Knowledge embedding method for automobile system fault knowledge graph

Through the fusion of pre-trained language model and entity embedding vectors, combining multi-scale feature interaction and dynamic prediction and completion layers, the shortcomings of knowledge graphs in capturing complex interactive information of automotive system failures are solved, and the expansion and improvement of knowledge embedding is achieved, and the improvement of knowledge embedding is achieved. The accuracy of fault diagnosis and the practicality of knowledge graphs.

CN119067206APending Publication Date: 2024-12-03CHONGQING UNIV

Patent Information

Application Number
CN202411042729.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing knowledge graphs have shortcomings in capturing complex interactive information between entities and relationships in automotive system failures, resulting in incomplete knowledge embedding and inefficient computing, and cannot effectively improve the integrity and real-timeness of the knowledge graph.

Method used

The pre-trained language model and entity embedding vector fusion are used to enhance the semantic representation of entities through multi-scale feature interaction methods, and the probability prediction of entities and relationships is used to achieve knowledge of automotive system failure knowledge graphs. Embed and expand and perfect.

Benefits of technology

Improve the accuracy of entity representation and complex relationship understanding ability, enhance the accuracy and efficiency of fault diagnosis, ensure the continuous learning and improvement of knowledge graphs, adapt to the rapid response of unknown faults, and significantly improve the practicality of automotive system fault knowledge graphs Sex and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067206A_ABST
    Figure CN119067206A_ABST
Patent Text Reader

Abstract

The invention discloses a knowledge embedding method for an automobile system fault knowledge graph. The process of the method is divided into three stages of a text enhanced entity embedding vector layer, a multi-scale feature interaction layer and a dynamic prediction and completion layer. The text enhanced entity embedded vector layer performs deep semantic information extraction on an entity description text in a knowledge graph by utilizing a pre-training model, introduces an attention mechanism, and fuses a text embedded vector and a graph embedded vector; the multi-scale feature interaction layer captures local-to-global relationship features through convolution operations of different scales, provides multi-level information analysis capability, and increases residual paths to promote gradient flow in a deep network; and the dynamic prediction and completion layer predicts missing entities and relationships in the knowledge graph through a probability score prediction function so as to provide an information prediction basis for completing the knowledge graph. According to the method, complex interaction between entities and relations is effectively captured, and continuous learning and perfection of the knowledge graph are achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph technology, and in particular to a knowledge embedding method for an automobile system fault knowledge graph. Background Art

[0002] The knowledge graph was proposed by Google in 2012 to improve the user's search experience. In essence, the knowledge graph is a semantic network that reveals the relationship between entities and can formally describe real-world affairs and their relationships. The knowledge graph consists of multiple triples, and the general structure of the triple is "entity-relationship-entity" or "entity-attribute value". Existing large-scale knowledge graphs such as Freebase, WordNet and YAGO are widely used in downstream applications such as recommendation systems, intelligent question answering and information retrieval. Due to the difficulty and imperfection of knowledge acquisition and extraction, existing knowledge graphs are often incomplete. For example, 75% of the entities in the Freebase database lack nationality information and 94% of the entities have no parent information, which limits the role of knowledge graphs in downstream tasks.

[0003] Knowledge graph embedding is a technology used to represent entities and relationships in knowledge graphs by mapping the entities and relationships of knowledge graphs into low-dimensional continuous vector spaces. This representation method enables machines to more efficiently process and reason about complex relationships in knowledge graphs, and is an important knowledge management and reasoning tool.

[0004] In the prior art, the translation model represented by TransE has a simple structure and strong interpretability, but can only effectively model simple relationship patterns and cannot extract the interaction information between entities and relationships. The model based on tensor decomposition can model more complex relationships and obtain feature interactions between entity relationships, but it has a large number of parameters and low computational efficiency. In recent years, neural network-based models have been widely used in link prediction tasks. The convolutional neural network model represented by ConvE uses two-dimensional convolution to capture triple features and has the advantage of a small number of parameters, but it can only obtain interaction information at the stacking of entity relationships and has limited feature extraction capabilities.

[0005] How to capture the complex interactions between entities and relationships, realize the knowledge embedding and expansion of the automobile system fault knowledge graph, and improve the integrity of the knowledge graph has become a problem that needs further research and resolution. Summary of the invention

[0006] In view of the deficiencies of the above-mentioned prior art, the present invention provides a knowledge embedding method for automobile system fault knowledge graph, which obtains text-enhanced entity embedding vectors by fusing pre-trained language models and entity embedding vectors, and adopts a multi-scale feature interaction method to enhance the semantic representation of entities, enhance the model's understanding and representation capabilities of complex relationships between entities, and proposes dynamic prediction and completion to achieve knowledge embedding expansion of automobile system fault knowledge graph and improve the integrity and real-time performance of the knowledge graph.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A knowledge embedding method for automobile system fault knowledge graph includes the following steps:

[0009] Obtaining entity description text in the knowledge graph of automobile system failure; the entity description text is used to describe entity and entity relationship, the entity includes a head entity and a tail entity, the head entity is a feature word describing the automobile system component, the tail entity is a feature word describing the failure event, and the entity relationship is a feature word describing the relationship between the automobile system component and the failure event;

[0010] Using a pre-trained text-enhanced entity embedding layer network, semantic information is extracted from the entity description text to obtain a corresponding text embedding vector; using a graph structure embedding learning method, the embedded representation information of entities and entity relationships in the entity description text is extracted to obtain a corresponding graph embedding vector; then, using an attention mechanism, the text embedding vector and the graph embedding vector are dynamically weighed and fused to obtain a text-enhanced entity embedding vector;

[0011] Using a pre-trained multi-scale feature interaction layer network, interleaved feature reorganization is performed on different entity embedding vectors to obtain a feature reorganization matrix, and residual convolution feature extraction operations of different scales are performed on the reorganized feature matrix through multi-scale residual convolution to extract convolution feature matrices of different scales; then, feature fusion and feature extraction are performed on the convolution feature matrices of different scales to obtain the final feature representation matrix of the entity description text;

[0012] Using the pre-trained dynamic prediction and completion layer network, the probability score prediction values ​​of the entities and entity relationships in the entity description text are calculated according to the final feature representation matrix as the probability prediction results of completing the entities and entity relationships in the knowledge graph of automobile system failures.

