Parviz Malekzadeh

An accurate estimate of remaining useful life (RUL) in health monitoring is a critical challenge. Deep learning as an effective solution to predict RUL has been applied frequently in recent years. However, the high computational load of very deep networks and inconsistent window lengths in real applications are a big deal. This paper proposes a multi-window stacking method with simple architectures to overcome these issues. In this method, raw multivariate time series are preprocessed with different window lengths to create new samples. Simple networks are trained for these samples. In health monitoring in training phase machines work to failure, while in testing phase monitoring data are cut down in early cycles, therefore, available run times are shorter in the test dataset. In our proposed method, RUL predictions of test data are replaced with prior smaller window predictions, when window lengths are increased. All multiple predictions as predictors are entered into a ridge regression to determine the optimized weights of each prediction. Experimental results demonstrate the superiority of the proposed method in comparison with other state-of-the-art models on the same dataset.