STATISTICAL ANALYSIS AND PREDICTION MODEL FOR THROUGHPUT IN WIRELESS MULTIMEDIA SENSOR NETWORK USING DEEP LEARNING WITH LONG SHORT TERM MEMORY ARCHITECTURE

• Main Author: Arfan Wicaksono, Mokhamad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/61879
• Institution: Institut Teknologi Bandung

The escalation of Internet of Things (IoT) technology is spreading throughout many devices that are used to solve problems in society. Multimedia IoT (M-IoT) is a current IoT trend that focuses on applying multimedia sensors and directly involve multimedia data. One of the approaches to implement infrastructure-layer M-IoT is through Wireless Sensor Network (WSN). The performance of WSN can be assessed by measuring the communication throughput. The current technology enables us to collect communication throughput data in WSN for particular needs. Some opportunities encountered from the collected data are to analyze it in order to have an understanding of the data and to make a prediction model for it. In this research, the throughput data analysis is conducted in a multimedia WSN to discover the characteristic of it. Then, the deep learning model is designed to predict future throughput by referring to Long Short Term Memory (LSTM) architecture. The WSN system that is used in this research is Software-Defined Network (SDN) based with modified Constrained Application Protocol (CoAP) by adding sequence length parameter as a communication protocol and with ability provided by the controller to receive throughput information. The influence of testing parameters, i.e., number of devices and sequence length, towards throughput is analyzed. Then, the regression model is implemented to predict the throughput value based on the testing parameter. The regression model is combined with a deep learning model with various complexity levels and referring to manyto- many LSTM architecture. The evaluation is conducted by measuring accuracy and execution duration result in the implementation. The result shows that the performance is slightly increased by combining the regression model with the deep learning model, the more complex architecture shows better results in accuracy but worse in execution time, and underfitting is still indicated in all the models.