A Physics-Visual Collaborative Fault Diagnosis Method for Rolling Bearings Fusing Transient Clipping and Manifold Decoding

Abstract

To address the challenges in industrial edge computing, where complex, intense noise tends to induce overfitting in deep models and computational bottlenecks constrain performance, this paper proposes a lightweight physics-visual collaborative diagnostic architecture for bearing fault diagnosis. At the physical perception level, an adaptive truncation mechanism based on a normal-distribution prior is introduced to accurately remove out-of-bounds impulses, thereby restoring the signal’s generalized stationarity. Subsequently, a high-fidelity two-dimensional time-frequency manifold is mapped utilizing logarithmically enhanced STFT. At the visual decoding end, a lightweight GAP-CNN is constructed by replacing the fully connected layer with global average pooling. This topological optimization endows the network with spatial translation invariance, drastically reducing the parameter scale while mechanistically circumventing the local noise memory effect. Empirical validations across diverse dynamic operating conditions demonstrate that the proposed algorithmic architecture achieves accuracies of 95.2% and 97.4% under baseline and low-load conditions, respectively, and exhibits superior noise-resistant generalization robustness in the presence of complex, intense impulsive interference. Furthermore, integrating the dynamic laws of rotating machinery (E ∝ v2), this study quantitatively reveals the perceptual limits of the model under low-speed weak degradation conditions (with an accuracy of 93.6%), objectively elucidating the natural physical constraints of the generalized visual diagnostic paradigm dictated by extreme signal-to-noise ratios.