Fault diagnosis method and system for railway signal equipment

Abstract

The invention provides a fault diagnosis method and system for railway signal equipment, and relates to the technical field of fault diagnosis, and the method comprises the steps: obtaining multi-source heterogeneous data of the railway signal equipment, and the multi-source heterogeneous data comprises equipment operation parameters, environmental factors and maintenance records; carrying out fusion processing on the multi-source heterogeneous data by utilizing a factorization machine to obtain a fused data set; based on the fused data set, constructing a railway signal equipment knowledge graph by using an ontology construction method and a knowledge graph technology; based on the knowledge graph of the railway signal equipment, obtaining the correlation between the fault root cause of the railway signal equipment and the equipment through knowledge reasoning, and generating a fault diagnosis rule base; according to the fault diagnosis rule base and the fusion data set, constructing a railway signal equipment fault prediction model; and inputting real-time data of the railway signal equipment into the railway signal equipment fault prediction model, and outputting a fault diagnosis result and fault cause analysis of the railway signal equipment.

Description

Technical Field

[0001] This invention relates to fault diagnosis technology, and more particularly to fault diagnosis methods and systems for railway signaling equipment. Background Technology

[0002] Due to the wide variety, complex structure, and harsh working environment of railway signaling equipment, equipment failures occur frequently, posing safety hazards and economic losses to railway operations. Therefore, timely and accurate diagnosis of railway signaling equipment failures is of great significance for improving equipment availability and ensuring safe train operation.

[0003] Traditional fault diagnosis of railway signaling equipment mainly relies on manual experience and periodic maintenance, which suffers from low diagnostic efficiency, low accuracy, and over-maintenance. With the increasing complexity and intelligence of railway signaling equipment, there is an urgent need to introduce advanced fault diagnosis technologies to achieve real-time monitoring and intelligent analysis of equipment status. Summary of the Invention

[0004] The present invention provides a fault diagnosis method and system for railway signaling equipment, which can at least solve some of the problems in the prior art.

[0005] A first aspect of the present invention,

[0006] Provide fault diagnosis methods for railway signaling equipment, including:

[0007] Acquire multi-source heterogeneous data of railway signaling equipment, including equipment operating parameters, environmental factors, and maintenance records; wherein, the equipment operating parameters include the equipment's voltage, current, and temperature parameters, the environmental factors include the temperature, humidity, and vibration parameters of the environment in which the equipment is located, and the maintenance records include the equipment's historical maintenance data and the operation records of maintenance personnel;

[0008] The multi-source heterogeneous data is fused using a factorization machine to obtain a fused dataset. Based on the fused dataset, a railway signaling equipment knowledge graph is constructed using ontology construction methods and knowledge graph technology. The railway signaling equipment knowledge graph includes the component composition, working principle, and fault cases of railway signaling equipment.

[0009] Based on the knowledge graph of the railway signaling equipment, the root causes of the faults of the railway signaling equipment and the correlation between the equipment are obtained through knowledge reasoning, and a fault diagnosis rule base is generated. According to the fault diagnosis rule base and the fused dataset, a fault prediction model of the railway signaling equipment is constructed. Real-time data of the railway signaling equipment is input into the fault prediction model of the railway signaling equipment, and the fault diagnosis results and fault cause analysis of the railway signaling equipment are output.

[0010] In one alternative implementation,

[0011] The multi-source heterogeneous data is fused using a factorization machine to obtain a fused dataset, including:

[0012] Multi-source heterogeneous sample data is input into an initial factorization machine model. For each feature of the multi-source heterogeneous sample data, the features of the multi-source heterogeneous sample data are mapped to a latent space of a preset dimension through latent vectors to obtain latent vector representations. The second-order interaction between features in the multi-source heterogeneous data is modeled by calculating the inner product of the latent vector representations to obtain interaction features.

[0013] Based on the latent vector representation and the interaction features, the predicted value of the initial factorization machine model is calculated. The predicted value consists of a global bias term, a first-order linear term, and a second-order interaction term. Taking maximizing the difference between the predicted values ​​of positive and negative samples output by the initial factorization machine model as the objective function, the parameters of the initial factorization machine model are trained and updated using the gradient descent algorithm to obtain the trained factorization machine model.

[0014] The multi-source heterogeneous data is fused using a trained factorization machine model to obtain a fused feature vector, which represents information about equipment operating parameters, environmental factors, and maintenance records. A fused dataset of railway signaling equipment is constructed based on the fused feature vector, in which each sample contains the fused feature vector and the corresponding fault label data.

[0015] In one alternative implementation,

[0016] Based on the fused dataset, a railway signaling equipment knowledge graph is constructed using ontology construction methods and knowledge graph technology, including:

[0017] Based on the fused dataset, an ontology construction method is used to define the top-level concepts of the railway signaling equipment knowledge graph. The top-level concepts include equipment concepts, component concepts, fault type concepts, fault cause concepts, and fault repair concepts. The equipment concepts include signal lights, track circuits, and switch machines. The component concepts include light-emitting units, relays, and motors. The fault type concepts include signal light malfunction and switching failure. The fault cause concepts include component aging and environmental interference. The repair scheme concepts include component replacement and parameter adjustment.

[0018] Determine the hierarchical and semantic relationships between different top-level concepts. The hierarchical relationships include the subordinate relationship between the component concept and the equipment concept, the relationship between the fault type concept and the component concept and the equipment concept respectively, and the relationship between the maintenance plan concept and the fault type concept and the fault cause concept respectively. The semantic relationships include the causal relationships between fault and equipment, fault and fault cause, and fault and maintenance plan.

[0019] Named entity recognition technology is used to extract instances of equipment, components, faults, causes, and maintenance solutions from the fused dataset, and these instances are mapped to knowledge graph nodes based on concept hierarchy definitions. The fused dataset is traversed to extract attribute information of node instances and add it as attributes of knowledge graph nodes. Relation extraction technology is used to mine semantic associations between instances in the fused dataset, construct relation edges between instances, and complete the remaining relation edges in the knowledge graph through ontology reasoning mechanism to form a knowledge network of node associations. The railway signaling equipment knowledge graph is represented in ontology description language form, and instance nodes and instance edges are stored in triple form.

[0020] In one alternative implementation,

[0021] Based on the railway signaling equipment knowledge graph, the root causes of faults in the railway signaling equipment and the relationships between equipment are obtained through knowledge reasoning, generating a fault diagnosis rule base, including:

[0022] The fault tracing rules and device association impact rules are defined using a data log rule language. Horn clauses are used to represent the recursive causal chain between fault causes and the transitive relationship of physical connections and logical dependencies between devices. Logical programming methods are used to reason about the root causes of faults. The data log rule consists of a header and a body. The header is an atomic formula and the body is the conjunction of the atomic formulas. The Horn clauses are logical clauses.

[0023] For each type of fault node in the knowledge graph, a bottom-up reasoning strategy based on depth-first search is adopted. By matching and executing fault tracing rules, the root cause of the fault is obtained. The device association and impact rules are recursively called to determine the associated devices of the fault. Based on the root cause of the fault and the associated devices, common root cause patterns and association transmission paths of different fault types are extracted to construct a fault diagnosis decision tree. Through the reverse diagnosis path from the leaf node to the root node, structured diagnosis rules are generated. The structured diagnosis path includes fault phenomena, intermediate causes, root causes, and affected devices.

[0024] Using machine learning toolkits, correlation patterns between faults and causes, equipment, and monitoring quantities are mined from historical fault case data to generate data-driven diagnostic rules. These data-driven diagnostic rules include fault conditions, causes, affected equipment, and key equipment. Structured diagnostic paths and data-driven diagnostic rules are integrated to obtain fault diagnosis rules. The credibility of multiple fault diagnosis rules is ranked based on evidence weights to construct a fault diagnosis rule library.

