Developing Physics-Informed Deep Learning Frameworks for Predictive Maintenance of Rotating Machinery

Abstract

Predictive maintenance of rotating machinery is essential for improving industrial reliability, minimising downtime, and reducing maintenance costs in modern Industry 4.0 environments. Although deep learning approaches have demonstrated promising performance in machinery fault diagnosis, their dependence on large labelled datasets and limited physical interpretability restricts practical industrial deployment. This study proposes a novel Physics-Informed Deep Learning (PIDL) framework that integrates governing mechanical dynamics with a hybrid CNN–LSTM architecture for intelligent fault diagnosis and Remaining Useful Life (RUL) prediction of rotating machinery. The proposed framework combines vibration signal processing techniques, including Fast Fourier Transform (FFT) and Short-Time Fourier Transform (STFT), with physics-informed neural networks (PINNs) to embed system dynamic constraints directly into the learning process. Experimental validation was conducted using vibration datasets containing healthy and faulty bearing conditions under varying operating speeds. The proposed PIDL model achieved a fault classification accuracy of 98.4% and reduced RUL prediction error by 21% compared with conventional deep learning models, including CNN, LSTM, and Transformer-based approaches. The incorporation of physics-based constraints significantly improved robustness, generalisation capability, and interpretability under noisy and limited-data conditions. The results demonstrate that the proposed PIDL framework provides an efficient, scalable, and reliable predictive maintenance solution for smart industrial monitoring systems and next-generation intelligent manufacturing applications.