Transfer Learning of RSSI to Improve Indoor Localisation Performance

TL;DR

This work tackles the challenge of limited annotated BLE RSSI data for in-home indoor localisation by introducing a Conditional GAN (ConGAN)–based RSSI augmentation framework and a transfer-learning extension (T-ConGAN) to share RSSI information across different homes and protocols. By pre-training on a broad RSSI source (SPHERE) and fine-tuning to target homes, the approach achieves improved room-level localisation, with notable gains for minority classes such as stairs and outdoor areas. The study demonstrates that RSSI data can be effectively transferred across houses, significantly boosting macro F1 scores (up to 12.2% overall) and achieving up to 51% improvements in challenging zones with minimal target data. This cross-house RSSI augmentation enables scalable, low-effort deployment for in-home health monitoring systems, addressing data collection bottlenecks and enhancing robustness in real-world settings.

Abstract

With the growing demand for health monitoring systems, in-home localisation is essential for tracking patient conditions. The unique spatial characteristics of each house required annotated data for Bluetooth Low Energy (BLE) Received Signal Strength Indicator (RSSI)-based monitoring system. However, collecting annotated training data is time-consuming, particularly for patients with limited health conditions. To address this, we propose Conditional Generative Adversarial Networks (ConGAN)-based augmentation, combined with our transfer learning framework (T-ConGAN), to enable the transfer of generic RSSI information between different homes, even when data is collected using different experimental protocols. This enhances the performance and scalability of such intelligent systems by reducing the need for annotation in each home. We are the first to demonstrate that BLE RSSI data can be shared across different homes, and that shared information can improve the indoor localisation performance. Our T-ConGAN enhances the macro F1 score of room-level indoor localisation by up to 12.2%, with a remarkable 51% improvement in challenging areas such as stairways or outside spaces. This state-of-the-art RSSI augmentation model significantly enhances the robustness of in-home health monitoring systems.

Figures (3)

• Figure 1: Transfer learning framework for RSSI augmentation.
• Figure 2: ConGAN architectures for the generator and discriminator. The red text highlights the architectural differences between the ConGAN during the pre-training and transfer processes. All Conv1d and ConvTranspose1d layers use a kernel size of 5, except for the final layer of the discriminator, which uses a kernel size equal to the width of the previous layer's output to produce a single value, indicating whether the RSSI input is real or generated. The red asterisks represent the dynamic width of the Conv1d output, which varies based on the input shape, as both the kernel size and output dimension are fixed.
• Figure 3: Line plots of the actual RSSI (a) compared with augmented RSSI using different methods: (b) Domain-Expert with noise adding, (c) Domain-Expert with periodic dropping, (d) SMOTE, (e) ConGAN, (f) T-ConGAN, and (g) T-ConGAN-SPHERE. The plots consist of RSSI signals from 11 APs in residence B. Each colour represents a different AP, and we compare some example RSSI from Bathroom in a 4s window.