Improved 3D Wireless Indoor Localization with Deep Learning Algorithms

This paper presents a novel deep neural network (DNN) approach for indoor localization problems. In this method, we use the received signal strength indicator (RSSI) values representing the distance measurements as the features and train the DNN. However, due to multipath fading, noise, and the limited dynamic range of the RSSI measures, near-optimal 3D localization based on a path-loss model and multiliterate has become highly challenging. To cope with challenges, we will investigate using Gaussian filters with deep learning (DL) algorithms to improve RSSI-based indoor 3D localization. Gaussian filters filter and classify the RSSI data using DNN algorithms under different conditions. We developed and tested seven different neural network (NN) models to provide a comprehensive analysis. In addition, we performed the grid search algorithms to find the best parameters for the model. Consequently, the hyperparameters were tuned, including the batch size, number of epochs, weight initialization mode, and the optimizer, and the results were presented. According to the results, the best accuracy of 96.2% was given by the net_2s, an NN with two hidden layers containing 1024 and 256 neurons in each layer. This accuracy was obtained by training net_2s for 100 epochs with the Adagrad optimizer. The model was trained with a batch size of 4, and uniform weight initialization was used to initialize the network weights before training.