[0013] As a preferred solution, in the text-enhanced entity embedding layer network, the text-enhanced entity embedding vector obtained is expressed as:

[0014] e enhanced =softmax(W·[egraph ;e text ]+b)·[e graph ;e text ];

[0015] In the formula, e enhanced is the entity embedding vector after text enhancement, W is the learned weight, b is the bias term, and e text is the text embedding vector obtained by extracting semantic information, e graph is the graph embedding vector extracted by the graph structure embedding learning method, [e graph ;e text ] is the concatenation of the graph embedding vector and the text embedding vector extracted by the graph structure embedding learning method.

[0016] As a preferred solution, during the training process of the text-enhanced entity embedding layer network, the loss function L of the text-enhanced entity embedding layer network is minimized. e The loss function L of the text-enhanced entity embedding layer network is obtained by training the target; e for:

[0017]

[0018] In the formula, N is the number of training samples, The entity embedding vector predicted by the text-enhanced entity embedding layer network during training, Embedding vector for the preset labeled target entity, ||·|| 2 is the square of the Euclidean distance;

[0019] As a preferred solution, in the multi-scale feature interaction layer network, the specific method of performing interleaved feature reorganization on different entity embedding vectors is:

[0020] Taking every two different entity embedding vectors as a reorganization group, the elements of the two entity embedding vectors in the reorganization group are reorganized in an interlaced manner using the checkerboard reorganization method; the interlaced feature reorganization method for the two entity embedding vectors in each reorganization group is expressed as:

[0021] M=reshape(interleave(e1,e2));

[0022] Where M is the feature reorganization matrix obtained after feature reorganization, e1 is an entity embedding vector in the reorganized group, e2 is another entity embedding vector in the reorganized group, interleave(e1,e2) is a function that interleaves the elements of e1 and e2, and reshape represents a reshaping function that reshapes the interleaved vector into a matrix form.

[0023] As a preferred solution, in the multi-scale feature interaction layer network, the specific method of performing residual convolution feature extraction operations of different scales on the reorganized feature matrix through multi-scale residual convolution is:

[0024] Each reorganized feature matrix is ​​processed by residual one-dimensional convolution, residual two-dimensional convolution and residual dilated convolution to obtain the convolution feature matrix of the corresponding scale; where:

[0025] Residual one-dimensional convolution, which is defined as: P1 = ReLU (M + Conv1D (M))

[0026] Where P1 is the convolution feature matrix obtained after one-dimensional convolution, Conv1D(M) is the one-dimensional convolution applied to the reorganized feature matrix M, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function;

[0027] Residual 3D convolution, defined as: P2 = ReLU(M + Conv3D(reshape(M)))

[0028] Where P2 is the convolution feature matrix obtained after three-dimensional convolution, Conv3D(reshape(M)) is the application of three-dimensional convolution to the reorganized feature matrix M, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function;

[0029] Residual dilated convolution, which is defined as: P3 = ReLU (M + DilatedConv (M, rate = 2))

[0030] Where P3 is the convolution feature matrix obtained after dilated convolution, DilatedConv(M,rate=2) is the dilated convolution applied to the reorganized feature matrix M, with a dilation rate of 2, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function.

[0031] As a preferred solution, in the multi-scale feature interaction layer network, the specific method of performing feature fusion and feature extraction on the convolution feature matrices of different scales is:

[0032] Perform feature fusion operation on residual convolution features of different scales:

[0033] P final =FusionLayer(P1,P2,P3);

[0034] Where P final is the obtained fusion feature representation matrix, P1, P2, P3 are the outputs of residual one-dimensional convolution, residual three-dimensional convolution and residual dilated convolution respectively, FusionLayer represents the feature fusion operation based on multi-layer perceptron or self-attention mechanism;

[0035] Then the fusion feature representation matrix P final Perform further feature extraction:

[0036] P extracted =ReLU(FC(P final ));

[0037] Where P extracted is the final feature representation matrix of the entity description text; FC represents the fully connected layer, and ReLU is the activation function.

[0038] As a preferred solution, during the training process of the multi-scale feature interaction layer network, the loss function L of the multi-scale feature interaction layer network is minimized. p The loss function L of the multi-scale feature interaction layer network is obtained by training the target; p for:

[0039] L p =βL feature +(1-β)L task ;

[0040] Where, L feature is the feature fidelity loss, L task is the task guidance loss; β is the loss adjustment hyperparameter; where:

[0041]

[0042] Where m is the number of convolutional layers contained in the multi-scale residual convolution, P j is the output of the jth convolutional layer, To decode P through the decoder network j The reconstructed features obtained by decoding, α j is the reconstruction error weight of the jth convolutional layer, and MSE represents the mean square error operation;

[0043]

[0044] Where N is the number of training samples, C is the number of fault cause categories, and y i,c is the true label of training sample i in fault cause category c, p i,c The probability that the multi-scale feature interaction layer network predicts that the training sample i belongs to the fault cause category c, and p i,c =softmax(P final ).

