The switch from conventional to predictive maintenance promises industrial companies significant efficiency gains. In practice, however, the successful use of predictive maintenance often fails due to great uncertainties in the assessment of a system's health status. This project shows how the current megatrend of the Internet of Things can be utilized to develop a viable and robust estimator of a system's remaining useful life. In particular, various filter and smoothing methods are implemented and tested to account for sensor noise. Additionally, a recurrent neural network model with a stacked LSTM architecture is defined which allows remaining useful life estimation without the necessity of an underlying structural fault propagation model.
Please see "README Predictive Maintenance Project.pdf" for the full discussion.