MMRCNN-KKAN-GMAT: a multi-modal multi-scale framework with KKAN attention and augmented transformer for rotating machinery fault diagnosis

Abstract

To address the degradation of diagnostic accuracy caused by insufficient fault data and noise interference in practical applications, a novel rotating machinery fault diagnosis framework is introduced in this work. First, the one-dimensional raw signals, envelope signals, and two-dimensional continuous wavelet transform images of multi-sensor vibration signals are employed as multimodal inputs to achieve the complementary feature learning across different modalities. Then, a Multi-Modal Multi-Scale Residual Convolutional Neural Network is utilized for discriminative feature extraction from the multimodal inputs, while a kurtosis-guided KAN hybrid attention module is designed to perform joint channel and spatial attention calibration on the fused features, thereby highlighting the fault-related information and suppressing irrelevant components. Finally, a Gated Memory-Augmented Transformer is employed to model the global context of the fused sequential features, further enhancing the feature representation capability and temporal dependency modeling. Experiments are conducted on both bearing and gearbox datasets. The proposed method achieves 99.87% accuracy on the SEU gearbox dataset under multi-sensor fusion, and 99.57% accuracy on the HUST bearing dataset with only 10% training samples under single-sensor conditions, demonstrating its strong anti-noise performance and high diagnostic accuracy even with limited training data.