A Hybrid Adaptive Fusion Deep Learning Model for Fault Diagnosis of Rotating Machinery Under Noisy Conditions

Rotating machinery is essential in modern industry.Arobust noise-resistant method is proposed to improve diagnostic accuracy under intense noise conditions. Initially, time-domain signals are transformed into the time-frequency domain using the Synchrosqueezing Short-Time Fourier Transform to reduce the impact of noise. A novel hybrid adaptive fusion deep learning model is then introduced, incorporating two new modules: the heterogeneous convolution adaptive fusion block and the global-local attention adaptive fusion block. These modules enable the integration of heterogeneous convolution operators and complementary attention mechanisms, optimizing component importance for identifying subtle features despite noise. Additionally, the traditional Multilayer Perceptron in the classification layer is replaced with Kolmogorov–Arnold Networks to improve diagnostic accuracy. Case studies demonstrate that the method has strong noise resistance under challenging signal-to-noise ratio conditions.