KD-ViT: A Lightweight Method for Online Wear State Recognition of Key Friction Pairs in Axial Piston Pump

Online wear state recognition of key friction pairs in the axial piston pump is of great significance for stable operation and predictive maintenance of the whole hydraulic system. Edge computing (EC) meets the real-time and low-cost requirements of online wear state recognition whereas two challenges limit its application. One is that current fault diagnosis methods only focus on local fault information, causing inaccuracy and poor generalization ability in different working conditions. The other is that the computing power and storage of EC devices are limited. Therefore, a lightweight knowledge-distilled vision transformer (ViT) is proposed for online wear state recognition. A novel time-frequency domain stacked and channel-weighted pooling structure is proposed to directly process raw time series. To realize high accuracy and high generalization ability, a ViT-based teacher model is pretrained to learn local and global information. To narrow model capacity gap and adapt to the limited resource of the edge node, a novel student model with a simplified self-attention mechanism is proposed to mimic the structure of the ViT and learn from the pretrained teacher model through knowledge distillation. An edge node with functions of signal acquisition, data preprocessing, and wear state recognition is designed and the distilled student model is deployed into it. Comparison with other state-of-the-art methods, ablation experiment, and online verification experiment demonstrate that the proposed method trades off wear state recognition performance and hardware limitations.