An Investigation on Machine Learning and Statistical Hybrid Framework to Predict Remaining Useable Life of Polyurethane Conveyor Belts

Abstract

The correct determination of Remaining Useful Life (RUL) is also required to allow predictive maintenance of industrial conveyor belt systems, especially in pharmaceutical manufacturing facilities where the reliability of operations and contamination prevention are vital. This paper describes a hybrid prognostic modeling strategy of RUL prediction of polyurethane (PU) conveyor belts, which run with varying load, drop height, and current to the motor. Vibration RMS, current deviation, and impact force measurements were used to obtain experimental degradation data that was converted to a composite health indicator that reflected the progressive cumulative damage. The first model of wear-out behaviour to be developed was a Weibull-based statistical reliability model, which was developed with the aim of characterizing wear-out behaviour and offering baseline life prediction. In order to model nonlinear and time-varying dynamics of degradation, Random Forest regression and Long Short-Term Memory (LSTM) networks were applied to estimate RUL using data. The hybrid model used an optimized weighting strategy to combine statistical and machine learning predictions. The quantitative analysis showed that the standalone Weibull model had a high R 2 of 0.91 and a low RMSE of 18.4 cycles, whereas the models using the random forest and LSTM had high R 2 values of 0.96 and 0.98 and low RMSE values of 11.2 and 7.6 cycles, respectively. The advanced hybrid framework was better than all other models reaching an R 2 of 0.99, an MEA of 4.3 cycles, and RMSE of 5.1 cycles, which greatly enhanced the stability of prediction and the lowering of late stage estimation variance. The suggested solution offers a powerful and scalable system of smart condition-based maintenance of pharmaceutical conveyor belt systems