Compound Fault Diagnosis for Industrial Machinery: Data-Driven Methods, Challenges, and LLM-Enhanced Perspectives

Abstract

Compound faults occur when several local faults exist at the same time in industrial machinery and affect condition-monitoring signals together. Their signatures become mixed through modulation, structural transmission, operating-condition variation, and sensor response, which makes feature separation and state recognition more difficult. This paper reviews data-driven compound fault diagnosis from a scenario-oriented perspective. It first summarizes the task formulation and three main difficulties: coupled fault signatures, combinatorial growth of fault states, and distribution shifts across operating conditions and machines. The review then groups representative methods into three areas: signal representation, multi-label and fault-semantic modeling, and generalization under changing domains and unseen fault combinations. The paper also examines the emerging role of large language models, focusing on retrieval-augmented knowledge integration, evidence-grounded reasoning, tool-using diagnostic agents, and deployment reliability. It then points to several directions for future work, including compositional generalization, cross-machine transfer, multimodal sensing, industrial foundation models, and autonomous maintenance. Overall, the review argues that practical compound fault diagnosis depends on combining signal-level intelligence with reusable fault semantics and verifiable knowledge-based reasoning.