TrustMamba: calibrated uncertainty quantification for reliable and label-efficient bearing fault diagnosis

Abstract

Vibration-based bearing fault diagnosis informs maintenance decisions, but a predicted fault label is useful in practice only when its associated uncertainty is reliably quantified. Existing deep learning-based diagnostic models are often overconfident, poorly calibrated, and dependent on large labeled datasets, which limits their utility in risk-aware condition monitoring. This study proposes TrustMamba, a framework that reports calibrated uncertainty as part of the diagnostic output. Its Adaptive Uncertainty Module decomposes epistemic and aleatoric uncertainty and, together with post-hoc temperature scaling, aligns reported confidence with observed correctness so that each diagnosis is accompanied by a calibrated reliability estimate. A dual-stream representation pathway then combines a 2D selective state-space scan with a multi-scale convolutional mixer to capture both global temporal dependencies and local transient signatures in vibration signals. A Cost-aware Sample Selection strategy further embeds an asymmetric misclassification-cost prior into the labeling process, concentrating limited annotation effort on high-risk samples and reducing the measurement and annotation costs required to build a reliable diagnostic model. On the Case Western Reserve University, Jiangnan University, and Southeast University datasets, TrustMamba attains expected calibration errors (ECE) as low as 0.11%–0.30% while maintaining near-ceiling accuracy, and it retains the lowest calibration error among competing methods under low signal-to-noise measurement conditions down to 0 dB. The framework provides reliable, well-calibrated, and label-efficient bearing fault diagnosis for risk-aware condition monitoring.