Prognostics and Health Management (PHM) models aim to estimate remaining useful life (RUL) of complex systems, enabling lower maintenance costs and increased availability. A substantial body of work considers the development and testing of new models using the NASA C-MAPSS dataset as a benchmark. In recent work, the use of ensemble methods has been prevalent. This paper proposes two adaptations to one of the best-performing ensemble methods, namely the Convolutional Neural Network - Long Short-Term Memory (CNN-LSTM) network developed by Li et al. (IEEE Access, 2019, 7, pp 75464-75475)). The first adaptation (adaptable time window, or ATW) increases accuracy of RUL estimates, with performance surpassing that of the state of the art, whereas the second (sub-network learning) does not improve performance. The results give greater insight into further development of innovative methods for prognostics, with future work focusing on translating the ATW approach to real-life industrial datasets and leveraging findings towards practical uptake for industrial applications.

@en

Estimating the RUL (Remaining Useful Life) of machinery is a useful tool for maintenance and performance operations. This results in lower costs, improved safety and operational improvements.
This paper proposes two adaptations to the CNN-LSTM network provided by Li et al. \cite{Li2019APrediction}, as well as exploring reproducibility, accuracy and sensitivity of the original DAG (Directed Acyclic Graph) network. The network at hand is an ensemble network combining LSTM and CNN neural networks to provide an accurate regression RUL prediction using the NASA CMAPSS dataset \cite{NasaNasaReprository}.
The Adaptable Time Window (ATW) adaptation increases the amount of time cycles that can be predicted and increases the accuracy, allowing for earlier predictions and better RUL predictions. Allowing state-of-the-art predictions accuracy for complex datasets. The Sub-network training adaptions did not surpass the accuracy of the original network with the current implementation settings, however is promising for further research.