Optimizing Safety Logic Device Selection with Machine Learning: A Classifier-Based Comparison

In Industry 4.0 manufacturing, selecting appropriate safety logic devices during machinery risk assessment is critical yet complex. This study applies supervised machine learning classification to predict safety logic device categories. We assembled an expert-informed dataset of 306 machine cases, each described by safety-related parameters and labelled with one of four target classes (Relay, CR30, GMX, GLX). Using the WEKA toolkit, we evaluated 32 classifiers across four families (rule-based, instance-based, neural network, and tree/forest) using training-set evaluation and 5and 10-fold cross-validation. On the holdout training data, classifiers achieved very high accuracy (average 97.1%), but cross-validation accuracies dropped to only 58–59%, indicating overfitting. Tree/forest ensembles (Random Forest, Random Tree, OptimizedForest) performed best overall (95.9% holdout test, 69.1% 5-fold cross-validation, 68.8% 10-fold cross-validation) compared to rule-based, instance-based, and multi-layer perceptron models. These results suggest that machine learning can effectively guide safety device selection, potentially reducing design time and cost in industrial safety engineering, while highlighting the need for expert oversight and larger datasets.