Knowledge Distillation for a Lightweight Deep Learning-Based Indoor Positioning System on Edge Environments

Deep learning has become an attractive choice for indoor positioning systems (IPS) because of its ability to be more accurate in dealing with phenomena such as multipath and shadowing. However, mobile and edge devices in an IPS environment usually have limited resources, making deep learning deployment challenging. Our research aim is to implement knowledge distillation for a deep neural network (DNN)-based IPS that is lightweight in an edge environment. The first step in our research is to collect a dataset regarding IPS, which contains distance data to three beacons as well as the position of the mobile phone in the room and determines the location based on the two outputs, namely Position X and Position Y. Then we design knowledge distillation for DNN to realize the compression model. We prove our compressed model by comparing the student model with the teacher model and the student model without the teacher model. We use several test metrics, including accuracy, loss, number of model parameters (Params), and compression ratio (CR). The test results show that the accuracy of the student with the teacher model outperforms the student without the teacher model, namely 0.8 compared to 0.776 for Position X and 0.96 compared to 0.88 for Position Y. Our research concludes that our knowledge distillation model for IPS can perform compression with $\mathrm{CR}=10.2 \times$ while having better performance than the model without knowledge distillation.