Attention-enhanced Koopman operator network for bearings fault diagnosis

Abstract As a core component of mechanical equipment, the fault diagnosis of rolling bearings is of great significance for the reliable operation of industrial systems. Traditional diagnostic methods have inadequate adaptability under complex working conditions since they rely on manual feature extraction and shallow machine learning models. Although deep learning methods can automatically extract features, they have high computational complexity, poor physical interpretability, and lower noise robustness. A bearing fault diagnosis method based on an attention-enhanced Koopman operator network (AKON) is proposed in this paper. Firstly, a Koopman feature space is constructed through a convolutional encoder to map the nonlinear dynamic system of bearing vibration into a linear evolution. Furthermore, to enhance the temporal correlation and discriminability of features, a temporal self-attention module is introduced to dynamically weight fault-sensitive time steps through scaled dot-product attention. Finally, to overcome the deficiency of relying on a single loss in conventional designs, a multi-task joint optimization mechanism is proposed, which incorporates reconstruction loss, dynamic loss, and classification loss. Experiments on the bearing datasets of QUT, CWRU and the Paderborn University show that the proposed method significantly outperforms traditional methods and mainstream deep learning models in terms of noise robustness and fault classification accuracy, achieving 98.11%, 99.90% and 95.62% accuracy on the three datasets respectively.