Intelligent fault diagnosis for gearbox via multi-head attention with Kolmogorov–Arnold network using multi-view feature fusion

Abstract Gearbox systems operating under dynamic mechanical loads are susceptible to progressive degradation and even catastrophic failures inherently, posing substantial risks to operational continuity and structural integrity. Conventional fault diagnosis methodologies face considerable challenges in processing vibration signals characterized by pronounced nonlinearities and nonstationary behaviors. To address these limitations, this paper introduces a novel diagnostic framework that integrates Kolmogorov–Arnold networks (KAN) with multi-head attention mechanisms. The proposed approach leverages complementary multi-view feature extraction encompassing temporal, spectral, time–frequency, and entropy-based signal characterizations, facilitating comprehensive fault pattern representation. Unlike conventional architectures that employ fixed activation functions, the model utilizes learnable spline-based transformations via KAN, enabling superior approximation of complex vibrational signatures. The hierarchical architecture incorporates feature mapping layers for latent space projection, multi-head attention modules for dependency modeling, and nonlinear classification via spline-optimized transformations. The framework’s diagnostic efficacy is evaluated systematically through comprehensive experimentation on two distinct gearbox datasets. Superior performance characteristics are demonstrated across varied operational regimes, with particular emphasis on diagnostic robustness and generalization capacity. Enhanced resilience to variable operating conditions is achieved through synergistic multi-view feature fusion, while nonlinear signal dynamics are effectively captured via the KAN computational paradigm, achieving generalizability that outperforms conventional deep learning architectures.