Abstract
Due to an increase in Internet of Things (IoT) based applications and location-based services amongst millions of devices and gateways over the network, a lot of research is being conducted in analyzing various wireless positioning methods in Low Power Wide Area Network (LPWAN). Fingerprinting based Localization is one such technique with short-range radio frequency and addresses problems with multi-path in indoor positioning. With the increase in complex infrastructure layout and millions of devices being connected, fingerprinting approach has challenges like spatial ambiguity, long distances, low-bandwidth, scalability, cost and size constraints. Considering these challenges and the problem at hand, this work aims at predicting accurate location using deep neural networks, for the data collected using LoRaWAN communication protocol. LoRa, a low power wide-range technology has been used to collect signal strength data at different locations from the sensors and gateways, in an indoor university building. Indoor localization using deep neural networks has achieved 1.2-2.0 [m] of mean distance error to predict accurate latitude and longitude. Additionally, to validate the comprehensive approach, a publicly available outdoor data-source collected in the city of Antwerp, Belgium is used to minimize localization error. The approach uses the neural network algorithms like Artificial Neural Network (ANN), Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) in comparison with their traditional k-NN approach. The results have demonstrated that mean distance error of 191.52 [m] from LSTM has out-performed results from the k-NN algorithm. From the results, LSTM estimates sequential data more accurately, based on the user’s current and previous position in the linear trajectory. Analysis of different hyper-tuning parameters of DeepML has helped to optimize the results, while reducing problems like vanishing gradient descent and overfitting in an indoor dataset. The whole approach has been implemented in different hardware accelerators like GPU and TPU in a python-based framework called TensorFlow. Google Colab has been used to train the models. The statistical results have been plotted using Matlab. Hardware devices like Dragino LoRa gateway, Arduino, GPS LoRa shield and a DHT11 sensor have been used in this study.