A cross-teaching multi-output neural network prediction model for physical assets

The prediction of remaining useful life (RUL) is central to Prognostics and Health Management (PHM), and for an accurate prediction, one needs to build an effective prognostics model. However, in building a prognostics model there are some issues to be addressed, such as: assets having more than one fault mode, experiencing unexpected degradation, and some assets possessing more degradation information than others. Previous prognostic studies on assets designed a prediction model to identify fault classifications and another separate model for RUL estimation. We demonstrate the performance and potential advantages of a multi-output neural network prognostics framework, which accounts for more than one fault mode and produces simultaneously fault classifications and RUL predictions for turbofan engines. Using the multi-input-multi-output (MIMO) framework as benchmark results, this multi-output framework contains one final prediction model which could improve the results while being computationally more efficient. For further fault classification and RUL improvements, we introduced cross-teaching between two deep learning models within the multi-output framework. We compared the performance of the cross-teaching multi-output framework with the multi-output neural network framework. The proposed methodologies were applied to the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS)’s FD003, FD004, and the New Commercial Modular Aero-Propulsion System Simulation (N-CMAPSS)’s DS02 datasets. The results indicated that the cross-teaching multi-output neural network framework performed better overall in terms of root mean squared error (RMSE) across all datasets.