[0025] In one alternative implementation,

[0026] Based on the fault diagnosis rule base and the fused dataset, a fault prediction model for railway signaling equipment is constructed, including:

[0027] Using natural language processing technology, the fault diagnosis rules in the fault diagnosis rule base are transformed into vector representations. The semantic similarity of fault concepts is learned through word embedding. Fault symptoms and monitoring condition elements in the fault diagnosis rules are mapped into semantic vectors. Through vector operations, a fault representation vector that integrates the semantic features of the diagnosis rules is constructed.

[0028] The fusion feature vectors and fault label data in the fusion dataset are balanced. Based on the sample distribution characteristics of different fault modes, an adaptive synthetic minority class oversampling technique is used to enhance the fault samples. The fault duration is calculated based on the sample timestamp and used as the time attribute of the synthetic sample to construct time series monitoring data.

[0029] Based on time series monitoring data, a gated cyclic unit network is used to model the time series feature vector. The fault representation vector and the time series feature vector are concatenated to construct a feature space containing rule knowledge and monitoring data. A weighted cross-entropy loss function is used to adaptively adjust the penalty weight according to the number of samples of each fault category, and an early stop method is introduced to control the number of iterations to obtain a fault prediction model for railway signaling equipment.

[0030] In one alternative implementation,

[0031] Based on time series monitoring data, a gated recurrent unit network is used to model time series feature vectors, including:

[0032] A gated recurrent unit network is used to perform sequence modeling on time series monitoring data. The time series monitoring data at each time step is mapped to a real-valued dense vector through an embedding layer, which serves as the input of the gated recurrent unit. In the gated recurrent unit, the input vector of the current time step and the hidden state of the previous time step are adaptively weighted using update gate and reset gate. The update gate controls the proportion of the hidden state information of the previous time step retained in the current time step, and the reset gate controls the proportion of the input vector of the current time step incorporated into the current hidden state.

[0033] By updating the gate and resetting the gate to aggregate the sequence dependencies between time steps, long-term context information in the time series monitoring data is determined; the steps of using the gated recurrent unit network to perform sequence modeling on the time series monitoring data and subsequent steps are repeated until the last time step of the time series monitoring data, to obtain the hidden state sequence of each time step of the gated recurrent unit network, which represents the dynamic feature information of the time series monitoring data at different time scales;

[0034] In a gated recurrent unit network, an attention query vector is generated using the hidden state of the last time step. The attention query vector is an aggregated representation of global context information. The similarity between the attention query vector and the hidden states of each time step in the gated recurrent unit network is calculated using an attention scoring function to obtain attention weights that represent the relevance of the hidden states of each time step to the current sequence modeling task. The hidden states of each time step in the gated recurrent unit network are linearly combined according to the attention weights to obtain a temporal feature vector that integrates global temporal features.

[0035] In one alternative implementation,

[0036] Also includes:

[0037] The correlation propagation algorithm is applied layer by layer to the fault prediction model of the railway signaling equipment. The correlation weight of each dimension of the real-time data of the railway signaling equipment to the fault diagnosis result is calculated through back propagation. For each layer of the gated cyclic unit network, the connection weight of each dimension of the real-time data of the railway signaling equipment to the neuron and the correlation score of the neuron to the fault diagnosis result are calculated.

[0038] The correlation weights of each dimension of the real-time data of the railway signaling equipment to the prediction category are calculated according to the layer-by-layer correlation propagation algorithm until the back propagation of all layers is completed, so as to obtain the correlation weight distribution of each time step of the input time series to the fault prediction result; the correlation weight distribution is normalized and visualized to determine the correlation weight visualization result.

[0039] The Shapley additive interpretation algorithm is applied to the railway signaling equipment fault prediction model. This involves calculating the marginal contribution of the model's input features to the prediction function. For each input feature, all feature subsets are enumerated, the marginal contribution of that feature within the subset is calculated, and the Shapley value of that feature is obtained according to the Shapley addition formula. This yields the distribution results of the Shapley values ​​for key features under different fault modes. The visualization results of the correlation weights and the Shapley value distribution results are then used to obtain the explanation of the prediction mechanism of the railway signaling equipment fault prediction model.

[0040] A second aspect of the present invention,

[0041] Provides a fault diagnosis system for railway signaling equipment, including:

[0042] The first unit is used to acquire multi-source heterogeneous data of railway signaling equipment. The multi-source heterogeneous data includes equipment operating parameters, environmental factors, and maintenance records. The equipment operating parameters include the voltage, current, and temperature parameters of the equipment. The environmental factors include the temperature, humidity, and vibration parameters of the environment in which the equipment is located. The maintenance records include the historical maintenance data of the equipment and the operation records of maintenance personnel.

[0043] The second unit is used to fuse the multi-source heterogeneous data using a factorization machine to obtain a fused dataset; based on the fused dataset, an ontology construction method and knowledge graph technology are used to construct a railway signaling equipment knowledge graph, which includes the component composition, working principle and fault cases of railway signaling equipment.

[0044] The third unit is used to obtain the root causes of faults in the railway signaling equipment and the correlation between equipment through knowledge reasoning based on the knowledge graph of the railway signaling equipment, and generate a fault diagnosis rule base; construct a fault prediction model for railway signaling equipment based on the fault diagnosis rule base and the fused dataset; input the real-time data of the railway signaling equipment into the fault prediction model for railway signaling equipment, and output the fault diagnosis results and fault cause analysis of the railway signaling equipment.

[0045] A third aspect of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] Memory used to store processor-executable instructions;

[0049] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0050] Fourth aspect of the embodiments of the present invention,

[0051] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0052] This invention fully utilizes multi-source heterogeneous data from railway signaling equipment, including equipment operating parameters, environmental factors, and maintenance records. Through factorization, it fuses the data, effectively integrating equipment status information and providing comprehensive and reliable data support for subsequent fault diagnosis. Furthermore, it introduces knowledge graph technology to construct a domain knowledge base containing the component composition, working principles, and fault cases of railway signaling equipment. This system organizes and represents expert knowledge regarding equipment structure, behavior, and fault mechanisms, providing rich semantic information and reasoning basis for fault diagnosis.

[0053] Based on a knowledge graph of railway signaling equipment, this system uses knowledge reasoning to obtain the root causes of equipment failures and the relationships between equipment, automatically generating a fault diagnosis rule base. This formalizes domain knowledge and diagnostic experience, improving the intelligence and interpretability of fault diagnosis. Combining the fault diagnosis rule base and a multi-source fusion dataset, a fault prediction model for railway signaling equipment is constructed. Leveraging the feature learning and modeling capabilities of deep learning technology, it mines fault patterns and evolution laws from equipment operation data, enabling early warning and timely diagnosis of equipment failures. Inputting real-time data from railway signaling equipment into the fault prediction model enables real-time monitoring of equipment status and fault diagnosis, while also providing fault cause analysis. This provides an intuitive and actionable basis for equipment maintenance decisions, significantly improving the efficiency and accuracy of fault diagnosis. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the fault diagnosis method for railway signaling equipment according to an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the fault diagnosis system for railway signaling equipment according to an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1 This is a flowchart illustrating the fault diagnosis method for railway signaling equipment according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0059] S101. Acquire multi-source heterogeneous data of railway signaling equipment, wherein the multi-source heterogeneous data includes equipment operating parameters, environmental factors, and maintenance records; wherein the equipment operating parameters include the voltage, current, and temperature parameters of the equipment, the environmental factors include the temperature, humidity, and vibration parameters of the environment in which the equipment is located, and the maintenance records include the historical maintenance data of the equipment and the operation records of maintenance personnel;

[0060] In one alternative implementation,

[0061] The multi-source heterogeneous data is fused using a factorization machine to obtain a fused dataset, including:

[0062] Multi-source heterogeneous sample data is input into an initial factorization machine model. For each feature of the multi-source heterogeneous sample data, the features of the multi-source heterogeneous sample data are mapped to a latent space of a preset dimension through latent vectors to obtain latent vector representations. The second-order interaction between features in the multi-source heterogeneous data is modeled by calculating the inner product of the latent vector representations to obtain interaction features.

