Real‐Time Detection of Undesired Human Interventions in Robotic Work Cells Using a Convolutional Neural Network‐Based Novel Architecture and Reliability Analysis With Explainable Artificial Intelligence

ABSTRACT This study presents a task‐specific deep‐learning framework for real‐time detection of undesired human interventions in robotic work cells, based on a customized convolutional neural network (CNN) architecture termed MerdanNet. Unlike general‐purpose lightweight models, MerdanNet integrates progressive dropout, systematic batch normalization, and compact hierarchical depth to ensure reliable performance in low‐data, safety‐critical environments. The data set comprises 588 original images of human hand, upper body, and foot categories, expanded to 2940 samples through targeted augmentation (mirroring, grayscale transformation, and controlled noise). Three models, MerdanNet, YOLOv8, and MobileNet, were trained and evaluated using accuracy, precision, recall, and F 1‐score metrics; MerdanNet achieved the highest performance, with an accuracy of 98.76%. Beyond static evaluation, the framework was validated in a closed‐loop robotic setup where detections directly triggered safety actions. Experiments across 90 trials confirmed consistent activation of slowdown and emergency stop functions with an average end‐to‐end latency of 92 ms (95th percentile: 124 ms), well within industrial safety thresholds. Interpretability was assessed using Grad‐CAM and LIME, which revealed meaningful attention patterns and provided diagnostic insights, though quantitative explainable artificial intelligence (XAI) evaluation remains a target for future work. While the data set is limited in diversity, the study highlights this as a current limitation and outlines future directions, including expanded data collection, synthetic stress‐test data sets, bounding box annotations for detector benchmarking, and transfer learning approaches. Overall, the findings demonstrate that combining tailored CNN architectures with XAI and closed‐loop validation can yield deployable, transparent, and robust safety modules for industrial robotic environments.