Lessons Learned and Practical Solutions for Equipment Health Management in the Chemical Industry

The paper reports lessons learned and practical solutions from a three-year research project for Health Management of an existing fleet of electrical and rotating equipment assets in air separation units (ASUs) operated by Linde Gas. Even though a rich data lake of historical process data from Linde chemical process plants for air separation was available, challenges were encountered with regard to data integration and data quality. Process measurements and fault events were not consistently documented and named across the fleet. Further challenges, typical of this type of initiatives, consisted in the scarcity of historical fault events and heterogeneity across the fleet of plants with regard to plant design, instrumentation, operating and environmental conditions. We report first findings from using Natural Language Processing (NLP) approaches to deal with inconsistent semantics. We further describe a hybrid approach for a recommender system, comprising anomaly detection using autoencoders, statistical feature extraction methods that are robust to fleet variability, and a k-nearest neighbors (k-NN) approach for a human-in-the-loop recommender system. The system leverages historical downtime events, but also continuously learns from new plant outages during operations. Lessons learned and recommendations are provided both for brownfield situations as well as for the engineering and operations of new plants.