[0045] As a preferred solution, in the dynamic prediction and completion layer network, the calculated probability score prediction value includes:

[0046] The probability score prediction value φ of the tail entity: φ(P extracted ,t)=softmax(W t ·P extracted );

[0047] The probability score prediction value ψ of the head entity: ψ(P extracted ,h)=softmax(W h ·P extracted );

[0048] The probability score prediction value θ of the entity relationship: θ(P extracted ,r)=softmax(W r ·P extracted );

[0049] Where P extracted is the final feature representation matrix of the entity description text, t, h, r represent the tail entity, head entity and entity relationship respectively, W t , W h , W r They are the weight matrices of the tail entity, head entity, and entity relationship, respectively.

[0050] As a preferred solution, in the dynamic prediction and completion layer network, the weight matrices of the tail entity, the head entity and the entity relationship are obtained by training the dynamic prediction and completion layer network with the goal of minimizing their respective loss functions, and their respective loss functions are:

[0051] Probability score prediction loss function L for tail entity tail for:

[0052] In the formula, The model predicts that the head entity and the tail entity corresponding to the relationship of the sample are the true tail entities t i The probability of

[0053] The probability score prediction loss function L for the head entity head for:

[0054] In the formula, The model predicts that the head entity corresponding to the tail entity and relationship of the sample is the real head entity h i The probability of

[0055] Probability score prediction loss function L for entity relations relation for:

[0056] In the formula, The model predicts that the relationship between the head entity and the tail entity of a sample is the true relationship r i The probability of

[0057] As a preferred solution, in the process of knowledge embedding for the automobile system fault knowledge graph, adversarial training is also performed on the dynamic prediction and completion layer network to update the weight matrix in the dynamic prediction and completion layer network; the specific method is:

[0058] In the process of adversarial training of the dynamic prediction and completion layer network, the gradient of each probability score prediction loss function with respect to the weight matrix is ​​calculated respectively, and the weight matrix is ​​iteratively updated until the number of iterations reaches the preset iteration number threshold, the training is terminated, and the update of the weight matrix is ​​completed; wherein, the update of the weight matrix in each iteration is expressed as:

[0059] W t ′=W t -η·▽ Wt L tail ;

[0060] W h ′=W h -η·▽ Wh L head ;

[0061] W r ′=W r -η·▽ Wr L relation ;

[0062] Where η is the learning rate of adversarial training; Wt L tail Represents the probability score prediction loss function L of the tail entity tail The weight matrix W of the tail entity t The gradient of Wh L head Denotes the probability score prediction loss function L of the head entity head The weight matrix W of the head entity h The gradient of Wr L relation The probability score prediction loss function L representing the entity relationship relation About the weight matrix W of entity relations r gradient.

[0063] Compared with the prior art, the present invention has the following technical effects:

[0064] (1) The present invention obtains deep semantic information through semantic information extraction, uses graph structure embedding to analyze the relationship between entities, and uses the attention mechanism to dynamically fuse and adjust the entities learned from graph structure embedding with the text vectors extracted from the language model, thereby improving the accuracy and richness of entity representation. In addition, through in-depth semantic analysis, it helps the automotive system to better identify and classify fault types, thereby improving the accuracy and efficiency of fault diagnosis.

[0065] (2) The present invention enhances the feature interaction between entities and relationships by utilizing feature reorganization and staggered arrangement, so that the knowledge embedding model can better capture complex fault relationships. It also uses a multi-scale convolutional network and a residual convolutional network to extract features at different scales, effectively preventing information loss and gradient vanishing, enhancing feature transfer and gradient flow, and thus effectively enhancing the knowledge embedding model's ability to understand and represent complex relationships between entities.

[0066] (3) The present invention dynamically predicts and completes the missing entities and relationships in the knowledge graph through a dynamic prediction and completion layer, thereby adapting to the update and change of the knowledge graph, realizing continuous learning and improvement of the knowledge graph, ensuring the adaptability and response speed of fault diagnosis when facing unknown faults, and significantly improving the practicality and reliability of the automotive system fault knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0068] Figure 1 It is a flow chart of the knowledge embedding method for automobile system fault knowledge graph of the present invention;

[0069] Figure 2 An example diagram of a workflow for completing and updating a knowledge graph using the knowledge embedding method of the present invention for an automobile system fault knowledge graph. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0071] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0072] The data in the automotive system fault knowledge graph is obtained based on the vehicle fault information. Compared with the text in the general field, these fields have the following characteristics: (1) Professionalism: automotive system fault data contains a large number of professional terms and technical details; (2) Unstructuredness: fault descriptions often exist in the form of free text, containing rich but complex information; (3) Complexity of relationship patterns: There are complex dependencies and influence relationships between components and faults in the automotive system; (4) Incomplete information: There is often missing information in the knowledge graph, such as unknown fault causes or unlabeled component relationships, and it needs to be updated in a timely manner.

[0073] However, the knowledge graph extension models in the prior art have the following main problems: for translation models, when there are complex relationships of one-to-many, many-to-one, and many-to-many types, the modeling ability is limited and cannot fully capture the complex interactions between entities and relationships; for models based on tensor decomposition, it is usually necessary to allocate a large number of parameters to each relationship and entity in the knowledge graph, which increases the number of model parameters. The model training and inference process may require a lot of computing resources and time, and in the case of sparse data, the model is difficult to generalize to unseen data, increasing the risk of overfitting; for convolutional neural network models, the main focus is on the extraction of local information of triples, and the global context may be ignored; the model interpretability is relatively low, especially in terms of the interpretability of model decisions.

