(9b) Machine Learning Assisted Preventive Maintenance

ABSTRACT

Majority of manufacturing and chemical plants rely on the knowledge of experienced maintenance staff to plan, schedule, implement, and even predict routine and emergency maintenance needs to avoid unplanned downtime. As many of these knowledgeable staff members begin to reach retirement age, there is a growing demand to document, standardize, and accurately time preventive maintenance, instead of following a predetermined time schedule. This demand is increased by the growing availability of different techniques, such as ultrasonic, vibration and heat analysis, to help predict failures.

The staggering amount of data for predicting maintenance has prompted this research by creating a need for a suite of machine learning methods to analyze and utilize all the different inputs available to predict failures. The present study uses natural language processing (NLP) to read scheduled maintenance plans and analyze maintenance staff notes on how equipment has failed, and what corrective measures were taken afterwards. This information is coupled with time series analysis of data from the time leading up to that failure from the ensemble of sensors in the plant, to define key predictors of the failure using Self Organizing Maps (SOMs) and k-Nearest Neighbors (KNN). A deep neural network (DNN) is trained on the observed trends of key predictors using a combination of Long Short Term Memory (LSTM) and Convolutional Neural Networks (CNN) and system health (healthy/fault prone/failing/failed). The trained system is then used against real-time data from the sensors to predict an upcoming failure, its severity, and prescribe possible remedial actions ahead of failure.

The results from this project showed that many common equipment failures can be defined and predicted by coupling together data from many different inputs surrounding the equipment. The prediction accuracy also improves as more data becomes available to train the model, and when more sensors and inputs are available. The results of this project should encourage industry reliability teams to increase the sensors available for critical equipment and to apply machine learning techniques to handle all the inputs for improved failure prediction.