Abstract:Rail trains are a vital component of railway passenger transport and regional economic development. The traction drive system serves as the power source for rail trains. Due to the complex composition of system components and the working environment being affected by coupled multi-physical fields such as electric, thermal, and magnetic influences, long-term operation often leads to wear, aging, and sudden anomalies in certain components. Failure to perform timely maintenance can result in significant economic losses and casualties. Thus, rapid and efficient fault diagnosis of the traction drive system is crucial for ensuring the safe and stable operation of rail trains. Existing fault diagnosis methods for traction drive systems can be divided into online and remote diagnostics. Online diagnostics typically rely on structured time-series data for analysis. However, as the system continues to operate, a significant amount of unstructured knowledge about faults accumulates, which is vital for enhancing the depth and breadth of fault diagnosis, yet online diagnostic methods are unable to effectively utilize this knowledge. Traditional remote diagnostics mainly involve experts using their experiential knowledge combined with fault logs, maintenance manuals, etc. However, expert consultations heavily depend on human experience, and the storage methods for fault logs are flat, lacking intuitiveness and difficult to query, which leads to inefficient utilization of unstructured knowledge. Therefore, there is an urgent need to introduce new methods to fully utilize unstructured fault knowledge and improve the efficiency of fault diagnosis. Knowledge graphs are a type of semantic network-based knowledge base, primarily extracting entities and relationships from unstructured knowledge for storage in a directed graph of nodes and edges. In the context of remote fault diagnosis systems, the introduction of knowledge graphs can standardize and unify the storage and expression of unstructured knowledge such as fault corpora and expert knowledge dispersed across various documents. This enables efficient knowledge querying and reasoning, which is critical for enhancing the utilization of unstructured knowledge, reducing dependency on expert experience, and ultimately improving diagnostic efficiency and shortening maintenance decision-making times. Currently, knowledge graphs have been widely applied in many research areas such as medicine, industrial manufacturing, and agriculture, showcasing their potential applications. However, most of these applications only consider the knowledge graph construction process within their specific fields, with few studies reported on using knowledge graphs for remote fault diagnosis in rail train traction drive systems. Moreover, the information extraction stage of building system knowledge graphs faces issues with low accuracy, especially given the diverse sources of fault corpora and the relatively rare domain-specific terminology in traction drive systems, making traditional knowledge graph construction methods challenging to apply directly. In summary, this paper proposes a remote fault diagnosis method based on domain-adapted knowledge graphs. Firstly, it introduces a domain-adapted segmentation method to construct a system fault lexicon, addressing the issue of large errors in splitting specialized phrases under a general lexicon. Secondly, it addresses the lack of interactivity in traditional information extraction processes causing disconnection between subtasks by introducing a dual BiLSTM combination method for information extraction, which shares features between entity extraction and relationship extraction to enhance their potential interconnection. The paper's experiments use long-term operational fault logs from a specific train model as data sources to validate the overall method's effectiveness; it also analyzes the effects of each component of the proposed method through ablation studies. Finally, based on the proposed method, a remote fault diagnosis system has been established, offering a new solution for rapid and efficient maintenance decision-making.

Fault Tracing on Transformer Substations Via Graph Kernel on Event Graphs

Graph-Based Method for Fault Detection in the Iron-Making Process.

Single-Ended Fault Location for Hybrid Transmission Line Using Embedded Artificial Intelligence

A Concurrent Fault Diagnosis Method of Transformer Based on Graph Convolutional Network and Knowledge Graph

Construction of transformer substation fault knowledge graph based on a depth learning algorithm

STAGED: A Spatial-Temporal Aware Graph Encoder–Decoder for Fault Diagnosis in Industrial Processes

Generalized Transformer in Fault Diagnosis of Tennessee Eastman Process

A Novel Label-Aware Global Graph Construction Method and Spiking-Coded Graph Neural Network for Intelligent Process Fault Diagnosis

The Study of Knowledge Graph Framework for Faults Disposal of Power Transformer

Enhancing Power Transformer Fault Diagnosis Through Dynamic Knowledge Graph Reasoning

Knowledge Graph Construction and Decision Support Towards Transformer Fault Maintenance

FaultTracer: interactive visual exploration of fault propagation patterns in power grid simulation data

Dynamic Fault Analysis in Substations Based on Knowledge Graphs

Power Transformer Fault Prediction Based on Knowledge Graphs

Fault diagnosis of power transformers using graph convolutional network

FL-TN:A Fault Location Algorithm Based on Tree Topology for Smart Grid

Dynamic Fault Characteristics Evaluation in Power Grid

GridOnto: Knowledge Representation and Extraction for Fault Events in Power Grid

Remote Fault Diagnosis Method for Traction Drive Systems Based on Domain-Adapted Knowledge Graphs

Fault Diagnosis of Transformer Based on KPCA and Elman Neural Network

Distribution Network Fault Identification Based on the Improved Transformer Network