[0074] To this end, the present invention proposes a knowledge embedding method for an automobile system fault knowledge graph, which can be divided into three stages: a text-enhanced entity embedding vector layer, a multi-scale feature interaction layer, and a dynamic prediction and completion layer; wherein, the text-enhanced entity embedding vector layer extracts deep semantic information from the entity description text in the knowledge graph by utilizing a pre-trained model, and introduces an attention mechanism to fuse the text embedding vector and the graph embedding vector, thereby enhancing the richness and accuracy of the entity representation; the multi-scale feature interaction layer captures relational features from local to global through convolution operations of different scales, provides multi-level information parsing capabilities, increases residual paths to promote gradient flow in deep networks, and prevents information loss and gradient vanishing problems during training; the dynamic prediction and completion layer predicts missing entities and relationships in the knowledge graph through a probability scoring prediction function, so as to provide an information prediction basis for completing the knowledge graph.

[0075] Specifically, if Figure 1 As shown, the knowledge embedding method for automobile system fault knowledge graph of the present invention specifically includes the following steps:

[0076] Obtaining entity description text in the knowledge graph of automobile system failure; the entity description text is used to describe entity and entity relationship, the entity includes a head entity and a tail entity, the head entity is a feature word describing the automobile system component, the tail entity is a feature word describing the failure event, and the entity relationship is a feature word describing the relationship between the automobile system component and the failure event;

[0077] Using a pre-trained text-enhanced entity embedding layer network, semantic information is extracted from the entity description text to obtain a corresponding text embedding vector; using a graph structure embedding learning method, the embedded representation information of entities and entity relationships in the entity description text is extracted to obtain a corresponding graph embedding vector; then, using an attention mechanism, the text embedding vector and the graph embedding vector are dynamically weighed and fused to obtain a text-enhanced entity embedding vector;

[0078] Using a pre-trained multi-scale feature interaction layer network, interleaved feature reorganization is performed on different entity embedding vectors to obtain a feature reorganization matrix, and residual convolution feature extraction operations of different scales are performed on the reorganized feature matrix through multi-scale residual convolution to extract convolution feature matrices of different scales; then, feature fusion and feature extraction are performed on the convolution feature matrices of different scales to obtain the final feature representation matrix of the entity description text;

[0079] Using the pre-trained dynamic prediction and completion layer network, the probability score prediction values ​​of the entities and entity relationships in the entity description text are calculated according to the final feature representation matrix as the probability prediction results of completing the entities and entity relationships in the knowledge graph of automobile system failures.

[0080] Figure 2 An example of a workflow for completing and updating a knowledge graph using the knowledge embedding method for automobile system fault knowledge graph of the present invention is shown. In conjunction with the accompanying drawings, the knowledge embedding method for automobile system fault knowledge graph of the present invention is described in more detail through three stages: text-enhanced entity embedding vector layer, multi-scale feature interaction layer, and dynamic prediction and completion layer.

[0081] 1. Text-enhanced entity embedding vector layer

[0082] The text-enhanced entity embedding vector layer improves the accuracy and richness of entity representation by combining the graph structure intrusion of the entity and the semantic information extracted from the text description, thereby better understanding and processing complex automobile fault descriptions and helping the system better identify and distinguish fault types.

[0083] 1.1 Text Embedding Vector Extraction

[0084] Deep semantic information can be extracted from the entity description text of the knowledge graph to capture the detailed characteristics, functions and associations of the entity with other entities. The entity description text can be converted into a dense vector through a pre-trained language information extraction model (such as BERT, GPTH or XLNet), which encodes the rich semantic information contained in the text.

[0085] The text embedding vector can be expressed as:

[0086] e text =LM(desc);

[0087] In the formula, desc is the entity description text, LM is the pre-trained language information extraction model, and e text It is the text vector obtained by extracting semantic information.

[0088] 1.2 Entity Embedding Vector Fusion

[0089] Combining the entities learned from graph structure embedding and the text vectors extracted from the language model, an attention mechanism (such as Transformer or its variants) is used to dynamically adjust the degree of fusion of the two information sources to optimize the effect of entity representation. The attention mechanism is used to capture long-distance dependencies, dynamically weigh the importance of information in the graph embedding vector and the text vector, and generate the final entity embedding vector.

[0090] The process of entity embedding vector fusion can be as follows:

[0091] e enhanced =softmax(W·[e graph ;e text ]+b)·[e graph ;e text ];

[0092] In the formula, e enhanced is the entity embedding vector after text enhancement, W is the learned weight, b is the bias term, and e text is the text embedding vector obtained by extracting semantic information, e graph is the graph embedding vector extracted by the graph structure embedding learning method, [e graph ;e text ] is the concatenation of the graph embedding vector and the text embedding vector extracted by the graph structure embedding learning method. Softmax is an activation function used to calculate the weight distribution of the two embeddings, thereby dynamically adjusting the contribution of each embedding.

[0093] The graph structure embedding is performed by analyzing the relationships between entities (such as dependencies or influences). Specific technologies may include graph neural networks (GNNs) or other graph embedding vector technologies, which can learn the embedding representation of entities from entities and their relationships (such as nodes and edges). At this time, the embedding of this entity in the graph structure is obtained. graph , which is usually based on the relationship of the entity to other related entities (such as cooling system components). graph and e text , through a trained weight matrix and bias term, dynamically adjust and fuse these two types of information, and finally obtain an enhanced entity embedding vector e that combines text description and graph structure enhanced , provides comprehensive information about "engine overheating" and its related components (such as cooling system, water pump).

