Predictive Maintenance of Manufacturing Industry Machines Using Remaining Useful Life (RUL) Prediction

In this paper, a lightweight Autoencoder-LSTM is provided to predict Remaining Useful Life (RUL) of bearings in real-life circumstances with noisy sensors and a small amount of labeled information. The method involves unsupervised health indicator (HI) construction and time modelling that is supervised. Through an autoencoder, trained on healthy data, vibration signals are converted into HIs on the basis of reconstruction error, which captures trends on degradation. These HI sequences are then inputted to an LSTM network that predicts RUL and monotonicity constraints are enforced to provide physically consistent, non-increasing predictions. With strict bearing-wise data splitting applied on the PRONOSTIA dataset to avoid leakage, the monotonicity-constrained model not only attains better RMSE (0.1700 vs 0.1819), but also non-physical prediction oscillations are removed, which is indicative of practical applicability in industry.