[0063] Based on the latent vector representation and the interaction features, the predicted value of the initial factorization machine model is calculated. The predicted value consists of a global bias term, a first-order linear term, and a second-order interaction term. Taking maximizing the difference between the predicted values ​​of positive and negative samples output by the initial factorization machine model as the objective function, the parameters of the initial factorization machine model are trained and updated using the gradient descent algorithm to obtain the trained factorization machine model.

[0064] The multi-source heterogeneous data is fused using a trained factorization machine model to obtain a fused feature vector, which represents information about equipment operating parameters, environmental factors, and maintenance records. A fused dataset of railway signaling equipment is constructed based on the fused feature vector, in which each sample contains the fused feature vector and the corresponding fault label data.

[0065] For example, fault diagnosis of railway signaling equipment requires comprehensive consideration of multiple influencing factors. These factors have wide-ranging and diverse data sources, forming multi-source heterogeneous data. The multi-source heterogeneous data studied in this paper mainly includes three categories: equipment operating parameters, environmental factors, and maintenance records. Among them, equipment operating parameters reflect the real-time operating status of the equipment, such as voltage, current, and temperature; environmental factors describe the characteristics of the environment in which the equipment is located, such as temperature, humidity, and vibration; and maintenance records provide historical maintenance information and manual operation records of the equipment.

[0066] Collect multi-source heterogeneous data from railway signaling equipment, including equipment operating parameters, environmental factors, maintenance records, and corresponding fault tag data. Preprocess the collected data, including data cleaning and feature extraction.

[0067] The structure of the factorization machine model is defined, including an input layer, an embedding layer, a feature interaction layer, and an output layer. For each feature of the multi-source heterogeneous sample data, the embedding layer maps the features to a latent space of a preset dimension (e.g., k-dimensional), obtaining a latent vector representation. The second-order interaction between features in the multi-source heterogeneous data is modeled by calculating the inner product of the latent vector representations, resulting in the interaction features.

[0068] Based on the latent vector representation and interaction features, the predicted values ​​of the factorization machine model are calculated. The predicted values ​​consist of a global bias term, a first-order linear term, and a second-order interaction term, as shown in the following formula:

[0069]

[0070] in, This represents the predicted value from the factorization machine model, where w0 is the global bias term, and w i Let i be the weight of the i-th feature. <v i v j > is the inner product of the latent vectors of the i-th feature and the j-th feature, x i x j Let i and j represent the i-th feature and the j-th feature, respectively, and n represent the number of features.

[0071] Construct a training dataset, generating a corresponding positive and negative sample pair for each sample. Define the objective function as maximizing the difference between the predicted values ​​of the positive and negative samples:

[0072]

[0073] Where S is the training sample set, x + and x - Positive and negative samples, respectively Sample prediction function.

[0074] The objective function is optimized using the gradient descent algorithm, and the model parameters are iteratively updated until convergence, resulting in the trained factorization machine model.

[0075] A pre-trained factorization machine model is used to fuse multi-source heterogeneous data. The features of each sample are input into the factorization machine model, and the fused feature vector representation is obtained through an embedding layer and a feature interaction layer. The fused feature vector contains information such as equipment operating parameters, environmental factors, and maintenance records. A fused dataset for railway signaling equipment is constructed based on the fused feature vector. Each sample in the fused dataset contains a fused feature vector and corresponding fault label data.

[0076] Fusion datasets can be used to train fault diagnosis models, fault prediction models, and other technologies to improve the intelligent operation and maintenance of railway signaling equipment. They can also be used for data analysis and visualization to uncover patterns and trends within the data.

[0077] This solution effectively integrates data from different sources and of different types through a factorization machine model, resulting in a fused feature vector and a fused dataset containing rich information, laying a data foundation for subsequent intelligent operation and maintenance applications.

[0078] S102. The multi-source heterogeneous data is fused using a factorization machine to obtain a fused dataset; based on the fused dataset, a railway signaling equipment knowledge graph is constructed using ontology construction methods and knowledge graph technology, the railway signaling equipment knowledge graph containing the component composition, working principle, and fault cases of railway signaling equipment;

[0079] In one alternative implementation,

[0080] Based on the fused dataset, a railway signaling equipment knowledge graph is constructed using ontology construction methods and knowledge graph technology, including:

[0081] Based on the fused dataset, an ontology construction method is used to define the top-level concepts of the railway signaling equipment knowledge graph. The top-level concepts include equipment concepts, component concepts, fault type concepts, fault cause concepts, and fault repair concepts. The equipment concepts include signal lights, track circuits, and switch machines. The component concepts include light-emitting units, relays, and motors. The fault type concepts include signal light malfunction and switching failure. The fault cause concepts include component aging and environmental interference. The repair scheme concepts include component replacement and parameter adjustment.

[0082] Determine the hierarchical and semantic relationships between different top-level concepts. The hierarchical relationships include the subordinate relationship between the component concept and the equipment concept, the relationship between the fault type concept and the component concept and the equipment concept respectively, and the relationship between the maintenance plan concept and the fault type concept and the fault cause concept respectively. The semantic relationships include the causal relationships between fault and equipment, fault and fault cause, and fault and maintenance plan.

[0083] Named entity recognition technology is used to extract instances of equipment, components, faults, causes, and maintenance solutions from the fused dataset, and these instances are mapped to knowledge graph nodes based on concept hierarchy definitions. The fused dataset is traversed to extract attribute information of node instances and add it as attributes of knowledge graph nodes. Relation extraction technology is used to mine semantic associations between instances in the fused dataset, construct relation edges between instances, and complete the remaining relation edges in the knowledge graph through ontology reasoning mechanism to form a knowledge network of node associations. The railway signaling equipment knowledge graph is represented in ontology description language form, and instance nodes and instance edges are stored in triple form.

[0084] For example, railway signaling equipment involves multiple professional fields and includes complex structural hierarchies and operating mechanisms. In order to systematically apply domain knowledge to the fault diagnosis process, this application adopts ontology construction methods and knowledge graph technology to formally represent and semantically organize the concepts, instances and their associations of railway signaling equipment.

[0085] First, based on the characteristics of railway signaling equipment, the top-level concepts of the knowledge graph are defined. Through analysis of domain expert knowledge and relevant standards and specifications, the top-level concepts are divided into five categories: equipment concepts, component concepts, fault type concepts, fault cause concepts, and fault repair concepts. Specifically, equipment concepts include signaling equipment types such as signal lights, track circuits, and switch machines; component concepts include equipment components such as light-emitting units, relays, and motors; fault type concepts include common fault phenomena such as signal light malfunction and switching failures; fault cause concepts include factors that cause faults such as component aging and environmental interference; and repair solution concepts include fault handling measures such as component replacement and parameter adjustment.

[0086] Building upon the definition of top-level concepts, the hierarchical and semantic relationships between different concepts are further determined. Hierarchical relationships reflect the subordinate and inclusive relationships between concepts, primarily including: the subordinate relationship between the component concept and the equipment concept, i.e., a component is a constituent unit of equipment; the association between the fault type concept and both the component and equipment concepts, i.e., a specific fault phenomenon is associated with a corresponding component or equipment; and the association between the maintenance plan concept and both the fault type and fault cause concepts, i.e., specific maintenance measures target specific fault phenomena and causes. Semantic relationships reflect the causal and dependent logical relationships between different concepts, primarily including: the causal relationship between fault and equipment, fault and fault cause, and fault and maintenance plan, i.e., a specific fault in a piece of equipment caused by a specific reason requires corresponding maintenance measures.

