Ruhland T, Müller K, Lübke M, Franchi N (2026)
Publication Language: English
Publication Type: Journal article, Letter
Publication year: 2026
DOI: 10.1109/LSENS.2026.3696864
Low-cost MEMS accelerometers are widely deployed for industrial condition monitoring, yet their data streams lack intrinsic quality metrics. This can lead to false alarms caused by sensor artifacts rather than genuine process anomalies, and to diminished trust in condition monitoring systems. This work introduces a lightweight, unsupervised framework for computing a continuous Quality of Sensing (QoS) score. A feedforward autoencoder learns a spectral fingerprint from power spectral density (PSD) and time-domain features during nominal operation – the reconstruction error is then mapped to a normalized QoS parameter. The proposed approach isolates intrinsic MEMS degradation mechanisms. The framework is validated against physically grounded noise simulation: thermo-mechanical noise, vibration rectification error (VRE), and flicker-induced bias instability. The autoencoder detects all three degradation mechanisms, with VRE producing the steepest response. Experiments under laboratory shaker excitation, electric motor vibration, and idle conditions show that real operational data yield the optimal detection sensitivity. The lightweight architecture supports future edge-oriented deployment and improves the trustworthiness of Industrial IoT sensor systems.
APA:
Ruhland, T., Müller, K., Lübke, M., & Franchi, N. (2026). Sensor Data Quality Scoring for MEMS Accelerometers Via Spectral Anomaly Detection. IEEE Sensors Letters. https://doi.org/10.1109/LSENS.2026.3696864
MLA:
Ruhland, Tim, et al. "Sensor Data Quality Scoring for MEMS Accelerometers Via Spectral Anomaly Detection." IEEE Sensors Letters (2026).
BibTeX: Download