Machine Learning-Driven Automated Selection of Safety Logic Devices in Industrial Control Systems

Abstract

The design of machine control systems requires the correct selection of safety logic devices to ensure functional safety and compliance with international standards such as EN ISO 13849-1 and IEC 62061. This process is typically based on expert knowledge and manual evaluation of design parameters, which can be time-consuming and error-prone. In this study, machine learning techniques are applied to automate and improve the selection of safety logic devices using real industrial data originating from various types of machinery and automated manufacturing real-world projects. This work introduces a significantly extended
industrial dataset comprising 670 labelled machine configurations derived from real anonymized engineering projects and performs a comprehensive comparison of ten representative ML algorithms implemented in WEKA. The main novelty of the study is a unified large-scale comparative evaluation of heterogeneous machine learning classifiers on real industrial decision data, enabling joint assessment of scalability, generalization, interpretability, and computational efficiency under identical experimental conditions. The results demonstrate that increasing dataset size considerably enhances model stability and generalization. The Averaged 2- Dependence Estimator (A2DE) achieved the highest performance with an accuracy of 86% and Kappa = 0.81, followed by REPTree and Random Forest classifiers. Rule-based methods such as PART and NNge maintained strong interpretability with competitive predictive power. The findings confirm that probabilistic and ensemble algorithms provide reliable and practically applicable solutions for data-driven decision support in industrial safety engineering, paving the way for deployable, explainable, and adaptive decision-support tools in smart manufacturing environments.