[1F5] Application of a lightweight transformer in computationally efficient diagnosis of rotating machine faults
T Mian¹, A Tripathi² and P Kundu¹,²
¹KU Leuven, Belgium
²Vellore Institute of Technology, India
Artificial intelligence (AI)-based strategies have played an important role in the intelligent diagnosis of rotating machines. Recently, transformers have shown promising performance in this field by modelling global dependencies in time-series signals. Also, for computationally efficient applications, their lightweight versions are gaining much attention. However, the memory and computational requirements of such models remain a challenge for deployment on resource-constrained devices such as industrial edge units or embedded systems having no high-end graphics processing units (GPUs). In this work, an optimised lightweight transformer was employed through activation checkpointing (ACP) and post-training dynamic quantisation (PDQ) to enable memory-efficient training, thus enabling it for deployment in resource-limited scenarios. The performance evaluation of the proposed model was carried out on various bearing faults, demonstrating a substantial memory-efficient performance in constrained environments. The method demonstrated a perfect classification for faults and for the location of specific fault conditions with 100% accuracy. The work lays the foundation for integrating lightweight transformers into a real-time low-power diagnostic system and provides a modular structure suitable for future expansion through generative AI-based data augmentation.