[0087] After constructing the conceptual hierarchy of the knowledge graph, it is necessary to extract and fuse specific instance information from multi-source heterogeneous data. The fused dataset provides rich equipment status data, environmental monitoring data, and operation and maintenance record data, containing a large amount of knowledge such as equipment instances, fault cases, and handling experience. This application adopts named entity recognition technology, and by constructing a domain dictionary and feature templates, identifies and extracts key entities such as equipment, components, faults, causes, and maintenance plans from the text fields of the fused dataset, and maps them to corresponding nodes in the knowledge graph according to the conceptual hierarchy definition. To enrich the node information, the fused dataset is traversed to extract the attribute information of each entity, such as equipment model parameters, fault occurrence time, and maintenance plan operation steps, and these are added to the attribute list of the knowledge graph nodes.

[0088] After constructing instance nodes, it is necessary to mine the relationships between different instances and construct instance edges that reflect their semantic relationships. Traditional relation extraction methods mostly adopt rule-based or statistical approaches, which have insufficient generalization performance. In recent years, relation extraction technology based on deep learning has made significant progress. This application uses pre-trained language models such as BERT and sequence labeling models such as BiLSTM-CRF to construct a relation extractor for the railway signaling equipment domain. Taking sentences from the fused dataset as input, the semantic relationships between entity pairs in the sentences are extracted to form directed edges between instances. At the same time, using an ontology reasoning mechanism, based on the relation constraints defined at the concept level, unmentioned relation edges in the knowledge graph are supplemented to form a logically closed knowledge network.

[0089] Finally, to facilitate the storage, retrieval, and application of knowledge graphs, they need to be formally represented. This application uses the Web Ontology Language (OWL) published by the W3C organization, describing concepts, instances, attributes, and relationships in the knowledge graph in the form of RDF (Resource Description Framework) triples. Each triple consists of three parts: a subject, a predicate, and an object, corresponding to nodes, edges, and nodes in the knowledge graph, respectively. Through the set of triples, the structure and semantic information of the knowledge graph can be fully expressed. The represented knowledge graph is then stored in a graph database such as Neo4j to achieve persistent knowledge storage and efficient retrieval.

[0090] This application constructs a railway signaling equipment knowledge graph, with equipment, components, faults, causes, and maintenance plans as core concepts. It integrates multi-source heterogeneous data to extract and organize equipment instances and the relationships between them. On one hand, the knowledge graph, in a structured form, preserves knowledge in the railway signaling equipment domain, providing comprehensive and rich background information for fault diagnosis. On the other hand, the hierarchical semantic relationships between concepts and the causal logical relationships between instances in the knowledge graph provide a foundation for fault tracing and knowledge reasoning. This application uses an ontology construction method to define the top-level framework of the knowledge graph, utilizes named entity recognition and relation extraction technologies to obtain specific instances from the data, and represents and stores them using an ontology language and graph database. This constructs a railway signaling equipment domain knowledge base that is universal, semantic, and scalable, laying the foundation for subsequent knowledge reasoning and intelligent diagnosis.

[0091] S103. Based on the railway signaling equipment knowledge graph, obtain the root causes of the railway signaling equipment failures and the associations between the equipment through knowledge reasoning, and generate a fault diagnosis rule base; construct a railway signaling equipment failure prediction model according to the fault diagnosis rule base and the fused dataset; input the real-time data of the railway signaling equipment into the railway signaling equipment failure prediction model, and output the fault diagnosis results and fault cause analysis of the railway signaling equipment.

[0092] In one alternative implementation,

[0093] Based on the railway signaling equipment knowledge graph, the root causes of faults in the railway signaling equipment and the relationships between equipment are obtained through knowledge reasoning, generating a fault diagnosis rule base, including:

[0094] The fault tracing rules and device association impact rules are defined using a data log rule language. Horn clauses are used to represent the recursive causal chain between fault causes and the transitive relationship of physical connections and logical dependencies between devices. Logical programming methods are used to reason about the root causes of faults. The data log rule consists of a header and a body. The header is an atomic formula and the body is the conjunction of the atomic formulas. The Horn clauses are logical clauses.

[0095] For each type of fault node in the knowledge graph, a bottom-up reasoning strategy based on depth-first search is adopted. By matching and executing fault tracing rules, the root cause of the fault is obtained. The device association and impact rules are recursively called to determine the associated devices of the fault. Based on the root cause of the fault and the associated devices, common root cause patterns and association transmission paths of different fault types are extracted to construct a fault diagnosis decision tree. Through the reverse diagnosis path from the leaf node to the root node, structured diagnosis rules are generated. The structured diagnosis path includes fault phenomena, intermediate causes, root causes, and affected devices.

[0096] Using machine learning toolkits, correlation patterns between faults and causes, equipment, and monitoring quantities are mined from historical fault case data to generate data-driven diagnostic rules. These data-driven diagnostic rules include fault conditions, causes, affected equipment, and key equipment. Structured diagnostic paths and data-driven diagnostic rules are integrated to obtain fault diagnosis rules. The credibility of multiple fault diagnosis rules is ranked based on evidence weights to construct a fault diagnosis rule library.

[0097] For example, based on the construction of a knowledge graph of railway signaling equipment, this application further utilizes knowledge reasoning technology to explore the root causes and related effects of equipment failures, and automatically generates a structured fault diagnosis rule base to support intelligent diagnosis decision-making.

[0098] Traditional fault diagnosis often employs rule-based or case-based methods, relying on domain expert experience for knowledge acquisition and representation. Furthermore, diagnostic rules are typically flat and lack in-depth characterization of fault mechanisms. In recent years, knowledge graph reasoning technology has rapidly developed, enabling the modeling of the hierarchy and interrelationships of domain knowledge and the acquisition of implicit knowledge through chain-like reasoning. This application employs a reasoning method based on Datalog rule language, utilizing semantic relationships within the railway signaling equipment knowledge graph to define fault tracing rules and equipment association influence rules, thereby mining causal and relational chains of faults.

[0099] Data log rule language is a declarative logic programming language that represents knowledge and rules through predicate logic. This application uses data log rules to define inference rules for fault tracing and correlation propagation, formally represented through Horn clauses. A Horn clause is a special type of logical clause consisting of a head and a body. The head is a single atomic formula, and the body is the conjunction of atomic formulas. Fault tracing rules use multiple Horn clauses to characterize the recursive causal relationship between fault causes, such as "equipment component aging leads to component failure," "component failure leads to equipment failure," etc. Equipment correlation and impact rules use Horn clauses to characterize the propagation relationship of physical connections and logical dependencies between devices, such as "equipment A supplies power to equipment B," "equipment B failure affects the operation of equipment C," etc.

[0100] Based on the defined reasoning rules, this application employs a bottom-up reasoning strategy. Starting with a fault node in the knowledge graph, it recursively executes fault tracing and association impact rules to obtain the root cause of the fault and associated devices. Specifically, for each type of fault node, a depth-first search algorithm is first used to trace the causal chain step by step by matching and executing fault tracing rules until the root cause node, i.e., the root cause of the fault, is reached. Then, with the fault node as the center, device association impact rules are recursively called. Based on the device topology and logical dependencies, the scope of the fault's impact, i.e., the fault-associated devices, is determined. Through this reasoning process, the complex mechanisms and propagation paths behind the fault can be revealed.

[0101] To further refine diagnostic knowledge and support intelligent diagnostic decision-making, this application automatically generates structured diagnostic rules based on fault tracing and impact analysis. First, the reasoning paths for fault tracing are summarized, common root cause patterns for different fault types are extracted, and a fault-cause diagnostic decision tree is constructed. Then, the transmission paths of fault associations are abstracted, key links in fault propagation are extracted, and a fault-equipment diagnostic decision tree is constructed. Finally, the diagnostic paths of fault phenomena, intermediate causes, root causes, and affected equipment are structurally represented, forming a diagnostic rule template of "fault phenomenon - intermediate cause - root cause - affected equipment". Based on this template, fault nodes in the knowledge graph are matched one by one, automatically instantiating and generating a series of diagnostic rules.

