. The weight matrix that models the path, Activation functions for path modeling components; The relationship between entities is classified through the relationship classification layer, and the formula is expressed as: ； In the formula, is the output of the model, are the node representation and weight matrix of the final layer of the relation classification layer, b is the bias vector.

2. The method for constructing an operation and maintenance knowledge graph according to claim 1, characterized in that: Extract parameters from the pre-trained BiRTE model and use the extracted parameters as the parameters for initializing the PRGC model, including: The embedding layer weights, context encoding layer weights, and classifier weights corresponding to the task are extracted from the pre-trained BiRTE model, and the extracted weights are used as parameters for initializing the PRGC model.

3. The method for constructing an operation and maintenance knowledge graph according to claim 1, characterized in that: Multi-task learning models include DistMult model, TransE model and MTRL model; A multi-task learning model is used to filter out noisy operation and maintenance knowledge triplets from the extracted operation and maintenance knowledge triplets, including: The DistMult model is used to obtain the semantic score of the operation and maintenance knowledge triples; The TransE model is used to obtain the logical scores of the operation and maintenance knowledge triples; According to the semantic score and logical score, the MTRL model is used to filter out noisy operation and maintenance knowledge triples.

4. The method for constructing an operation and maintenance knowledge graph according to claim 3, characterized in that: In the DistMult model, the entities and relations in the operation and maintenance knowledge triples are embedded into a predefined low-dimensional vector space, and the semantic scores of the operation and maintenance knowledge triples are calculated based on the entities and relations in the low-dimensional vector space. The semantic score calculation formula is: ； In the formula, f 2 is the semantic score, are the head entity, relation and tail entity in the low-dimensional vector space, respectively. T Indicates transpose.

5. The method for constructing an operation and maintenance knowledge graph according to claim 3, characterized in that: In the TransE model, the entities and relations in the operation and maintenance knowledge triples are embedded into a pre-defined low-dimensional vector space, and the relations are regarded as the translation from the head entity to the tail entity. The logical score of the operation and maintenance knowledge triples is calculated based on the entities and relations in the low-dimensional vector space. The calculation formula for the logic score is: ； In the formula, f 1 is the logical score, || || is the norm, L n for L 1 or L 2-norm, They are the head entity, relation and tail entity in the low-dimensional vector space respectively.

6. A device for constructing an operation and maintenance knowledge graph, characterized in that: include: The extraction module preprocesses the operation and maintenance data and uses the pre-trained PRGC model to extract the operation and maintenance knowledge triples from the pre-processed operation and maintenance data. The PRGC model is trained using the pre-processed training data. During the training, parameters are extracted from the pre-trained BiRTE model and the extracted parameters are used as the parameters for initializing the PRGC model. The BiRTE model is trained using the unpre-processed training data. The filtering module uses a multi-task learning model to filter out noisy operation and maintenance knowledge triplets from the extracted operation and maintenance knowledge triplets; The construction module uses the filtered operation and maintenance knowledge triples to build the operation and maintenance knowledge graph; Among them, the PRGC model includes a graph convolution layer, a path modeling component, and a relationship classification layer; in the PRGC model, the relationship between entities is modeled through the graph convolution layer, and the formula is expressed as: ； In the formula, H (l) is the first convolutional layer of the graph l The node representation of the layer, H (l+1) is the first convolutional layer of the graph l The nodes at the +1 layer represent is the adjacency matrix, is the degree matrix, W (l) is the weight matrix of the graph convolutional layer, σ Represents the sigmoid activation function; The path modeling component is used to model the association path between entities. The formula is expressed as follows: ； In the formula, P (l+1) Modeling the path component l +1 layer path representation, P (l) Modeling the path component l The path of the layer indicates that * represents the convolution operation. The weight matrix that models the path, Activation functions for path modeling components; The relationship between entities is classified through the relationship classification layer, and the formula is expressed as: ； In the formula, is the output of the model, are the node representation and weight matrix of the final layer of the relation classification layer, b is the bias vector.

7. A track operation and maintenance method, characterized in that: include: Adopt the method described in any one of claims 1 to 5 to construct an operation and maintenance knowledge graph; In response to rail transit related equipment failures, the maintenance plan for the faulty equipment is obtained from the operation and maintenance knowledge graph to repair the faulty equipment.

8. A track operation and maintenance device, characterized in that: include: The knowledge graph module uses the method described in any one of claims 1 to 5 to construct an operation and maintenance knowledge graph; The operation and maintenance module, in response to the failure of rail transit related equipment, obtains the maintenance plan of the faulty equipment from the operation and maintenance knowledge graph to repair the faulty equipment.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores one or more programs, and the one or more programs include instructions. When the instructions are executed by a computing device, the computing device executes any one of the methods described in claims 1 to 5 and 7.

10. A computer device, characterized in that: include: One or more processors and one or more memories, one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods described in claims 1 to 5 and 7.

Images