A Unified Label Noise-Tolerant Framework of Deep Learning-Based Fault Diagnosis via a Bounded Neural Network

Supervised fault diagnosis (FD) in mechanical systems, particularly in high-speed train (HST) components, faces significant challenges due to the presence of label noise in annotating large-scale monitoring data. This label noise introduces strict requirements for label noise tolerance and the learning capabilities of FD algorithms. This article presents a unified framework for label noise FD in HST components using a bounded neural network (BNN) to address this issue. The proposed framework consists of multiple basic models with shared weights, enabling the learning of global knowledge across sensor nodes and facilitating the estimation of local states to adapt to dynamic measurement networks. The BNN-based basic model incorporates implicit weighted learning and bounded loss mechanisms, which extract valuable insights from misannotated data. In addition, a tighter bound of loss is introduced, providing theoretical proof and enhancing the label noise tolerance of the BNN. A surrogate training strategy based on an alternative convex search (ACS) is established to ensure the stability of the BNN model during the early training stage. This strategy mitigates the risk of failure in the initial training phase of the BNN model. The feasibility and effectiveness of the proposed method are demonstrated through a real-field test conducted on a mechanical system of an HST component. The code is released on https://github.com/sudao-he/Bounded_Neural_Network.