[0102] Considering that knowledge-based reasoning rules rely heavily on domain models and expert experience, this application also introduces a data-driven rule mining method to improve the comprehensiveness and adaptability of the diagnostic rule base. Using machine learning toolkits such as Weka and Scikit-learn, correlation patterns between faults and their causes, equipment, and monitoring quantities are mined from historical fault case data of railway signaling equipment to form data-driven diagnostic rules. These rules are represented in the form of "fault condition - causing - affecting equipment - key monitoring quantity," such as "when the vibration amplitude of the equipment continues to increase, it may be due to unstable equipment foundation, affecting the stability of the track circuit, and the change in the track circuit current value needs to be closely monitored." By integrating knowledge-driven structured diagnostic rules and data-driven correlation patterns, a comprehensive diagnostic rule base with both interpretability and adaptability is formed.

[0103] During the generation of diagnostic rules, different rules may have different confidence levels and priorities. To improve the reliability of diagnostic decisions, this application employs evidence weighting to rank the rules. On one hand, the confidence level of structured diagnostic rules is calculated using the prior probability of fault nodes in the knowledge graph and the conditional probabilities of fault-cause and fault-equipment edges. On the other hand, the statistical significance of data-driven diagnostic rules is evaluated using indicators such as support and confidence levels. Based on this, the rules in the diagnostic rule base are weighted and ranked in conjunction with factors such as rule coverage and conflict level, forming a set of diagnostic rules with clear priorities and sufficient evidence.

[0104] This application proposes a fault diagnosis method for railway signaling equipment based on knowledge graph reasoning. Based on the equipment knowledge graph, fault tracing and correlation influence rules are defined using a data log rule language. Logical programming is used to reason about the root causes of faults, revealing the fault's occurrence mechanism and propagation path. Based on the reasoning results, structured diagnostic rules are generated from the bottom up, and data-driven correlation patterns are integrated to construct a comprehensive diagnostic rule base. An evidence weighting mechanism is introduced to rank the rules in the rule base, improving the reliability of diagnostic decisions. This method utilizes the semantic relevance of knowledge graphs to model the complex logic of faults and obtains implicit knowledge through chain-like reasoning, overcoming the problems of flattening and independence in traditional diagnostic rules. Simultaneously, the integration of knowledge-driven and data-driven rule generation mechanisms balances the interpretability and adaptability of diagnostic knowledge, laying the foundation for intelligent diagnostic decision-making.

[0105] In one alternative implementation,

[0106] Based on the fault diagnosis rule base and the fused dataset, a fault prediction model for railway signaling equipment is constructed, including:

[0107] Using natural language processing technology, the fault diagnosis rules in the fault diagnosis rule base are transformed into vector representations. The semantic similarity of fault concepts is learned through word embedding. Fault symptoms and monitoring condition elements in the fault diagnosis rules are mapped into semantic vectors. Through vector operations, a fault representation vector that integrates the semantic features of the diagnosis rules is constructed.

[0108] The fusion feature vectors and fault label data in the fusion dataset are balanced. Based on the sample distribution characteristics of different fault modes, an adaptive synthetic minority class oversampling technique is used to enhance the fault samples. The fault duration is calculated based on the sample timestamp and used as the time attribute of the synthetic sample to construct time series monitoring data.

[0109] Based on time series monitoring data, a gated cyclic unit network is used to model the time series feature vector. The fault representation vector and the time series feature vector are concatenated to construct a feature space containing rule knowledge and monitoring data. A weighted cross-entropy loss function is used to adaptively adjust the penalty weight according to the number of samples of each fault category, and an early stop method is introduced to control the number of iterations to obtain a fault prediction model for railway signaling equipment.

[0110] For example, fault prediction for railway signaling equipment involves analyzing equipment status data to identify potential faults in advance, thereby enabling timely maintenance and preventing equipment damage and accidents. This application constructs a fault prediction model that integrates diagnostic rule knowledge and monitoring data based on a fault diagnosis rule base and a fused dataset, achieving equipment status assessment and fault early warning based on heterogeneous information.

[0111] Traditional fault prediction methods primarily rely on monitoring data modeling, neglecting the valuable information inherent in expert knowledge and diagnostic rules. To incorporate structured knowledge from fault diagnosis rule bases into data-driven prediction models, this application employs natural language processing techniques to learn the semantic representation of diagnostic rules, thereby achieving feature space fusion between diagnostic rules and monitoring data.

[0112] First, word embedding is used to learn the semantic similarity of concept elements in fault diagnosis rules. Key concepts such as fault type, fault symptoms, and monitoring conditions from the rule base are represented by word embeddings, and low-dimensional dense vectors of these concepts are obtained by training on a domain corpus. The word embedding model can be trained quickly using tools such as Word2Vec. Word embedding vectors can characterize the semantic relationships between concepts; for example, "power failure" and "power supply anomaly" are relatively close in vector space.

[0113] Then, the structured elements in the fault diagnosis rules are mapped to semantic vectors. For each component of the rule, such as fault phenomenon, monitoring conditions, and associated equipment, it is mapped to a real-valued vector through a word embedding matrix based on the conceptual vocabulary it contains. Through vector concatenation, pooling, and other operations, the entire diagnosis rule can be mapped to a fixed-length semantic vector representation.

[0114] After obtaining the semantic vector representations of the diagnostic rules, vector operations are used to fuse the semantic features of the diagnostic rules to construct a comprehensive fault representation vector. Common fusion strategies include vector weighted averaging and max pooling. For example, for a fault like "signal light malfunction," the semantic vectors of relevant diagnostic rules (such as relay fault rules and power supply fault rules) are fused to form a comprehensive semantic representation of the fault. The fault representation vector incorporates structured knowledge from the diagnostic rule base, reflecting the semantic features and diagnostic logic of the fault.

[0115] Through the above steps, the fault diagnosis rule base is transformed into a semantic vector representation, realizing the mapping between diagnostic rules and numerical features. This allows it to be fused with the feature space of monitoring data for subsequent fault prediction model training. Learning diagnostic rule representations based on semantic embedding fully utilizes expert knowledge, compensates for the shortcomings of single data-driven methods, and provides more comprehensive feature information for fault prediction.

[0116] After obtaining the semantic representation of the fault diagnosis rules, it is necessary to further integrate monitoring data, construct time series features, and use a deep learning model to train the fault predictor.

[0117] Because the frequency of faults in real-world scenarios is far lower than that of normal conditions, fault samples are sparse and their distribution is imbalanced. To improve the accuracy and robustness of fault prediction, the sample data needs to be balanced. Adaptive Synthetic Sampling (ADASYN) is used to enhance fault samples, taking into account the sample distribution characteristics of different fault modes. ADASYN adaptively determines the number and location of synthesized samples based on the distribution of minority class samples, increasing the diversity of fault samples. When synthesizing new samples, the duration of the fault is calculated based on the timestamp of the original fault sample, and this duration is used as the time attribute of the synthesized sample to reflect the dynamic characteristics of the fault. Using the above methods, a fault monitoring dataset in time series format is constructed.

[0118] Based on the fused dataset, a Gated Recurrent Unit (GRU) network is used to model the time series data. GRU adaptively captures the long-term and short-term dependencies of time series data through a gating mechanism, effectively extracting feature representations in the time dimension. The multivariate time series data of each sample in the fused dataset is used as input to the GRU to learn time series feature vectors. At the last time step of the GRU network, the time series feature vectors are concatenated with the fault representation vector obtained in the previous step to form a joint feature representation that integrates diagnostic rule knowledge and monitoring data, which serves as input to the fault prediction model.

[0119] During model training, a weighted cross-entropy loss function is employed, adaptively adjusting class weights based on the number of samples in each fault category. Higher penalty weights are assigned to fault types with sparse samples, making the model focus more on minority class samples. Simultaneously, an early stopping strategy is introduced to monitor model performance on the validation set. If performance fails to improve after several consecutive epochs, training is terminated early to prevent overfitting. These methods optimize the model training process and improve its generalization ability for fault prediction.

