Efficient Classification of Live Sensor Data on Low-Energy IoT Devices with Simple Machine Learning Methods

This paper addresses the implementation of time series classification on Internet of Things (IoT) devices with minimal resource usage while maintaining required accuracy levels. Although IoT devices have grown more powerful, optimizing resource usage remains essential to reduce costs and extend battery life. This work explores the practical aspects of developing and deploying time series classification models on IoT devices, emphasizing non-neural network models due to their lower resource demands. A detection system utilizing multiple predictions and a voting-based mechanism is constructed to enhance prediction stability. The input data comprises multivariate sensor streams of the same length. All computations are performed on the device, leveraging edge computing principles without external server dependency. This approach demonstrates how to balance accuracy and resource efficiency in IoT-based time series classification, offering a practical framework for similar applications.