Center margin loss-based uncertainty-aware fault diagnosis for rotating machines to identify unseen faults

Recently, deep-learning-based uncertainty-aware fault-diagnosis methods have been developed to enhance the trustworthiness of fault-diagnosis results. However, these methods often fail to identify unseen faults, largely due to their lack of ability to extract discriminative features. In addition, existing methods are computationally intensive, as they require multiple iterations of model training or intricate probabilistic computations to calculate uncertainty. To address these challenges, this article proposes a novel uncertainty-aware machine fault diagnosis method named center margin loss-based fault diagnosis (CMLFD). The proposed method extracts highly discriminative features by regulating the distances between class centers and deep features during training. Furthermore, the method easily calculates the uncertainty of the input data with the need for only a single deterministic model training process, by comparing the center distances of the deep features and class boundaries. The effectiveness of the proposed method is validated through experimental studies on two rotating machine datasets. The results demonstrate that the method can successfully identify unseen faults, while maintaining high diagnostic performance.