Fault diagnosis of aero-engines based on a dual-branch CNN-KAN network

In response to the challenges of high model computation complexity and insufficient interpretability in bearing fault diagnosis for aviation engines, this paper proposes a dual-branch diagnostic model that integrates wavelet-initialized CNN with Kolmogorov-Arnold Network (KAN). By embedding Shannon, Morlet, and Gaussian wavelet basis functions into the wavelet convolution layer as physical prior knowledge, a domain-adaptive feature extraction module is constructed. This model combines the explicit mathematical modeling ability of the KAN branch to achieve collaborative learning of local time-domain features and global nonlinear dynamic features in vibration signals. Experiments on the HIT aviation engine bearing dataset demonstrate a testing accuracy of 99.39%, with a performance degradation rate of only 0.21% when the sample size is reduced to 15%, while also providing hierarchical interpretability. This study offers an efficient solution for fault diagnosis in critical aviation safety systems.