Rolling bearings are common and critical rotating components in mobile engineering machinery and vehicle systems, directly affecting equipment reliability and safety. However, in real-world fault detection, both sensor-collected vibration signals and training labels are often contaminated by noise, resulting in globally noisy training data and degraded performance of data-driven models. The purpose of this study is to develop an effective model to improve the accuracy and robustness of bearing fault diagnosis in noisy environments.
Stochastic Configuration Network (SCN) is a simple, efficient and fast single-hidden-layer feedforward neural network, offering good flexibility for fault detection. Based on SCN, robustness strategies and an adaptive output weight optimization method are introduced, and the network is extended to a deep architecture, thereby enhancing the model’s robustness and learning capability, forming the Robust Adaptive Deep SCN (RADSCN).
To validate the effectiveness and robustness of the proposed method, experiments were conducted under various noise conditions, using benchmark classification data sets and real-world fault diagnosis data sets. Results show that the model maintains good generalization and robustness even with noisy data.
This paper proposes a novel RADSCN, marking a significant advancement in robustness enhancement for fault detection. The proposed approach expands the application potential of data-driven fault diagnosis models in complex real-world environments and provides fresh perspectives for the future development of fault detection technologies and industrial maintenance.
