Energy-Efficient Implementation of Explainable Feature Extraction Algorithms for Smart Sensor Data Processing
Julian Schauer, Payman Goodarzi, Andreas Schütze, Tizian Schneider · 2024
With the rising demand for smart sensor technologies in data-driven Machine Learning applications such as predictive maintenance and condition monitoring, the challenge lies in creating energy-efficient systems that can run self-sufficiently. This paper presents an approach to implement explainable feature extraction algorithms as Deep Neural Networks (DNN) to meet energy-efficient implementation requirements. A hardware-level evaluation of the DNN representation is demonstrated and compared to common feature extraction methods, and significant improvements are highlighted. With the new approach, the energy consumption of the feature extraction methods decreased by a minimum of 78.3 % compared to the standard implementation.