[0120] The final fault prediction model integrates structured knowledge from the fault diagnosis rule base and temporal characteristics of monitoring data. Through end-to-end training, it learns a joint representation of diagnostic rules and data features, achieving fault prediction based on heterogeneous information. This model can predict the likelihood of equipment failure within a future period based on real-time monitoring data and historical fault diagnosis knowledge, providing a basis for preventative maintenance and operational decisions.

[0121] This application proposes a deep learning prediction model that integrates fault diagnosis rules and monitoring data. It employs semantic embedding technology to learn the semantic representation of the diagnosis rules, constructs a time-series monitoring dataset through sample balancing and temporal feature modeling, and trains an end-to-end fault predictor using a GRU network and a weighted loss function. This method fully leverages the complementary information of structured knowledge and data, improving the accuracy and interpretability of predictions in sparse fault scenarios, and providing a new approach for predictive maintenance of railway signaling equipment.

[0122] In one alternative implementation,

[0123] Based on time series monitoring data, a gated recurrent unit network is used to model time series feature vectors, including:

[0124] A gated recurrent unit network is used to perform sequence modeling on time series monitoring data. The time series monitoring data at each time step is mapped to a real-valued dense vector through an embedding layer, which serves as the input of the gated recurrent unit. In the gated recurrent unit, the input vector of the current time step and the hidden state of the previous time step are adaptively weighted using update gate and reset gate. The update gate controls the proportion of the hidden state information of the previous time step retained in the current time step, and the reset gate controls the proportion of the input vector of the current time step incorporated into the current hidden state.

[0125] By updating the gate and resetting the gate to aggregate the sequence dependencies between time steps, long-term context information in the time series monitoring data is determined; the steps of using the gated recurrent unit network to perform sequence modeling on the time series monitoring data and subsequent steps are repeated until the last time step of the time series monitoring data, to obtain the hidden state sequence of each time step of the gated recurrent unit network, which represents the dynamic feature information of the time series monitoring data at different time scales;

[0126] In a gated recurrent unit network, an attention query vector is generated using the hidden state of the last time step. The attention query vector is an aggregated representation of global context information. The similarity between the attention query vector and the hidden states of each time step in the gated recurrent unit network is calculated using an attention scoring function to obtain attention weights that represent the relevance of the hidden states of each time step to the current sequence modeling task. The hidden states of each time step in the gated recurrent unit network are linearly combined according to the attention weights to obtain a temporal feature vector that integrates global temporal features.

[0127] For example, in fault prediction tasks, time series monitoring data contains important information about the evolution of equipment status. To fully exploit the temporal features in time series data, this application uses a gated recurrent unit (GRU) network to model the time series data and introduces an attention mechanism to learn the importance of features at different time steps, extracting feature representations that fuse global temporal information.

[0128] GRU is a special type of recurrent neural network that adaptively captures long-short-term dependencies in time series through gating mechanisms. Compared to traditional RNNs, GRU introduces update and reset gates to control the flow of information, mitigating the vanishing and exploding gradient problems, and is better able to model sequence dependencies over long time spans.

[0129] In the process of extracting temporal features, the first step is to use an embedding layer to map the multivariate monitoring data at each time step into a real-valued dense vector. The embedding layer can be seen as a dimensionality reduction mapping, which transforms the original high-dimensional monitoring data into a low-dimensional representation with condensed information, thus helping to improve the learning efficiency of the model.

[0130] The update gate controls the proportion of hidden state information from the previous time step retained in the current time step, while the reset gate controls the proportion of the input vector from the current time step incorporated into the current hidden state. Through adaptive weighting of these two gates, GRU can flexibly capture dynamic changes and long-term dependencies in time series.

[0131] The final hidden state is obtained by weighting the hidden state from the previous time step and the candidate hidden states using the update gate, achieving selective forgetting of historical information and selective memorization of current information. By recursively applying the GRU update formula along the time dimension, the hidden state sequence of time series monitoring data at each time step in the GRU network can be obtained. The hidden state sequence contains dynamic feature information of the time series data at different time scales.

[0132] To further extract global feature representations of the time series, this application introduces an attention mechanism on top of the GRU network to adaptively aggregate hidden state information from different time steps. The hidden state of the last time step is used as the attention query vector, and its similarity to the hidden states of each time step is calculated using an attention scoring function.

[0133] The temporal feature vector aggregates key information from different time steps in the time series data. Through adaptive weighting using an attention mechanism, it highlights salient features relevant to the fault prediction task, suppresses interference from redundant information, and improves the relevance and effectiveness of the feature representation. Finally, the obtained temporal feature vector is concatenated with the semantic representation of the fault diagnosis rules to form a fusion feature containing temporal information and structured knowledge, which is used for downstream fault prediction tasks.

[0134] This application proposes a temporal feature extraction method based on GRU and attention mechanisms. It utilizes a GRU network to model the dynamic features of time-series monitoring data, adaptively captures long- and short-term temporal dependencies through update and reset gates, and introduces an attention mechanism to aggregate salient feature information from different time steps. This method fully exploits the temporal patterns and global contextual information in the time-series data. The learned temporal feature vectors can effectively characterize the health status and degradation trends of equipment, providing rich temporal feature support for fault prediction.

[0135] In one alternative implementation,

[0136] Also includes:

[0137] The correlation propagation algorithm is applied layer by layer to the fault prediction model of the railway signaling equipment. The correlation weight of each dimension of the real-time data of the railway signaling equipment to the fault diagnosis result is calculated through back propagation. For each layer of the gated cyclic unit network, the connection weight of each dimension of the real-time data of the railway signaling equipment to the neuron and the correlation score of the neuron to the fault diagnosis result are calculated.

[0138] The correlation weights of each dimension of the real-time data of the railway signaling equipment to the prediction category are calculated according to the layer-by-layer correlation propagation algorithm until the back propagation of all layers is completed, so as to obtain the correlation weight distribution of each time step of the input time series to the fault prediction result; the correlation weight distribution is normalized and visualized to determine the correlation weight visualization result.

[0139] The Shapley additive interpretation algorithm is applied to the railway signaling equipment fault prediction model. This involves calculating the marginal contribution of the model's input features to the prediction function. For each input feature, all feature subsets are enumerated, the marginal contribution of that feature within the subset is calculated, and the Shapley value of that feature is obtained according to the Shapley addition formula. This yields the distribution results of the Shapley values ​​for key features under different fault modes. The visualization results of the correlation weights and the Shapley value distribution results are then used to obtain the explanation of the prediction mechanism of the railway signaling equipment fault prediction model.

[0140] For example, when applying deep learning models to predict faults in railway signaling equipment, it is insufficient to focus solely on the model's predictive performance; it is also necessary to explain and analyze the model's decision-making mechanism. This application proposes a prediction mechanism explanation method based on the Relevance Propagation algorithm and the Shapley Additive Explanations (SHAP) algorithm for the constructed fault prediction model, aiming to reveal the causal relationship between input features and prediction results, thereby improving the model's interpretability and reliability.

[0141] First, the correlation propagation algorithm is used to calculate the relevance weights of the input features to the prediction results layer by layer. The correlation propagation algorithm distributes the relevance scores of the prediction results to the input features layer by layer through backpropagation, thus obtaining a quantitative association between the input features and the prediction results.

[0142] Specifically, for each layer of the Gated Recurrent Unit (GRU) network, the connection weights from each dimension of the real-time monitoring data of railway signaling equipment to the neurons and the correlation scores of the neurons to the fault diagnosis results are calculated. By applying the correlation propagation formula layer by layer, the correlation weight of each feature dimension of the input layer to the prediction result can be obtained. For time series data, information propagation along the time dimension also needs to be considered. In the GRU network, the information flow between different time steps is controlled by a gating mechanism; therefore, the effects of the reset gate and update gate need to be additionally considered when calculating the correlation weights.

