Artificial Neural Networks to Analyze Energy Consumption and Temperature of UAV On-Board Computers Executing Algorithms for Object Detection

Abstract

When incorporating object detection models into unmanned aerial vehicles (UAVs) on-board computers, two aspects are relevant aspects. Firstly, the energy consumption required by the computer on board the UAV during the mission, since low-cost electric UAVs currently have low flight autonomy. Moreover, during the mission, the computer’s processor may suffer overheating caused by the running algorithm, which may directly impair the continuity of a given task or burn the computer. In this study, we aim to estimate the energy consumption and make temperature predictions of a computer embedded in UAVs for missions involving object detection. We propose a method, Analyzing Energy Consumption and Temperature of On-board computer of UAVs via Neural Networks (ETOUNN), which uses a multilayer perceptron (MLP) network to estimate the energy consumption and a long short-term memory (LSTM) network for predicting temperature. Our experiment relied on a Raspberry Pi 4 8 GB computer running nine popular models of object detectors (deep neural networks): eight of which are pre-trained models of the YOLO family, and one Mask R-CNN network. Regarding energy consumption, we compared our method to multivariate and simple regression-based on two metrics: mean squared error (MSE) and the $R^2$ regression score function. As for temperature prediction and considering the same metrics, ETOUNN was compared to the Autoregressive Integrated Moving Average (ARIMA), the Neural Basis Expansion Analysis for interpretable Time Series forecasting (N-BEATS), and a gated recurrent unit (GRU) network. In both comparisons, our method presented superior performances, showing that it is a promising strategy.