Reservoir Neural Networks and time series forecasting: application to failure prediction and predictive maintenance in aerospace.

Predictive maintenance is a budding field in the aerospace industry with certain caveats that have made implementation sluggish, especially considering the need for data availability and sensor technologies which lack wide commercial implementation industry-wide. This thesis proposes a ESN-GWO system, a Reservoir Neural Network whose configuration is optimized by means of a Grey Wolf Optimizer that explores the hyperparameters in search of the configuration best suited to forecast a given input signal. Once an accurate prediction is generated, the analysis to determine whether maintenance is required could begin. The necessary requirements for the system to generate these predictions are defined, with specific suggestions as to how a prediction can be improved through reservoir computing. Three different datasets are studied: the Mackey-Glass equation, the CMAPSS dataset and the power consumption from a resort facility. The preliminary results from these time series assist to determine certain common rules that improve the quality of the predictions and focus the optimization towards hyperparameter solutions that may allow for a faster approach to predictive maintenance.