Artificial intelligence for equipment operation and maintenance (O&M) has advanced rapidly in recent years, yet the field still lacks a reusable large-scale infrastructure. The central bottleneck is not simple data scarcity, but the long-standing fragmentation, heterogeneity, and weak interconnection among relevant resources. Directly constructing a single large-scale signal-label dataset is intrinsically difficult, because fault samples are scarce and high-quality labels depend heavily on expert knowledge and costly inspection or verification. At the same time, many valuable resources already exist in dispersed forms, including small open signal-label datasets, mechanism knowledge embedded in papers and standards, engineering maintenance texts and records, and reusable algorithms and tools. Here we present MuxHub, a unified multimodal infrastructure framework for aggregating these abundant but scattered heterogeneous resources. Through standardized datasets, dataset-paper relation graphs, intelligent agents, and benchmarking leaderboards, MuxHub organizes them into an interconnected and continually extensible system of research assets. To bridge non-text monitoring data and textual knowledge, MuxHub introduces a thought-chain-based multimodal data architecture and related bridging mechanisms, thereby supporting cross-modal alignment, multimodal sample fusion, and the construction of knowledge graphs and ontologies. MuxHub provides a route for transforming fragmented resources into a unified large-scale foundation for the training, retrieval, reasoning, and evaluation of multimodal foundation models and agents for equipment O&M, while improving the discoverability, reusability, and comparability of resources and remaining open to the continual incorporation of new assets.
Huang et al. (Mon,) studied this question.