Answers are grounded in this paper and cite their sources.
Intelligent fault diagnosis of industrial machinery is challenged by heterogeneous sensor modalities, inconsistent physical units, and the scarcity of labeled data across operating domains. Existing semi-self-supervised learning (semi-SSL) methods for fault diagnosis are typically evaluated on single-domain datasets and lack mechanisms to align fault representations across different machinery types and sensor configurations. This paper presents the Universal Machinery Fault Diagnosis Framework (UMFD), a semi-SSL framework that unifies vibration and acoustic signals from seven publicly available datasets spanning bearing, gearbox, and electrical motor machinery into a single diagnostic model. UMFD integrates three components. First, Unit Normalized Selective Embedding (UNSE), which standardizes multi-source signals into a common velocity-domain representation via physics-grounded unit conversion and alternating source interleaving. Second, Cross-Domain Contrastive Alignment (CDCA), which enforces domain-invariant fault representations by aligning the same fault signals across different datasets using fault labels during pre-training. Third, masked signal modeling (MSM), which promotes generalizable temporal feature extraction by reconstructing masked signal segments without labels. The resulting universal dataset comprises approximately 457 million single channel samples and 913 million two channel samples across four classes which consist of healthy, bearing faults, gearbox faults, and induction motor faults. UMFD achieves 97.60 % ± 0.74 accuracy with 100 % of fine-tuning labels, and reaches accuracy comparable to the strongest baseline’s fully labeled performance using only 25 % of fine-tuning labels. This work represents the first tier of a proposed three-stage scaling roadmap toward a Large Signal Model (LSM) for industrial condition monitoring.
No takes yet. Share an insight, caveat, or question.
Mert Sehri (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: