(91e) Investigating Machine Learning Techniques for Effective Predictive Maintenance in Industrial Systems

With the advent of Industry 4.0, more advanced maintenance strategies are being used. Among these practices is predictive maintenance, which predicts when faults will occur in unit operations and identifies those faults and their root causes ahead of time. Using this information can enable maintenance scheduling to be optimised, process safety to be improved, and a reduction in downtime to improve resource allocation over conventional preventative strategies. Combined with the advancement of cyber-physical systems (CPS), industrial data is becoming ever more readily available, making predictive maintenance increasingly feasible. Nonetheless, additional research is required to extend current predictive strategies that only analyse single-unit operations, failing to consider the underlying links between sequential unit operations of a full industrial process. In this study, the objective was to perform a systematic review of different Machine Learning (ML) techniques to predict and classify future sensor data by taking a multivariate approach, which incorporates relationships between multiple unit operations across an entire industrial process. We aimed to leverage this additional data gained, to increase the accuracy of these predictions, as well as develop a more effective predictive methodology. This was accomplished using data from a continuous powder to tablet processing pilot plant, composed of 9 unit operations and equipped with a total of 47 sensors, mimicking the scale of industrial systems. The effectiveness of different ML models for forecasting this sensor data of degrading unit operations was evaluated over short and long-term time periods up to a point of failure. The study found that Long-Short-Term-Memory Neural Networks (LSTMs) produce the best forecasting results, capable of predicting varying intricate patterns of degradation months in advance. In contrast, more traditionally used linear and naive models fail to capture the complexities of the system, resulting in faults being predicted either too early or too late, corresponding to money being wasted on unnecessary maintenance. Using the present research, we develop the foundation for more sophisticated predictive maintenance, effectively forecasting data which can be used to pre-emptively mediate faults in an industrial process, helping us take the next step towards Industry 4.0.