Lightweight Unsupervised Model for Anomaly Detection on Microcontroller Platforms

Abstract

Maritime operations are contingent upon the efficiency of the equipment operating on the ship. This bears a considerable impact on not only the overall effectiveness of the vessel, but notably on the safety standards too. An essential cog in this machine is the establishment of effective anomaly detection, which underscores predictive maintenance of the machinery and therein, drastically reduce machine breakdown costs. To this end, this study introduces a groundbreaking approach, utilizing unsupervised learning focused on a low-cost microcontroller unit (MCU) to detect equipment abnormalities via vibration signals. The concentration of the study was towards outliers in vibration signals occurring in multiple machinery, with fans being the key point of our research. We collected an array of accelerometer data using a microcontroller, which was meticulously mounted on a fan for accurate readings. Exploring further, a robust, lightweight unsupervised learning model was developed, trained, and evaluated for precise anomaly detection. The performance of multiple autoencoder architectures with varying complexities was tested and analyzed by measuring the area under the receiver operating curve (AUC), promising active predictive maintenance. The denoising convolutional autoencoder stood out, achieving an impressive AUC of 0.993. Notably, this high-performing model only requires 41 parameters and 3.77 KB of RAM on an Arduino, proving its suitability for deployment on resource-limited edge devices. Compared to previous studies, our models deliver superior anomaly detection while using fewer parameters and computational resources. This research underscores the potential of ultra-lightweight unsupervised learning models to enable accurate and efficient predictive maintenance through on-device anomaly detection with microcontrollers.