A TinyML approach to non-repudiable anomaly detection in extreme industrial environments

Unforeseen failures of industrial assets may lead to unexpected downtime with a huge impact on critical business processes. Therefore, modern assets usually include several embedded sensors and processing units, allowing to monitor certain operational parameters, i.e., Condition Monitoring. The sensed data can be later analyzed using Machine Learning (ML) approaches to detect anomalies and anticipate failures. Furthermore, the Internet of Things can provide the tools to extend Condition Monitoring to legacy assets that do not have onboard sensing capabilities. In general, these IoT devices offer the opportunity to move ML processing closer to the monitored asset, thus reducing costs and simplifying the anomaly detection system. However, extreme industrial environments present harsh operating conditions and limited resources, further exacerbated by the reduced computation capabilities of most IoT devices. This paper proposes an ML-based anomaly detection system that uses a retrofitting kit based on a constrained and cost-effective IoT device. Despite its limited resources, the latter executes a state-of-the-art unsupervised anomaly detection algorithm locally, autonomously learning the normality behavior of the monitored asset. Furthermore, to improve the transparency of the monitoring process, we propose to leverage blockchain technology as a non-repudiable repository of information, also assessing the impact of such implementation choice in terms of costs and overhead