A reliable rolling bearing fault diagnosis method based on Titan

Abstract

The accuracy of fault diagnosis for rolling bearings degrades sharply when operating conditions shift. Existing high-precision classifiers often experience a drop of over 50% in predictive accuracy when speed or load fluctuates, which seriously jeopardizes the reliability of industrial equipment health monitoring. Titan, a recently proposed architecture for long-context language modelling, tackles a similar challenge of maintaining performance across varying contexts. TitanDiag adapts this mechanism to fault diagnosis. The underlying rationale is that a persistent memory accumulates evidence across operating conditions and stabilizes predictions when the current segment alone is ambiguous. The architecture places Titan's dual-path memory (a long-term store gated by surprise plus a short-term FIFO buffer) inside a Transformer encoder. The multi-view front-end provides three complementary representations for every vibration segment, namely the raw waveform, the Fourier magnitude spectrum, and the continuous wavelet transform scalogram. At inference, Monte Carlo dropout produces per-prediction uncertainty scores that align naturally with Titan's surprise metric. On the CWRU and PU bearing benchmarks, TitanDiag attains 99.25% accuracy on the challenging PU-C2 low-speed condition, where TimeMachine and TSCMamba drop to 41.68% and 60.20%, respectively. The mean error-detection AUROC reaches 0.9668, well above the best baseline of 0.9391, demonstrating that the memory-driven variance inflation produces uncertainty estimates that are closely aligned with actual misclassification patterns.