Posters

Abstract

Wireless Sensor Nodes (WSNs) equipped with accelerometers are a powerful and cost-effective means to monitor wind turbine blades for faults and degradations. The main limitation of currently available WSNs is the amount of energy available, usually provided by an energy harvester like a solar cell, limiting the sensing frequency, the amount of computing feasible and the available bandwidth to the base station. Modern ML algorithms provide ample opportunity for advanced blade monitoring schemes. It is, however, non-trivial for a condition-monitoring system (CMS) designer to identify the best still executable blade monitoring algorithm and acceptable sensing frequency under the given power constraints. We believe that wind turbine blade monitoring can be significantly improved by jointly optimizing algorithm used, algorithm partitioning between sensor node and base station as well as ratio between computation and communication on the WSN.  To be able to prototype different ML algorithms for turbine blade fault detection, we built - together with industry partners and using the flex5 simulator - an extensive database of simulated blade vibration signatures under different windloads and operating conditions. We present results from this algorithm design study. To implement the full optimization scheme, we have created an automated workflow starting from high- level monitoring algorithms all the way to actual WSN firmware. For each algorithm, all possible partitions are derived, the code to be executed on the WSN is compiled via ONNX/MLIR/LLVM into a binary and automatically executed on the target platform (STM32 Nucleo64 Development Board with ARM Cortex-M4-NOFP CPU). Its execution time and power consumption is measured (using STM32 Power Shield). The most efficient algorithm partitions are made available in a full-system simulator, allowing exploration of CMS parameters like solar cell size, energy storage capacity, sun irradiance or system duty cycle with respect to CMS availability over a full year. We present the workflow, detail the core components, show results from screening 600+ algorithm partitions and the effect of different CMS design choices. Field test are currently being planned together with commercial CMS designers. Details from the field tests may be presented.