[0143] Let h be the hidden state at time step t. t The correlation score is Reset door r t and update gate z t The relevance scores are respectively and Candidate hidden state The correlation score is Enter x t The correlation score is The correlation propagation formula can then be extended to:

[0144]

[0145]

[0146]

[0147] in, W represents the transpose of the weight matrix of the hidden state. h The weight matrix U represents the hidden state. h This represents the weight matrix from the hidden state at the previous time step to the current candidate hidden state. h represents the transpose of the weight matrix from the previous hidden state to the current candidate hidden state. t-1 W represents the hidden state at time t-1. r U r This represents the weight matrix from the hidden state at the previous time step to the reset gate, and the weight matrix from the input at the current time step to the reset gate.

[0148] W r U represents the weight matrix input to the reset gate at the current time step.r W represents the weight matrix from the hidden state in the previous time step to the reset gate. z U represents the weight matrix input to the update gate at the current time step. z This represents the weight matrix from the hidden state in the previous time step to the update gate. This represents the transpose of the weight matrix from the hidden state of the previous time step to the update gate.

[0149] Here, σ′ represents the derivative of the sigmoid function. By calculating the relevance score of the input data at each time step through backpropagation, the relevance weight distribution of each feature dimension to the prediction result in the complete time series can be obtained.

[0150] After obtaining the correlation weight distribution, it is normalized so that the sum of the weights is 1. The importance of different feature dimensions and time steps is presented intuitively through visualization methods such as heatmaps to help analyze the formation mechanism of the prediction results.

[0151] In addition to the correlation propagation algorithm, this application also introduces the Shapley additive interpretation algorithm to measure the contribution of input features to the prediction function. The Shapley value originates from game theory and is used to evaluate the contribution of each participant in a cooperative game. In the interpretation of machine learning models, the model can be viewed as a cooperative game, with input features representing the game participants and the model output representing the game payoff. The influence of each feature on the prediction result is measured by calculating its Shapley value.

[0152] Based on the size of the Shapley value, the importance of different features to the prediction results can be compared, and key influencing factors can be identified.

[0153] Since the number of feature combinations grows exponentially with the number of features, directly calculating the Shapley value is highly complex. To improve computational efficiency, this application employs a sampling-based approximation algorithm, estimating the Shapley value through Monte Carlo sampling. For each feature, a certain number of feature subsets are randomly sampled, the marginal contribution of the feature in the subset is calculated, and then the average of all sampling results is taken to obtain the estimated Shapley value of the feature.

[0154] By calculating the Shapley values ​​of each input feature in the fault prediction model, the importance distribution of key features under different fault modes can be obtained. Combining the Shapley value distribution with the visualization results of correlation weights can reveal the decision-making mechanism of the prediction model from both global and local perspectives, identify key factors affecting fault diagnosis, and provide guidance for model optimization and improvement.

[0155] This application proposes a predictive model interpretation method based on correlation propagation and Shapley value. By calculating the correlation weights of input features to the prediction results layer by layer, the importance distribution of each feature dimension and time step in the time series is obtained. Simultaneously, the Shapley value is used to measure the contribution of features to the prediction function, revealing key influencing factors. This method comprehensively utilizes the internal structure of deep learning models and game theory tools to explain the model's decision-making mechanism from multiple perspectives, improving the interpretability and credibility of fault prediction results and providing a basis for intelligent operation and maintenance decisions for railway signaling equipment.

[0156] Figure 2 This is a schematic diagram of the fault diagnosis system for railway signaling equipment according to an embodiment of the present invention, as shown below. Figure 2 As shown, the system includes:

[0157] The first unit is used to acquire multi-source heterogeneous data of railway signaling equipment. The multi-source heterogeneous data includes equipment operating parameters, environmental factors, and maintenance records. The equipment operating parameters include the voltage, current, and temperature parameters of the equipment. The environmental factors include the temperature, humidity, and vibration parameters of the environment in which the equipment is located. The maintenance records include the historical maintenance data of the equipment and the operation records of maintenance personnel.

[0158] The second unit is used to fuse the multi-source heterogeneous data using a factorization machine to obtain a fused dataset; based on the fused dataset, an ontology construction method and knowledge graph technology are used to construct a railway signaling equipment knowledge graph, which includes the component composition, working principle and fault cases of railway signaling equipment.

[0159] The third unit is used to obtain the root causes of faults in the railway signaling equipment and the correlation between equipment through knowledge reasoning based on the knowledge graph of the railway signaling equipment, and generate a fault diagnosis rule base; construct a fault prediction model for railway signaling equipment based on the fault diagnosis rule base and the fused dataset; input the real-time data of the railway signaling equipment into the fault prediction model for railway signaling equipment, and output the fault diagnosis results and fault cause analysis of the railway signaling equipment.

[0160] A third aspect of the present invention,

[0161] An electronic device is provided, comprising:

[0162] processor;

[0163] Memory used to store processor-executable instructions;

[0164] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0165] Fourth aspect of the embodiments of the present invention,

[0166] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0167] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method for railway signaling equipment, characterized in that, include: Acquire multi-source heterogeneous data of railway signaling equipment, including equipment operating parameters, environmental factors, and maintenance records; wherein, the equipment operating parameters include the equipment's voltage, current, and temperature parameters, the environmental factors include the temperature, humidity, and vibration parameters of the environment in which the equipment is located, and the maintenance records include the equipment's historical maintenance data and the operation records of maintenance personnel; The multi-source heterogeneous data is fused using a factorization machine to obtain a fused dataset. Based on the fused dataset, a railway signaling equipment knowledge graph is constructed using ontology construction methods and knowledge graph technology. The railway signaling equipment knowledge graph includes the component composition, working principle, and fault cases of railway signaling equipment. Based on the knowledge graph of the railway signaling equipment, the root causes of the faults of the railway signaling equipment and the correlation between the equipment are obtained through knowledge reasoning, and a fault diagnosis rule base is generated. According to the fault diagnosis rule base and the fused dataset, a fault prediction model of the railway signaling equipment is constructed. Real-time data of the railway signaling equipment is input into the fault prediction model of the railway signaling equipment, and the fault diagnosis results and fault cause analysis of the railway signaling equipment are output.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data is fused using a factorization machine to obtain a fused dataset, including: Multi-source heterogeneous sample data is input into an initial factorization machine model. For each feature of the multi-source heterogeneous sample data, the features of the multi-source heterogeneous sample data are mapped to a latent space of a preset dimension through latent vectors to obtain latent vector representations. The second-order interaction between features in the multi-source heterogeneous data is modeled by calculating the inner product of the latent vector representations to obtain interaction features. Based on the latent vector representation and the interaction features, the predicted value of the initial factorization machine model is calculated. The predicted value consists of a global bias term, a first-order linear term, and a second-order interaction term. Taking maximizing the difference between the predicted values ​​of positive and negative samples output by the initial factorization machine model as the objective function, the parameters of the initial factorization machine model are trained and updated using the gradient descent algorithm to obtain the trained factorization machine model. The multi-source heterogeneous data is fused using a trained factorization machine model to obtain a fused feature vector, which represents information about equipment operating parameters, environmental factors, and maintenance records. A fused dataset of railway signaling equipment is constructed based on the fused feature vector, in which each sample contains the fused feature vector and the corresponding fault label data.

