Adaptive noise-resilient fault diagnosis for rotating machinery via dynamic multiscale transformer networks

Structural health monitoring (SHM) of rotating machinery is critical for preventing catastrophic failures in industrial systems, yet challenges persist due to complex signal noise and variable operational conditions. This study proposes a novel framework integrating complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and a dynamic feedforward parallel transformer (DFP-Former) to address the aforementioned problems. CEEMDAN decomposes vibration signals into intrinsic mode functions, enhancing fault-related feature extraction. The DFP-former dynamically integrates global dependencies and local patterns through an overlapping spatial reduction attention mechanism and input-dependent depthwise convolution, while a multiscale feedforward convolution module captures discriminative features across frequencies. Extensive experiments on bearing and gearbox datasets demonstrate superior accuracy (99.24 and 99.40%) under varying noise levels (−6 to 6 dB) and imbalanced data, outperforming state-of-the-art methods. The proposed approach offers a scalable solution for real-time SHM applications, enabling reliable fault diagnosis in complex industrial environments.