Accurate indoor positioning remains a challenge due to the unreliability of satellite signals within building structures. This study presents the development of an indoor positioning system based on WiFi fingerprinting that employs a Multi-Task Neural Network (MTNN) to simultaneously predict both room and floor locations within the IT Building of State Polytechnic of Samarinda. The system was designed to leverage shared feature representations across the two classification tasks, thereby improving generalization compared to single-task approaches. Data were collected at five reference points per room across 12 rooms, yielding approximately 4,000 training samples and 2,000 test samples from 11 access points. A series of experiments evaluated baseline statistical features against engineered features including milliwatt signal conversion, binary integer pattern encoding, and signal dominance metrics. The best-performing model achieved room classification accuracy of 86.75% and floor classification accuracy of 98.56% using a two-layer shared architecture with engineered features. Fine-tuning experiments revealed that while the model could learn new room classes, catastrophic forgetting degraded performance on previously learned classes. A mobile application built with Flutter successfully integrated with a local Flask server for real-time location prediction.
Copyrights © 2026