3. The method according to claim 1, characterized in that, Based on the fused dataset, a railway signaling equipment knowledge graph is constructed using ontology construction methods and knowledge graph technology, including: Based on the fused dataset, an ontology construction method is used to define the top-level concepts of the railway signaling equipment knowledge graph. The top-level concepts include equipment concepts, component concepts, fault type concepts, fault cause concepts, and fault repair concepts. The equipment concepts include signal lights, track circuits, and switch machines. The component concepts include light-emitting units, relays, and motors. The fault type concepts include signal light malfunction and switching failure. The fault cause concepts include component aging and environmental interference. The repair scheme concepts include component replacement and parameter adjustment. Determine the hierarchical and semantic relationships between different top-level concepts. The hierarchical relationships include the subordinate relationship between the component concept and the equipment concept, the relationship between the fault type concept and the component concept and the equipment concept respectively, and the relationship between the maintenance plan concept and the fault type concept and the fault cause concept respectively. The semantic relationships include the causal relationships between fault and equipment, fault and fault cause, and fault and maintenance plan. Named entity recognition technology is used to extract instances of equipment, components, faults, causes, and maintenance solutions from the fused dataset, and these instances are mapped to knowledge graph nodes based on concept hierarchy definitions. The fused dataset is traversed to extract attribute information of node instances and add it as attributes of knowledge graph nodes. Relation extraction technology is used to mine semantic associations between instances in the fused dataset, construct relation edges between instances, and complete the remaining relation edges in the knowledge graph through ontology reasoning mechanism to form a knowledge network of node associations. The railway signaling equipment knowledge graph is represented in ontology description language form, and instance nodes and instance edges are stored in triple form.

4. The method according to claim 1, characterized in that, Based on the railway signaling equipment knowledge graph, the root causes of faults in the railway signaling equipment and the relationships between equipment are obtained through knowledge reasoning, generating a fault diagnosis rule base, including: The fault tracing rules and device association impact rules are defined using a data log rule language. Horn clauses are used to represent the recursive causal chain between fault causes and the transitive relationship of physical connections and logical dependencies between devices. Logical programming methods are used to reason about the root causes of faults. The data log rule consists of a header and a body. The header is an atomic formula and the body is the conjunction of the atomic formulas. The Horn clauses are logical clauses. For each type of fault node in the knowledge graph, a bottom-up reasoning strategy based on depth-first search is adopted. By matching and executing fault tracing rules, the root cause of the fault is obtained. The device association and impact rules are recursively called to determine the associated devices of the fault. Based on the root cause of the fault and the associated devices, common root cause patterns and association transmission paths of different fault types are extracted to construct a fault diagnosis decision tree. Through the reverse diagnosis path from the leaf node to the root node, structured diagnosis rules are generated. The structured diagnosis path includes fault phenomena, intermediate causes, root causes, and affected devices. Using machine learning toolkits, correlation patterns between faults and causes, equipment, and monitoring quantities are mined from historical fault case data to generate data-driven diagnostic rules. These data-driven diagnostic rules include fault conditions, causes, affected equipment, and key equipment. Structured diagnostic paths and data-driven diagnostic rules are integrated to obtain fault diagnosis rules. The credibility of multiple fault diagnosis rules is ranked based on evidence weights to construct a fault diagnosis rule library.

5. The method according to claim 1, characterized in that, Based on the fault diagnosis rule base and the fused dataset, a fault prediction model for railway signaling equipment is constructed, including: Using natural language processing technology, the fault diagnosis rules in the fault diagnosis rule base are transformed into vector representations. The semantic similarity of fault concepts is learned through word embedding. Fault symptoms and monitoring condition elements in the fault diagnosis rules are mapped into semantic vectors. Through vector operations, a fault representation vector that integrates the semantic features of the diagnosis rules is constructed. The fusion feature vectors and fault label data in the fusion dataset are balanced. Based on the sample distribution characteristics of different fault modes, an adaptive synthetic minority class oversampling technique is used to enhance the fault samples. The fault duration is calculated based on the sample timestamp and used as the time attribute of the synthetic sample to construct time series monitoring data. Based on time series monitoring data, a gated cyclic unit network is used to model the time series feature vector. The fault representation vector and the time series feature vector are concatenated to construct a feature space containing rule knowledge and monitoring data. A weighted cross-entropy loss function is used to adaptively adjust the penalty weight according to the number of samples of each fault category, and an early stop method is introduced to control the number of iterations to obtain a fault prediction model for railway signaling equipment.

6. The method according to claim 5, characterized in that, Based on time series monitoring data, a gated recurrent unit network is used to model time series feature vectors, including: A gated recurrent unit network is used to perform sequence modeling on time series monitoring data. The time series monitoring data at each time step is mapped to a real-valued dense vector through an embedding layer, which serves as the input of the gated recurrent unit. In the gated recurrent unit, the input vector of the current time step and the hidden state of the previous time step are adaptively weighted using update gate and reset gate. The update gate controls the proportion of the hidden state information of the previous time step retained in the current time step, and the reset gate controls the proportion of the input vector of the current time step incorporated into the current hidden state. By updating the gate and resetting the gate to aggregate the sequence dependencies between time steps, long-term context information in the time series monitoring data is determined; the steps of using the gated recurrent unit network to perform sequence modeling on the time series monitoring data and subsequent steps are repeated until the last time step of the time series monitoring data, to obtain the hidden state sequence of each time step of the gated recurrent unit network, which represents the dynamic feature information of the time series monitoring data at different time scales; In a gated recurrent unit network, an attention query vector is generated using the hidden state of the last time step. The attention query vector is an aggregated representation of global context information. The similarity between the attention query vector and the hidden states of each time step in the gated recurrent unit network is calculated using an attention scoring function to obtain attention weights that represent the relevance of the hidden states of each time step to the current sequence modeling task. The hidden states of each time step in the gated recurrent unit network are linearly combined according to the attention weights to obtain a temporal feature vector that integrates global temporal features.

7. The method according to claim 1, characterized in that, Also includes: The correlation propagation algorithm is applied layer by layer to the fault prediction model of the railway signaling equipment. The correlation weight of each dimension of the real-time data of the railway signaling equipment to the fault diagnosis result is calculated through back propagation. For each layer of the gated cyclic unit network, the connection weight of each dimension of the real-time data of the railway signaling equipment to the neuron and the correlation score of the neuron to the fault diagnosis result are calculated. The correlation weights of each dimension of the real-time data of the railway signaling equipment to the prediction category are calculated according to the layer-by-layer correlation propagation algorithm until the back propagation of all layers is completed, so as to obtain the correlation weight distribution of each time step of the input time series to the fault prediction result; the correlation weight distribution is normalized and visualized to determine the correlation weight visualization result. The Shapley additive interpretation algorithm is applied to the railway signaling equipment fault prediction model. This involves calculating the marginal contribution of the model's input features to the prediction function. For each input feature, all feature subsets are enumerated, the marginal contribution of that feature within the subset is calculated, and the Shapley value of that feature is obtained according to the Shapley addition formula. This yields the distribution results of the Shapley values ​​for key features under different fault modes. The visualization results of the correlation weights and the Shapley value distribution results are then used to obtain the explanation of the prediction mechanism of the railway signaling equipment fault prediction model.

8. A fault diagnosis system for railway signaling equipment, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire multi-source heterogeneous data of railway signaling equipment. The multi-source heterogeneous data includes equipment operating parameters, environmental factors, and maintenance records. The equipment operating parameters include the voltage, current, and temperature parameters of the equipment. The environmental factors include the temperature, humidity, and vibration parameters of the environment in which the equipment is located. The maintenance records include the historical maintenance data of the equipment and the operation records of maintenance personnel. The second unit is used to fuse the multi-source heterogeneous data using a factorization machine to obtain a fused dataset; based on the fused dataset, an ontology construction method and knowledge graph technology are used to construct a railway signaling equipment knowledge graph, which includes the component composition, working principle and fault cases of railway signaling equipment. The third unit is used to obtain the root causes of faults in the railway signaling equipment and the correlation between equipment through knowledge reasoning based on the knowledge graph of the railway signaling equipment, and generate a fault diagnosis rule base; construct a fault prediction model for railway signaling equipment based on the fault diagnosis rule base and the fused dataset; input the real-time data of the railway signaling equipment into the fault prediction model for railway signaling equipment, and output the fault diagnosis results and fault cause analysis of the railway signaling equipment.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images