[0094] 1.3 Loss Function

[0095] In the training process of the text-enhanced entity embedding layer network, the loss function L of the text-enhanced entity embedding layer network is minimized. e The goal is to train the model so that the prediction of the model is closer to the target value and the accuracy of the model prediction is improved, so that the entity embedding vector can find the optimal vector representation, so that the model can better capture and utilize these features. The loss function of the entity embedding vector is L e for:

[0096]

[0097] In the formula, N is the number of training samples, The entity embedding vector predicted by the text-enhanced entity embedding layer network during training, Embedding vector for the preset labeled target entity, ||·|| 2 is the square of the Euclidean distance.

[0098] 2. Multi-scale feature interaction layer

[0099] The purpose of the multi-scale feature interaction layer is to extract and fuse complex relational features between entities by applying convolution techniques at different scales. Through feature reorganization and multiple convolution operations, this layer can enhance the model's understanding of complex interactions between entities from different aspects of fault information. Feature fusion further integrates these multi-level information to form a comprehensive feature representation matrix, which provides support for more accurate prediction and classification. This is crucial for understanding and handling diverse and interrelated failure modes in automotive systems, and can help the system more comprehensively evaluate the impact of failures and thus develop more effective maintenance strategies.

[0100] 2.1 Feature Recombination

[0101] Taking every two different entity embedding vectors as a reorganization group, the elements of the two entity embedding vectors in the reorganization group are staggered and reorganized using the checkerboard reorganization method to form a new reorganized feature matrix M to enhance the interaction between entities and relationships. The staggered feature reorganization method for the two entity embedding vectors in each reorganization group is expressed as:

[0102] M=reshape(interleave(e1,e2));

[0103] Where M is the feature reorganization matrix obtained after feature reorganization, e1 is an entity embedding vector in the reorganized group, e2 is another entity embedding vector in the reorganized group, interleave(e1,e2) is a function that interleaves the elements of e1 and e2, and reshape represents a reshaping function that reshapes the interleaved vector into a matrix form.

[0104] The feature interactions between entities and relations are enhanced by staggered arrangement, so that the model can better capture complex fault relations.

[0105] 2.2 Multi-scale convolution extraction

[0106] By applying different types of convolution kernels and combining the residual network structure, a multi-scale residual convolution is formed to capture the multi-level features between entities and relations. The specific method of performing residual convolution feature extraction operations of different scales on the reorganized feature matrix through multi-scale residual convolution is as follows:

[0107] Each reorganized feature matrix is ​​processed by residual one-dimensional convolution, residual two-dimensional convolution and residual dilated convolution to obtain the convolution feature matrix of the corresponding scale; where:

[0108] Residual one-dimensional convolution, which is defined as: P1 = ReLU (M + Conv1D (M))

[0109] Where P1 is the convolution feature matrix obtained after one-dimensional convolution, Conv1D(M) is the one-dimensional convolution applied to the reorganized feature matrix M, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function;

[0110] Residual 3D convolution, defined as: P2 = ReLU(M + Conv3D(reshape(M)))

[0111] Where P2 is the convolution feature matrix obtained after three-dimensional convolution, Conv3D(reshape(M)) is the application of three-dimensional convolution to the reorganized feature matrix M, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function;

[0112] Residual dilated convolution, which is defined as: P3 = ReLU (M + DilatedConv (M, rate = 2))

[0113] Where P3 is the convolution feature matrix obtained after dilated convolution, DilatedConv(M,rate=2) is the dilated convolution applied to the reorganized feature matrix M, with a dilation rate of 2, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function.

[0114] At this time, convolution operations of different scales (such as one-dimensional, three-dimensional, and dilated convolutions) are applied to extract features at different levels, combined with a residual network structure to prevent information loss and gradient vanishing, thereby enhancing feature transfer and gradient flow.

[0115] 2.3 Feature Fusion

[0116] Perform feature fusion operation on residual convolution features of different scales to form a fusion feature representation matrix;

[0117] The feature fusion is defined as:

[0118] P final =FusionLayer(P1,P2,P3);

[0119] Where P final is the obtained fusion feature representation matrix, P1, P2, and P3 are the outputs of one-dimensional convolution, three-dimensional convolution, and dilated convolution, respectively. FusionLayer can be a feature fusion operation structure based on a multi-layer perceptron (MLP) or a self-attention mechanism to integrate features of various scales.

[0120] At this time, P final Provides a comprehensive perspective to understand the multi-level influencing factors of "engine overheating" failure, such as direct component failure, system-level design defects or improper operation.

[0121] 2.4 Feature Extraction

[0122] The fused features are further processed for feature extraction and further optimized to form the final feature representation matrix of the entity description text.

[0123] The feature extraction is defined as:

[0124] P extracted =ReLU(FC(P final ));

[0125] Where P extractedis the final feature representation matrix of the entity description text; FC is the fully connected layer, which is used to further transform the feature vector; ReLU is the activation function, which is used to introduce nonlinearity and enhance the expressiveness of the model.

[0126] At this time, P extracted The nonlinear expression capability of the model is emphasized, providing precise input for the dynamic prediction and completion layer, enabling the model to accurately predict and diagnose the specific causes and related impacts of “engine overheating”.

[0127] 2.5 Loss Function

[0128] In the training process of the multi-scale feature interaction layer network, the loss function L of the multi-scale feature interaction layer network is minimized. p The loss function L of the multi-scale feature interaction layer network is obtained by training. p for:

[0129] L p =βL feature +(1-β)L task ;

[0130] Where, L feature is the feature fidelity loss, L task is the task guidance loss; β is the loss adjustment hyperparameter, which is used to adjust the relative importance of the two losses.

