EdgeSL: Edge-Computing Architecture on Smart Lighting Control With Distilled KNN for Optimum Processing Time

Abstract

Our previous research applied a novel classification-integrated moving average (CIMA) method, an intelligence method that improves the performance of passive infrared (PIR) sensors in smart lighting to make control more comfortable for the user. However, intelligence, closely related to cloud deployments with large latency effects, contradicts the real-time nature demanded by smart lighting. This paper proposes edge smart lighting (EdgeSL) architecture, an edge-computing architecture for real-time improved CIMA smart lighting control. We developed novel intelligent control using permutation importance feature selection to improve the CIMA algorithm. Three model compression methods for the k-nearest neighbor (KNN) model are compared, including the novel knowledge distillation on KNN called DistilKNN. The KNN model is the basis of CIMA. The method allows the model to run in an edge-computing environment. The experiment was carried out by evaluating processing time in three different environments, namely edge, fog, and cloud architecture. Our test results show that DistilKNN has the best accuracy compared to other methods, including pruning and quantization, which is 0.93. After deploying the compressed model to the NodeMCU, edge computing has a lower average processing time than fog computing and cloud computing, namely 9.0, 60.1, and 207.6 ms, respectively. After going through the Shapiro-Wilk test, we learned that the three processing times are not normally distributed. So after testing with the Wilcoxon test, it is proved that the EdgeSL has the best performance, where the average processing time with the other architectures has significant differences.

Keywords

• Edge-computing
• model compression
• smart lighting
• knowledge distillation
• k-nearest neighbor