Machine Learning-based Classification of Online Industrial Datasets

We aim to incorporate data analytics into industrial process control by utilizing machine learning (ML) algorithms to classify the real-time data of online analyzers. Real-time visualization of results onto a front-end system (i.e., refinery control room) provides an extensive view of the production process, increasing efficiency of production. Selected ML classifiers are assessed according to the performance metrics based on individual scores. These parameters, along with the complexity of implementation, provide an adequate pointer for selecting a suitable classifier model to serve as a decision-making tool. In our use case, accurate categorization of measurements provides a cheap validation guideline that would otherwise be not possible. Computed metrics indicate a difficulty to classify the cases when the slight deviations (drifts) occur from real values. Based on the true positivity rate, linear SVM separation is desirable for data drift prediction (64 %), while k-Means is more successful in detecting outliers (65 %) and normal operation (99 %).