Deep Learning-Based Anomaly Detection Using MEMS Sensors

Reliable condition monitoring during operation is of paramount importance for predictive maintenance. However, implementing such systems on compact and resource-constrained hardware remains challenging. This work presents a microcontroller-based sensing device that integrates MEMS motion and environmental sensors for real-world anomaly detection in mobile robots. The device operates at an output data rate of 250 Hz and streams field data wirelessly during operation. A dataset of 140 complete runs was collected, including five types of abnormal wheel operating conditions. A deep learning model based on the MLSTM-FCN architecture was trained on the raw sensor data to classify the robot's operational condition. The model achieved an overall accuracy of 94.4 % and a macro-F1 score of 0.93 on the test set. The results show that MEMS-based sensing combined with deep learningbased methods enables accurate and reliable anomaly detection on low-power embedded platforms. This provides a practical foundation for future predictive maintenance applications in industrial and robotic systems.

Keywords: Embedded Electronics, MEMS Sensors, Anomaly Detection, Predictive Maintenance, Deep Learning.