Interpretable Machine Learning Framework for Quality Control in Resource-Constrained Industrial Settings

Machine Learning (ML) is currently seen as a major driver for industrial process digitalisation. Its ability to process large volumes of data, extract complex patterns in data, and achieve high performances, make it a viable solution to the more traditional, rule- and knowledge-based systems, which lack adaptability and scalability to the increasing complexity in industrial processes. However, despite general interest from industry, many manufacturing settings still rely on older hardware and software solutions, incompatible with ML implementations. Moreover, and given that the best performing ML models are often “black-boxes”, user trust in their outputs is still limited. This paper proposes and validates a ML-based framework for enhanced leak detection, developed in collaboration with Bosch Termotecnologia S.A. The solution is designed for constrained environments and targets implementation on PLC-controlled systems. Several interpretable ML models, translatable to simple logic and arithmetic operations were evaluated. Results showcase that multiple classifiers achieve MCC scores above 0.95 using only a subset of the full dataset, enabling a 40% reduction in overall leak detection time compared to the currently employed approach. Additionally, an interactive dashboard was developed using a Decision Tree model to support real-time monitoring and enhance interpretability for non-expert users.