Abstract
Under complex industrial operating conditions with variable speeds, variable loads and high-level background noise, vibration-based rolling bearing fault diagnosis is challenging as noise obscures fault-related features. Existing fault diagnosis methods rely primarily on fault labels for feature learning and fail to fully exploit environmental information embedded in noise-level labels. To address this issue, to the authors’ knowledge, this paper is the first to treat noise-level identification as an auxiliary classification task in bearing fault diagnosis, and a multi-task multi-scale channel-mixed residual network is accordingly proposed. Taking fault classification as the main task and noise-level identification as the auxiliary task, the approach transforms noise-level information into effective auxiliary supervisory signals. A multi-scale channel-mixed residual network acts as the shared feature-extraction backbone. A hierarchical parameter-sharing architecture balances cross-task feature sharing and task-specific representation. A channel–spatial attention module embedded with the Kolmogorov–Arnold network enhances task-relevant features, while a dynamic loss-weighting strategy coordinates joint optimization of the two tasks. Experiments conducted on the public CWRU dataset and laboratory-tested bearing dataset under diverse noise, load and rotational-speed conditions show that noise-level auxiliary supervision raises average fault classification accuracy by 1.85 percentage points over single-task models. Compared with mainstream multi-task diagnostic methods including multi-task attention convolutional network, multi-task ConvFormer and multi-task one-dimensional convolutional neural network, the proposed method achieves higher diagnostic accuracy with smaller performance degradation under fluctuating operating conditions, demonstrating favorable robustness for practical condition-monitoring applications.