[0131] Among them, the feature fidelity loss L feature It is used to ensure that the feature vector output by each convolutional layer maintains high quality and rich information. Its expression is:

[0132] Where m is the number of convolutional layers contained in the multi-scale residual convolution, P j is the output of the jth convolutional layer, To decode P through the decoder network j The reconstructed features obtained by decoding; α j is the reconstruction error weight of the jth convolutional layer, which can be adjusted according to the importance of the layer; MSE is the mean square error, which is used to measure the difference between the reconstructed features and the original features.

[0133] Task guidance loss L task To utilize the fused feature P final To guide the final task (e.g., predicting tail entities, head entities, or relations), the expression is:

[0134] Where N is the number of training samples, C is the number of fault cause categories, and y i,c is the true label of the training sample i in the fault cause category c (usually one-hot encoding), pi,c The probability that the multi-scale feature interaction layer network predicts that the training sample i belongs to the fault cause category c, and p i,c =softmax(P final ), which is used to characterize the probability that the entity failure event described in the entity description text belongs to a certain failure cause category.

[0135] 3. Dynamic prediction and completion layer

[0136] The main purpose of the dynamic prediction and completion layer is to use the complex feature representation matrix generated by the previous layer to dynamically predict and complete the entities and relationships in the knowledge graph. Through the probability score prediction function, this layer can not only predict the specific components and types of faults, but also identify and infer the potential relationships between faults, such as causes and effects. This dynamic prediction and completion function enables the knowledge graph to be continuously updated and expanded, improving the system's adaptability and response speed to new fault modes. This is very important for maintaining an accurate and timely fault diagnosis support system, ensuring the timeliness and accuracy of maintenance decisions, thereby improving the efficiency and reliability of the entire automotive maintenance system.

[0137] 3.1. Probability score prediction function

[0138] In the dynamic prediction and completion layer network, the calculated probability score prediction values ​​include:

[0139] The probability score prediction value φ of the tail entity: φ(P extracted ,t)=softmax(W t ·P extracted );

[0140] The probability score prediction value ψ of the head entity: ψ(P extracted ,h)=softmax(W h ·P extracted );

[0141] The probability score prediction value θ of the entity relationship: θ(P extracted ,r)=softmax(W r ·P extracted );

[0142] Where P extracted is the final feature representation matrix of the entity description text, t, h, r represent the tail entity, head entity and entity relationship respectively, W t , W h , W r They are the weight matrices of the tail entity, head entity, and entity relationship, respectively.

[0143] At this time, the probability score prediction value φ of the tail entity obtained represents the prediction probability of the possible tail entity (for example, a specific heat dissipation component failure, such as a radiator blockage); the probability score prediction value ψ of the head entity represents the prediction probability of the possible head entity (for example, "engine"); the probability score prediction value θ of the entity relationship represents the prediction probability of the possible relationship type (for example, "lead to", "influence").

[0144] 3.2 Training process and acquisition of weight matrix

[0145] A. Initialization

[0146] Before the dynamic prediction and completion layer network training begins, the weight matrix needs to be initialized. Random initialization methods are usually used, such as small random numbers, Xavier / Glorot initialization (keeping the input and output variances consistent) or He initialization (optimized for ReLU activation function), which can help prevent gradient problems in the early stages of training.

[0147] B. Forward Propagation

[0148] During the forward propagation of the model, the input feature vector P final is multiplied by the weight matrix and then possibly processed through an activation function such as softmax to produce the predicted output.

[0149] C. Loss function calculation

[0150] The loss function is used to evaluate the difference between the model output and the true label. The calculation result of the loss function can also be used to guide the reverse adversarial training update of the weight matrix in the future.

[0151] In the dynamic prediction and completion layer network, the weight matrices of the tail entity, head entity, and entity relationship are obtained by training the dynamic prediction and completion layer network with the goal of minimizing their respective loss functions. The respective loss functions are:

[0152] Probability score prediction loss function L for tail entity tail for:

[0153] In the formula, The model predicts that the head entity and the tail entity corresponding to the relationship of the sample are the true tail entities t i The probability of

[0154] The probability score prediction loss function L for the head entity head for:

[0155] In the formula, The model predicts that the head entity corresponding to the tail entity and relationship of the sample is the real head entity h i The probability of

[0156] Probability score prediction loss function L for entity relations relation for:

[0157] In the formula, The model predicts that the relationship between the head entity and the tail entity of a sample is the true relationship r i The probability of

[0158] E. Back-propagation adversarial training and iterative update of weight matrix

[0159] In the process of knowledge embedding for the automotive system fault knowledge graph, adversarial training is also performed on the dynamic prediction and completion layer network to update the weight matrix in the dynamic prediction and completion layer network. In this process, the gradient of the loss function with respect to the weight matrix is ​​calculated by the back propagation algorithm, which automatically calculates and propagates the gradient through the chain rule. Then, based on the calculated gradient, the weight matrix is ​​updated using the optimization algorithm. The optimization algorithm adjusts the weight according to the gradient direction and the set learning rate to reduce the prediction error.

[0160] The specific method is that in the process of adversarial training of the dynamic prediction and completion layer network, the gradient of each probability score prediction loss function with respect to the weight matrix is ​​calculated respectively, and the weight matrix is ​​iteratively updated until the number of iterations reaches the preset iteration number threshold, the training is terminated, and the update of the weight matrix is ​​completed; wherein, the update of the weight matrix in each iteration is expressed as:

[0161]

[0162] Where η is the learning rate of adversarial training; Represents the probability score prediction loss function L of the tail entity tail The weight matrix W of the tail entity t The gradient of Denotes the probability score prediction loss function L of the head entity head The weight matrix W of the head entity h The gradient of The probability score prediction loss function L representing the entity relationship relation About the weight matrix W of entity relations r gradient.

