Self-Supervised Representation Learning for Cross-Domain Industrial Fault Diagnosis

Abstract

Industrial fault diagnosis has been transformed by deep learning, but the field's central practical obstacle remains largely
unsolved: labeled fault data is scarce and expensive to collect, and a diagnostic model trained on one operating condition tends
to degrade sharply when applied to another. This article synthesizes the rapidly growing literature that addresses both
problems at once through self-supervised representation learning. We trace the general self-supervised methods, contrastive
instance discrimination, bootstrap-style non-contrastive learning, and masked reconstruction, that originated in computer
vision and speech, and follow their adaptation into pretext tasks defined directly on vibration and motor-current signals. We
then examine how these pretrained representations are combined with domain-adversarial and Wasserstein-based
distribution-alignment mechanisms, descended from the original domain-adversarial neural network formulation, to produce
diagnostic pipelines that remain accurate across varying loads, speeds, and even previously unseen fault types under open-set
conditions. Reported results across independent studies are consistent: self-supervised pretraining followed by lightweight
labeled fine-tuning substantially outperforms classical unsupervised feature learning under matched, restricted-label budgets,
and recent open-set domain-adaptation methods achieve F1-scores above 96% across three structurally different bearing and
gearbox benchmarks. We close by identifying the field's most pressing open challenges: a continued reliance on a small number
of laboratory benchmarks, an under-theorized interaction between pretraining and alignment stages, and a near-total
absence of evaluation on naturally occurring field faults.