Secure and Efficient Machine Learning-based Anomaly Detection for IoT-enabled Mechatronic Systems

IoT devices are increasingly used in mechatronic and cyber-physical systems for monitoring, control, and automation. As these systems become more connected, detecting network anomalies and attacks in real time becomes an important security requirement. At the same time, many IoT devices have limited computational and energy resources, which can restrict the use of computationally demanding machine learning models. This study compares machine learning methods for anomaly and intrusion detection in IoT-enabled mechatronic systems. A structured literature review was used to identify relevant approaches, including classical and ensemble learning, deep learning, soft computing, and hybrid reinforcement learning methods. The selected models were evaluated on the CIC-ToN-IoT dataset using binary and multi-class classification tasks. Accuracy, computational cost, and scalability were considered in the comparison. Most models achieved binary classification accuracies above 97%, whereas multi-class accuracy was lower, ranging from 68% to 72%. This difference reflects the difficulty of distinguishing attack types with similar traffic characteristics. Bagged Trees achieved the best overall balance between accuracy and computational cost. Fine Tree and Logistic Regression also provided competitive binary classification results with substantially lower computational requirements. Deep learning and hybrid reinforcement learning methods required more training and inference resources without a comparable increase in accuracy. Overall, the results favor lightweight machine learning methods for real-time anomaly detection in resource-constrained IoT-enabled mechatronic systems.