[0163] By utilizing the dynamic prediction and completion mechanism, real-time updating and expansion of knowledge can be achieved. It not only supports the updating of existing data, but also can dynamically add new fault instances and relationships, thereby maintaining the timeliness and accuracy of the system.

[0164] 4. Overview

[0165] In summary, the method of the present invention has the following technical advantages:

[0166] (1) The present invention obtains deep semantic information through semantic information extraction, uses graph structure embedding to analyze the relationship between entities, and uses the attention mechanism to dynamically fuse and adjust the entities learned from graph structure embedding with the text vectors extracted from the language model, thereby improving the accuracy and richness of entity representation. In addition, through in-depth semantic analysis, it helps the automotive system to better identify and classify fault types, thereby improving the accuracy and efficiency of fault diagnosis.

[0167] (2) The present invention enhances the feature interaction between entities and relationships by utilizing feature reorganization and staggered arrangement, so that the knowledge embedding model can better capture complex fault relationships. It also uses a multi-scale convolutional network and a residual convolutional network to extract features at different scales, effectively preventing information loss and gradient vanishing, enhancing feature transfer and gradient flow, and thus effectively enhancing the knowledge embedding model's ability to understand and represent complex relationships between entities.

[0168] (3) The present invention dynamically predicts and completes the missing entities and relationships in the knowledge graph through a dynamic prediction and completion layer, thereby adapting to the update and change of the knowledge graph, realizing continuous learning and improvement of the knowledge graph, ensuring the adaptability and response speed of fault diagnosis when facing unknown faults, and significantly improving the practicality and reliability of the automotive system fault knowledge graph.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described with reference to the preferred embodiments of the present invention, it should be understood by those skilled in the art that various changes may be made in form and details without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A knowledge embedding method for automobile system fault knowledge graph, characterized in that: The steps include: Obtaining entity description text in the knowledge graph of automobile system failure; the entity description text is used to describe entity and entity relationship, the entity includes a head entity and a tail entity, the head entity is a feature word describing the automobile system component, the tail entity is a feature word describing the failure event, and the entity relationship is a feature word describing the relationship between the automobile system component and the failure event; Using a pre-trained text-enhanced entity embedding layer network, semantic information is extracted from the entity description text to obtain a corresponding text embedding vector; using a graph structure embedding learning method, the embedded representation information of entities and entity relationships in the entity description text is extracted to obtain a corresponding graph embedding vector; then, using an attention mechanism, the text embedding vector and the graph embedding vector are dynamically weighed and fused to obtain a text-enhanced entity embedding vector; Using a pre-trained multi-scale feature interaction layer network, interleaved feature reorganization is performed on different entity embedding vectors to obtain a feature reorganization matrix, and residual convolution feature extraction operations of different scales are performed on the reorganized feature matrix through multi-scale residual convolution to extract convolution feature matrices of different scales; then, feature fusion and feature extraction are performed on the convolution feature matrices of different scales to obtain the final feature representation matrix of the entity description text; Using the pre-trained dynamic prediction and completion layer network, the probability score prediction values ​​of the entities and entity relationships in the entity description text are calculated according to the final feature representation matrix as the probability prediction results of completing the entities and entity relationships in the knowledge graph of automobile system failures.

2. The knowledge embedding method for automobile system fault knowledge graph according to claim 1 is characterized in that: In the text-enhanced entity embedding layer network, the text-enhanced entity embedding vector obtained is expressed as: and enhanced =softmax(W·[e graph ;and text ]+b)·[and graph ;and text ]; In the formula, e enhanced is the entity embedding vector after text enhancement, W is the learned weight, b is the bias term, and e text is the text embedding vector obtained by extracting semantic information, e graph is the graph embedding vector extracted by the graph structure embedding learning method, [e graph ;e text ] is the concatenation of the graph embedding vector and the text embedding vector extracted by the graph structure embedding learning method.

3. The knowledge embedding method for automobile system fault knowledge graph according to claim 2 is characterized in that: During the training process of the text-enhanced entity embedding layer network, the loss function L of the text-enhanced entity embedding layer network is minimized. e To train for the goal; The loss function L of the text-enhanced entity embedding layer network is e for: In the formula, N is the number of training samples, The entity embedding vector predicted by the text-enhanced entity embedding layer network during training, Embedding vector for the preset labeled target entity,||.|| 2 is the square of the Euclidean distance.

4. The knowledge embedding method for automobile system fault knowledge graph according to claim 1 is characterized in that: In the multi-scale feature interaction layer network, the specific method of performing interleaved feature reorganization on different entity embedding vectors is as follows: Taking every two different entity embedding vectors as a reorganization group, the elements of the two entity embedding vectors in the reorganization group are reorganized in an interlaced manner using the checkerboard reorganization method; the interlaced feature reorganization method for the two entity embedding vectors in each reorganization group is expressed as: M=reshape(interleave(e1,e2)); Where M is the feature reorganization matrix obtained after feature reorganization, e1 is an entity embedding vector in the reorganization group, e2 is another entity embedding vector in the reorganization group, interleave(e1, e2) is a function that interleaves the elements of e1 and e2, and reshape represents a reshaping function that reshapes the interleaved vector into a matrix form.

