(300d) Ready-to-Use Operational Machine Learning in the Process Industry

In this talk, we present a machine learning approach that empowers the industrial practitioner to quickly solve their domain specific big data problems, and utilize the operational insight gained to improve safety, quality, and reduce downtime. This approach provides visibility and control in the hands of the practitioners who understand the operational processes and equipment (without requiring a team of data scientists). The practitioners feed the machine learning system with time series operational data, and based on the predictive insights and alerts, determine the corrective action to take. An important advantage of this approach is “continuous learning” - the practitioner can use new learnings and events to adapt, retrain, and refine the model online.

The application of this “data scientist in a box” approach in the process industry is presented through three case studies. The first case study describes the use of Falkonry for proactive health monitoring of critical equipment by a global industrial equipment manufacturer. The model was trained on operational data with identified fault conditions. The framework was integrated with their MES system issuing alerts to operations as fault conditions were detected. The second case study focused on early detection of a stream being off spec in a refinery. The model was trained on both normal as well as identified upset conditions across processing units that led to the stream being off spec. We show how the model was able to provide advance warning of the off-spec condition, which would have otherwise resulted in expensive storage and reprocessing costs for the refinery. The third case study highlights the application of Falkonry by a global semiconductor manufacturer for improving throughput through predictive maintenance of fab tools. Predictive maintenance of fab tools can reduce unexpected downtime and yield substantial benefits over the lifetime of the asset. For each of the case studies, details of the implementation, specific fault scenarios predicted, as well as organizational learnings from the implementation are discussed.