On-line WSN SoC estimation using Gaussian Process Regression: An Adaptive Machine Learning Approach

Abstract

Wireless sensor networks (WSN) are low-resource devices that run on small batteries. The availability of battery energy, device drive cycles, and environmental conditions all have an impact on node lifetime. The state of charge (SoC) is an important factor in determining the amount of energy available in the batteries. Accurate SoC estimation is critical for device lifetime prediction and safe device operation. We present a novel approach for adaptive SoC estimation based on Gaussian Process Regression in this paper (GPR). The training data was obtained in a climate-controlled laboratory setting by using IEEE 802.15.4-based drive loads at various temperatures for three different batteries such as Lithium-Ion, Nickel-metal hydride, and Lithium-Polymer. To estimate the SoC, battery parameters such as voltage, capacity, and temperature were directly mapped to the corresponding models. For each battery parameter, the GPR model with hyper tuned Radial Bias Filter (RBF) was trained at temperatures ranging from 5 °C to 45 °C. For model accuracy, the proposed scheme was compared to polynomial regression and support vector machines (SVM). In this regard, the proposed model provided Mean Absolute Error (MAE) values of 2.53 percent, 2.54 percent, and 2 percent, respectively, and Root Mean Square Error (RMSE) values of 0.295, 0.292, and 0.35 for Nickel-metal hydride, Lithium-Polymer, and Lithium-Ion batteries at 25 °C. Our proposed lightweight GPR scheme is, to the best of our knowledge, the only active implementation on embedded platforms for SoC estimation of WSN. Finally, the model was rigorously tested on ARM Cortex M4-based microcontrollers to report real-time online SoC estimation on WSN nodes.

Keywords

• Artificial Neural Networks (ANN)
• Energy optimization
• Gaussian Process Regression (GPR)
• Internet of Things (IoT)
• State-of-Charge (SoC) estimation
• Support Vector Machine (SVM)