Robust Industrial Fault Diagnosis under Observation-Pattern Shift

Abstract

Vibration-based fault diagnosis is one of the most studied applications of deep learning to industrial signals, where models routinely reach near-perfect accuracy on matched test data. Real deployments, however, suffer packet loss, bus stalls, sensor failures, and outages, all of which leave the test-time signal partially observed in ways the training set never anticipated, and accuracy can drop by tens of percentage points. This regime is formalized as observation-pattern shift (OPS) and benchmarked on four industrially motivated mask families under a worst-case observation-pattern accuracy (WOPA) metric. The natural mask-augmentation defense leaves WOPA low while collapsing clean-test accuracy from above 99% to roughly 80%, a joint shortfall attributable to the training distribution sitting entirely at high severity. Cross-severity training-distribution coverage is identified as the corrective principle and instantiated three structurally different ways: a paired weak/strong cross-entropy recipe and two single-view variants. On CWRU at high mask severity all three restore clean accuracy above 99% and lift WOPA by roughly 2 to 4 points over the augmentation baseline (statistically resolved for the weak/strong recipe), with no recipe measurably outperforming the others over five seeds, pointing to severity coverage rather than any specific recipe as the main factor behind the gain. On SEU-bearing, where the baseline collapses to near chance, the same coverage recovers more than ten WOPA points and restores clean accuracy above 99%. On SEU-gear, where the baseline does not collapse, clean accuracy is still restored but WOPA does not measurably improve, consistent with a corrective rather than universally beneficial principle.