A confidence calibration method for reliable fault diagnosis based on input-dependent temperature scaling

In the field of fault diagnosis, most existing research primarily focuses on improving model accuracy and stability, while relatively little attention has been paid to model reliability and the quantification of uncertainty. When exposed to strong noise interference or fault patterns from unknown distributions, models often still produce deterministic diagnostic outputs without indicating the associated level of uncertainty. To address this issue, this paper proposes an adaptive post-hoc calibration method—Input-dependent Temperature Scaling (ITS)—built upon a Bayesian Convolutional Neural Network (B-CNN) to enhance confidence calibration. Unlike conventional global temperature scaling, ITS adaptively learns temperature parameters from input features, thereby enabling more fine-grained confidence adjustment. Within the B-CNN framework, ITS effectively integrates parameter uncertainty with input-dependent post-hoc calibration. Experimental results on the CWRU dataset and a test bench developed by our research group demonstrate that the proposed method can significantly reduce calibration error and improve model reliability, while maintaining stable classification accuracy.