Sensor faults can produce incorrect data and disrupt the operation of entire systems. In critical environments, such as healthcare, industrial automation, or autonomous platforms, these faults can lead to serious consequences if not detected early. This study explores how faults in MEMS microphones can be classified using lightweight ML models suitable for devices with limited resources. A dataset was created for this work, including both real faults (normal, clipping, stuck, and spikes) caused by issues like acoustic overload and undervoltage, and synthetic faults (drift and bias). The goal was to simulate a range of fault behaviors, from clear malfunctions to more subtle signal changes. Convolutional Neural Networks (CNNs) and hybrid models that use CNNs for feature extraction with classifiers like Decision Trees, Random Forest, MLP, Extremely Randomized Trees, and XGBoost, were evaluated based on accuracy, F1-score, inference time, and model size towards real-time use in embedded systems. Experiments showed that using 2-s windows improved accuracy and F1-scores. These findings help design ML solutions for sensor fault classification in resource-limited embedded systems and IoT applications.
Talayoglu, B, Vande Velde, J & da Silva, B 2025, 'Lightweight AI for Sensor Fault Monitoring', Electronics, vol. 14, no. 22, 4532, pp. 1-30. https://doi.org/10.3390/electronics14224532
Talayoglu, B., Vande Velde, J., & da Silva, B. (2025). Lightweight AI for Sensor Fault Monitoring. Electronics, 14(22), 1-30. Article 4532. https://doi.org/10.3390/electronics14224532
@article{a94b858770894159b9ff90d4f892643c,
title = "Lightweight AI for Sensor Fault Monitoring",
abstract = "Sensor faults can produce incorrect data and disrupt the operation of entire systems. In critical environments, such as healthcare, industrial automation, or autonomous platforms, these faults can lead to serious consequences if not detected early. This study explores how faults in MEMS microphones can be classified using lightweight ML models suitable for devices with limited resources. A dataset was created for this work, including both real faults (normal, clipping, stuck, and spikes) caused by issues like acoustic overload and undervoltage, and synthetic faults (drift and bias). The goal was to simulate a range of fault behaviors, from clear malfunctions to more subtle signal changes. Convolutional Neural Networks (CNNs) and hybrid models that use CNNs for feature extraction with classifiers like Decision Trees, Random Forest, MLP, Extremely Randomized Trees, and XGBoost, were evaluated based on accuracy, F1-score, inference time, and model size towards real-time use in embedded systems. Experiments showed that using 2-s windows improved accuracy and F1-scores. These findings help design ML solutions for sensor fault classification in resource-limited embedded systems and IoT applications. ",
keywords = "AI, Machine learning, Fault classification, Sensor faults, MEMS microphone",
author = "Bektas Talayoglu and \{Vande Velde\}, Jerome and \{da Silva\}, Bruno",
note = "Publisher Copyright: {\textcopyright} 2025 by the authors.",
year = "2025",
month = nov,
day = "19",
doi = "10.3390/electronics14224532",
language = "English",
volume = "14",
pages = "1--30",
journal = "Electronics",
issn = "2079-9292",
publisher = "MDPI AG",
number = "22",
}