Trustworthy Fault Diagnosis Paradigm: Learning Models to Abstain from Uncertain Diagnosis

The utilization of deep learning techniques has made significant progress in the performance of fault diagnosis models. However, this success is often undermined by a critical limitation: current models tend to generate predictions even in instances where the model exhibits low confidence regarding the samples. This phenomenon results in erroneous diagnoses, which is particularly problematic in various industrial contexts that necessitate a high degree of accuracy. To mitigate this challenge, we propose a trustworthy fault diagnosis paradigm that systematically rejects uncertain samples while adhering to a specified coverage rate constraint, thereby ensuring that the diagnoses made by the model are both accurate and dependable. Empirical evaluations conducted on one synthetic dataset and three real-world fault diagnosis datasets demonstrated the efficacy of the proposed approach.