Machine Learning for Process Optimization

Machine learning has been successfully applied in various domains, where it is utilized to classify messages as spam or not, detect faces in images, or identify individuals from facial images. Despite the broad range of applications that can be tackled using machine learning, the most natural use case for machine learning is in data-rich environments, where massive amounts of data are available for the training and evaluation of models. The process industries have only recently started to benefit from data-driven techniques, as these industries have historically relied on the development of physics-based models for the prediction of process behavior.

Although some process systems are highly controlled and automated, enabling the online acquisition of large amounts of process data, these data often remain unused, with the primary function being archiving. In addition, the Internet of Things (IoT) is enabling process systems to become equipped with more sensors and data-acquisition tools than ever before, increasing the opportunity for the development of data-driven diagnostic, fault-detection, and fault-tolerant mechanisms. With the advancement of affordable computing and storage resources, data are being generated and stored at an unprecedented level, further contributing to the development of data-driven applications. Nevertheless, it remains a challenge for engineers to find the balance between leveraging data effectively while also maintaining a deep physics-based understanding of the underlying process.