Energy Characterization And Optimization Of Embedded Data Mining Algorithms

In recent years, internet of things becomes a hotspot. To the internet of things, stream data mining is a key component of it. The internet of things is a network of all embedded devices based on sensors. The large number of embedded devices with sensors will generate a lot of data streams, which set great challenges for embedded devices to do data mining. The most important challenge is for energy efficiency. In stream data mining, the DTW-KNN framework is widely applied for classification in miscellaneous application domains. In this paper, we established an energy measurement testbed. The testbed consists of an embedded computing platform, a Current Conversion and Voltage Acquisition module, and an energy calculation module. The testbed adopt a typical low-power processor ARM STM32 to do stream data mining, while the real-time current and voltage are recorded to calculate the whole energy consumption. Through a case study of the DTW-KNN framework, we investigate multiple existing strategies to improve the energy efficiency without any loss of algorithm accuracy. The results show that the most energy-expensive step is the Dynamic Time Warping, which accounts for 96% on average of the total energy consumption. While k-Nearest Neighbors and normalization accounts for around 3% and 1%. Furthermore, we discuss the influence of three optimization methods with a public-available datasets. The results show that the energy reduction of the full optimization is not a simple addition of energy reduction of each optimization method. Datasets with long sequence gain more energy reduction, while short-sequence datasets have less reduction. The energy reduction of early abandon optimization method varies from 29.5% to 89.9%. Lower bound and indexing DTW method's performance is the most excellent and stable. The work of this paper aims to provide reference for the application of stream data mining on the embedded platform.