Forecasts of multivariate time series sampled from industrial machinery sensors

Abstract

Prognostics assesses and predicts future machine health, which includes detecting incipient failures and predicting remaining useful life. Several studies have treated prognostics from a time series forecasting perspective. The main goal of this study is to evaluate the performance of a set of methods in the prediction of future values on a dataset of time series collected from sensors installed in an industrial gas turbine. Forecasting methods tested include the use of multivariate and univariate neural networks (FNN and LSTM), exponential smoothing and ARIMA models. Results show that the use of ARIMA models to forecast on the studied dataset is the best default method to apply, and is the only forecasting method that consistently beats a simple naïve no-change model