Advanced deep learning techniques for offline and online predictive maintenance

In industry, predictive maintenance has recently emerged as a valuable solution to secure the equipment effectiveness, optimize production time and minimize maintenance costs. The complexity of the underlying relationship of equipment functioning data has lead machine learning techniques to be widely employed in this field. However, their application to address problems with scarce or unreliable reference information on the machine state still constitutes a barely explored research field. This Thesis presents offline and online predictive models based on deep learning algorithms for applications in an industrial context. After a comparative study on the forecasting performance of various statistical and machine learning techniques, three machine learning based solutions to predictive maintenance problems of filler machines are presented. The first one enables recipe-agnostic detection of anomalous functioning of filling valves and the calculation of their remaining useful life (RUL). The second solution employs the Online Evolving Spiking Neural Networks to detect anomalies in filler motors functioning in an online mode and predict the RUL. Finally, the last proposed model combines time series analysis representation with functional data clustering to detect anomalous cleaning processes of filler machines and classify the normal ones.