关键词: 可信故障诊断, 不确定性感知网络, 变分注意力, 不确定性量化及分解, 贝叶斯深度学习

Abstract: Deep learning-based fault diagnosis methods are limited by their black-box nature to give trustworthy and interpretable results. Most of the existing research on interpretable fault diagnosis focuses on developing interpretable modules to be embedded in deep models to give some physical meaning to the results, or using the results as a basis to infer the deeper logic of the model to make such decisions, with limited research on how to quantify the uncertainty in diagnostic results and explain their sources and composition. Uncertainty quantification and decomposition can not only provide confidence in diagnostic results, but also identify the source of unknown factors in the data, ultimately guiding the enhancement of the interpretability of diagnostic models. Therefore, Bayesian variational learning is proposed to be embedded into Transformer to develop an uncertainty-aware network for trustworthy mechanical fault diagnosis. A variational attention mechanism is designed and the corresponding optimization objective function is defined, which can model the prior and variational posterior distributions of attention weights, thus empowering the network to be aware of uncertainty. An uncertainty quantification and decomposition scheme is developed to achieve confidence characterization of diagnostic results and separation of epistemic and aleatoric uncertainty. Using fault diagnosis of planetary gearboxes as an example, the feasibility of the proposed method for trustworthy fault diagnosis is fully validated in an out-of-distribution generalization scenario where the test data contains unknown failure modes, unknown noise levels and unknown operating condition samples.

Key words: trustworthy fault diagnosis, uncertainty-aware network, variational attention, uncertainty quantification and decomposition, Bayesian deep learning