Lightweight Feature-based Priority Sampling for Industrial IoT Multivariate Time Series.

Sampling Industrial IoT data streams aims to generate a sample for future data analysis tasks. Several variables influence the efficacy of the constructed sample, including the sampling algorithm and its complexity, the sampling rate selected, and how the sampled data are processed at the gateway. In this paper, we propose a lightweight feature-based priority sampling technique for optimizing Industrial IoT multivariate time series prior to deep learning model classification. The selection of an effective sampling algorithm and rate, coupled with efficient data processing, poses a significant challenge with a key objective of balancing communication overhead reduction and precision maintenance. Our technique minimizes data transmission at the IoT device level, enhancing energy efficiency and improving classification performance by noise reduction through selective feature sampling. Comparative evaluation with existing sampling techniques using a benchmark dataset indicates superior performance in terms of data reduction and classification accuracy trade-offs. Notably, our approach enhances the accuracy of a ResNet model and reduces its processing time.