Predicting Machine RUL by Injecting Reflection-Padded Convolutional Outputs from Time Series Reconstruction into a Multi-Layer Perceptron

Predicting the Remaining Useful Life (RUL) of machinery is a critical component of Prognostics and Health Management (PHM), as it enables proactive maintenance, reduces downtime, and optimizes asset utilization. In this paper, we present a new approach for predicting the RUL, where time series reconstruction is employed and reflection-padded results of convolutional layers are further injected into a Multi-Layer Perceptron (MLP). Time-series reconstruction reorganizes one-dimensional temporal data into two-dimensional formats. It broadens the spatial and temporal receptive fields of convolutional layers so that spatial and temporal dependencies can be effectively captured. By applying reflection padding to preserve boundary features and increasing data diversity, this method is able to extract more robust features. The convolutional outputs of the modules with reflection-padding are then input into an MLP so that the nonlinear mappings of the MLP can give significant improvements regarding RUL estimation. Empirical evaluations conducted on the C-MAPSS dataset show that the proposed architecture is beneficial in terms of prediction accuracy and robustness, which outperforms traditional MLP and CNN models by significant margins and achieves the best performance among the previously proposed state-of-the-art models.