Energy-Efficient Wireless Accelerometer Using Hybrid Edge-Central AI for High-Tech Machine Condition Monitoring

Abstract

Machine condition monitoring and predictive maintenance are crucial technologies in modern industrial settings. Wireless sensor networks (WSNs) are commonly used to gather machine data with high flexibility and minimal installation effort. However, traditional WSN approaches that periodically or selectively transmit raw data either lack predictive capability or consume excessive energy. Furthermore, conventional static Edge-AI models running entirely on sensor nodes struggle to adapt to dynamic and complex industrial conditions due to limited labelled failure data and unpredictable machine dynamics. In this paper, we propose and evaluate a hybrid edge-central AI architecture. In this approach, sensor nodes perform the feature extraction as the first layer of the AI model, while deeper adaptive model layers operate at the central base station. This approach reduces energy consumption by limiting radio transmissions and enabling the use of complex, adaptive AI models. We validate the proposed architecture by implementing a set of common features on a typical ARM Cortex-M4 microcontroller used in wireless sensor nodes. We target the architecture of our previously developed wireless 1kS/s (kilo-sample per second) accelerometer. Results demonstrate that these features can be computed in only 32.5 ms and consume 32.43 μW. This represents a significant energy saving compared to raw measurement transmission (686.4 μW), highlighting the effectiveness and feasibility of our hybrid approach for industrial monitoring.