On-Sensor TinyML Event-Based Fault Detection Strategies on Wind Turbine Blades

With the advent of wireless low-power Internet of Things (IoT) devices, Structural Health Monitoring (SHM), in which structures or machines are monitored using sensors, has become increasingly important to reduce costs and increase structure longevity. On-sensor machine learning represents a significant leap forward in addressing the challenge of managing the copious amounts of data these SHM systems generate. Novel sensors such as the ISM330IS Inertial Measurement Unit (IMU) from STMicroelectronics feature an onboard Intelligent Sensor Processing Unit (ISPU) enabling some of the required data processing to be computed onboard the sensor, reducing the need to run the main processor and ensuring only interesting data is collected, e.g., when a fault occurs. Using an existing state-of-the-art monitoring system as a baseline, this work investigates the power consumption, bandwidth requirements, and latency of potential data acquisition strategies using an event-based tiny Machine Learning (tinyML) model for fault detection in the context of wind turbine blades. The results demonstrate that replacing the current time-based data acquisition strategy of the existing system with an event-based strategy using the ISPU could reduce power consumption by up to 75% while decreasing latency between fault occurrence and alerting of the remote monitoring system to near-zero, with an always-on power of just 1.29mW.