Construction of Knowledge Graph for Air Compressor Fault Diagnosis Based on a Feature-Fusion RoBERTa-BiLSTM-CRF Model

Diagnosing complex air compressor systems with traditional data-driven deep learning models often results in isolated fault diagnosis, ignoring correlations between concurrent faults. This paper introduces a knowledge graph construction approach for the air compressor fault diagnosis field, using after-sales business data as the source. We propose a model based on Robustly Optimized Bidirectional Encoder Representations from Transformers (RoBERTa), specifically tailored for constructing a knowledge graph for air compressor fault diagnosis. By integrating Whole Word Masking (WWM) technology, Bidirectional Long Short-Term Memory (BiLSTM), and Conditional Random Fields (CRFs), our approach effectively extracts specific entities from unstructured data. On our dataset, the model achieved an average accuracy of 0.7962 and an F1 score of 0.7956, demonstrating notable improvements in both accuracy and recall for entity recognition tasks. The extracted entities were subsequently stored in a Neo4j graph database, facilitating the construction of a domain-specific knowledge graph for air compressor fault diagnosis.