Joint Denoising and Classification for Fault Diagnosis Under Strong Noise

Industrial fault diagnosis often degrades when vibration signals are strongly contaminated by noise and only a few labeled samples are available. This paper proposes a joint denoising and classification framework for rolling bearing and gearbox diagnosis. An adaptive denoiser combines band-wise Wiener filtering with lightweight temporal smoothing to suppress noise while preserving fault-salient components. A dual time–frequency backbone with a learnable fusion gate aggregates complementary features, and a supervised contrastive head enlarges inter-class margins in the low-label regime. All modules are trained end-to-end under a single multi-task objective that couples time- and frequency-domain reconstruction with cross-entropy and contrastive terms. Experiments on three public datasets, under both clean and severely noisy conditions and with few-shot label budgets, show consistent gains in accuracy and robustness over strong baselines, indicating that the proposed method is practical for noise-robust and data-efficient industrial fault diagnosis.