5. The knowledge embedding method for automobile system fault knowledge graph according to claim 1 is characterized in that: In the multi-scale feature interaction layer network, the specific method of performing residual convolution feature extraction operations of different scales on the reorganized feature matrix through multi-scale residual convolution is: Each reorganized feature matrix is ​​processed by residual one-dimensional convolution, residual two-dimensional convolution and residual dilated convolution to obtain the convolution feature matrix of the corresponding scale; where: Residual one-dimensional convolution, which is defined as: P1 = ReLU (M + Conv1D (M)) Where P1 is the convolution feature matrix obtained after one-dimensional convolution, Conv1D(M) is the one-dimensional convolution applied to the reorganized feature matrix M, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function; Residual 3D convolution, defined as: P2 = ReLU(M + Conv3D(reshape(M))) Where P2 is the convolution feature matrix obtained after three-dimensional convolution, Conv3D(reshape(M)) is the application of three-dimensional convolution to the reorganized feature matrix M, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function; Residual dilated convolution, which is defined as: P3 = ReLU (M + DilatedConv (M, rate = 2)) Where P3 is the convolution feature matrix obtained after dilated convolution, DilatedConv(M, rate = 2) is the dilated convolution applied to the reorganized feature matrix M, the dilation rate is 2, the residual connection directly adds the reorganized feature matrix M to the convolution output, and ReLU is the activation function.

6. The knowledge embedding method for automobile system fault knowledge graph according to claim 5 is characterized in that: In the multi-scale feature interaction layer network, the specific method of performing feature fusion and feature extraction on the convolution feature matrices of different scales is as follows: Perform feature fusion operation on residual convolution features of different scales: P final =FusionLayer(P1,P2,P3); Where P final is the obtained fusion feature representation matrix, P1, P2, P3 are the outputs of residual one-dimensional convolution, residual three-dimensional convolution and residual dilated convolution respectively, and FusionLayer represents the feature fusion operation based on multi-layer perceptron or self-attention mechanism; Then the fusion feature representation matrix P final Perform further feature extraction: P extracted =ReLU(FC(P final )); Where P extracted The final feature representation matrix of the obtained entity description text; FC represents the fully connected layer and ReLU is the activation function.

7. The knowledge embedding method for automobile system fault knowledge graph according to claim 6 is characterized in that: During the training process of the multi-scale feature interaction layer network, the loss function L of the multi-scale feature interaction layer network is minimized. p The loss function L of the multi-scale feature interaction layer network is obtained by training the target; p for: L p =βL feature +(1-β)L task ; Where, L feature is the feature fidelity loss, L task is the task guidance loss; β is the loss adjustment hyperparameter; where: Where m is the number of convolutional layers contained in the multi-scale residual convolution, P j is the output of the jth convolutional layer, To decode P through the decoder network j The reconstructed features obtained by decoding, α j is the reconstruction error weight of the jth convolutional layer, and MSE represents the mean square error operation; Where N is the number of training samples, C is the number of fault cause categories, and y i,c is the true label of training sample i in fault cause category c, P i,c The probability that the multi-scale feature interaction layer network predicts that the training sample i belongs to the fault cause category c, and p i,c =softmax(P final ).

8. The knowledge embedding method for automobile system fault knowledge graph according to claim 1 is characterized in that: In the dynamic prediction and completion layer network, the calculated probability score prediction values ​​include: The probability score prediction value φ of the tail entity: φ(P extracted , t) = softmax(W t ·P extracted ); The probability score prediction value ψ of the head entity: ψ(P extracted , h)=softmax(W h ·P extracted ); The predicted probability score of the entity relationship θ: θ(P extracted , r)=softmax(W r ·P extracted ); Where P extracted is the final feature representation matrix of the entity description text, t, h, r represent the tail entity, head entity and entity relationship respectively, W t , W h , W r They are the weight matrices of the tail entity, head entity, and entity relationship, respectively.

9. The knowledge embedding method for automobile system fault knowledge graph according to claim 8 is characterized in that: In the dynamic prediction and completion layer network, the weight matrices of the tail entity, the head entity and the entity relationship are obtained by training the dynamic prediction and completion layer network with the goal of minimizing their respective loss functions. The respective loss functions are: Probability score prediction loss function L for tail entity tail for: In the formula, The model predicts that the head entity and the tail entity corresponding to the relationship of the sample are the true tail entities t i The probability of The probability score prediction loss function L for the head entity head for: In the formula, The model predicts that the head entity corresponding to the tail entity and relationship of the sample is the real head entity h i The probability of Probability score prediction loss function L for entity relations relation for: In the formula, is the probability that the model predicts that the relationship between the head entity and the tail entity of a sample is the true relationship ri, and 10. The knowledge embedding method for automobile system fault knowledge graph according to claim 9 is characterized in that: In the process of knowledge embedding for the automobile system fault knowledge graph, adversarial training is also performed on the dynamic prediction and completion layer network to update the weight matrix in the dynamic prediction and completion layer network; the specific method is as follows: In the process of adversarial training of the dynamic prediction and completion layer network, the gradient of each probability score prediction loss function with respect to the weight matrix is ​​calculated respectively, and the weight matrix is ​​iteratively updated until the number of iterations reaches the preset iteration number threshold, the training is terminated, and the update of the weight matrix is ​​completed; wherein, the update of the weight matrix in each iteration is expressed as: Where η is the learning rate of adversarial training; Represents the probability score prediction loss function L of the tail entity tail The weight matrix W of the tail entity t The gradient of Denotes the probability score prediction loss function L of the head entity head The weight matrix W of the head entity h The gradient of The probability score prediction loss function L representing the entity relationship relation The weight matrix w of entity relations r gradient.

Citation Information

Patent Citations

  • Satellite fault diagnosis method, device and equipment based on graph embedding and storage medium

    CN117851612A

  • Risk prediction method and apparatus, and device and storage medium

    WO2023065545A1

Cited By

  • Knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework

    CN120067341A