GreenPipe: Energy-Efficient Data-Processing Pipelines for Resource-Constrained Systems

Billions of resource-constrained systems, such as embedded devices and cyber-physical systems, are in operation worldwide. These systems process input data (e.g., sensor data) into control signals for actuators or human-readable information, thereby providing valuable services and insights. Modern software methods, such as machine learning, have the potential to enhance the performance of these systems even further. However, machine learning is often associated with excessive energy demand, which urgently needs to be resolved. To address this issue, we present GreenPipe, an approach that creates energy-efficient data-processing pipelines tailored for embedded systems known for their low power demand. GreenPipe combines traditional AutoML techniques with energy models and thereby enables the selection of energy-efficient and accurate data-processing pipelines. We implemented GreenPipe on an ARM Cortex-M4 platform and evaluated its performance and energy efficiency. We demonstrate GreenPipe’s capabilities through a comprehensive evaluation, including a practical realworld application for predicting machinery-bearing faults. Green-Pipe demonstrates that it can reduce the energy footprint by up to 90% while maintaining high accuracy.