Context-Enriching Feature Selection Method for Industrial Machine Learning

An industrial process consists of a complex network of thousands of elements interconnected by the flow of material, energy, and information. Each of these elements can be attached to several sensors producing different process signals. Building robust industrial machine learning (ML) models requires handling this multi-dimensional process data while incorporating the process domain knowledge in the feature selection activity. Despite the variety of methods that automate this feature selection task, in industrial practice, they are outperformed with the manual feature selection by domain experts as embedding the industrial domain knowledge. In this paper, we introduce a feature selection method for industrial ML that (1) automatically captures domain knowledge from process topology models that hold information on the plant process and (2) uses it to identify the relevant features (i.e., process signals) to a specified process element to which an ML model is being built. We performed an empirical case study on two industrial use cases to evaluate the effectiveness and efficiency of the proposed method in comparison to existing ones from the literature. In the first use case, our method improved the ML model performance with accuracy of 95% and recall of 91%, compared to a baseline model that achieved 70% and 5% respectively. It also decreased the training time by 81% with a simpler model. In the second use case, our method competed equally with the other methods with regards to model performance, however, it outperformed them in decreasing the training time by 64% with a simpler model.