TL;DR
This paper introduces WK-Pnet, a novel FM signal-based positioning system that uses Wavelet Packet Decomposition (WPD) to extract rich time-frequency features and knowledge distillation to create a lightweight, accurate model. This approach significantly reduces computational load and inference latency, making it ideal for resource-constrained IoT applications.
What Problem Does This Paper Address?
Accurate and efficient indoor and urban positioning is crucial for many applications, but traditional satellite-based systems like GPS struggle in these environments due to signal blockage and interference. While opportunistic signal positioning using ambient radio frequency (RF) signals offers a promising alternative, existing FM-based methods often face challenges: some require precise time synchronization and complex signal processing, while others, particularly those relying on deep learning, can be computationally demanding, hindering their deployment on resource-limited devices. This paper aims to bridge this gap by developing a highly accurate and computationally efficient FM positioning system.
Methods
The core of WK-Pnet lies in two key innovations:
- Wavelet Packet Decomposition (WPD): Instead of traditional time-frequency analysis methods like STFT, WK-Pnet employs WPD to decompose FM signals. WPD offers finer frequency band divisions and allows for multi-scale analysis, effectively capturing subtle signal variations and local features. This results in a richer and more informative representation of the FM signal for the positioning model.
- Knowledge Distillation: To address the computational burden of deep learning models, WK-Pnet utilizes knowledge distillation. A complex, high-capacity "teacher" model (based on ResNeXt) is trained first. Then, its learned knowledge (soft labels) is transferred to a smaller, more efficient "student" model. This process allows the student model to achieve performance comparable to the teacher model while being significantly lighter and faster, making it suitable for edge devices.
The overall framework involves an offline training phase where FM signals are collected and used to train the teacher and then the student model, and an online inference phase where the trained student model estimates the target's position.
Datasets
The paper utilizes a dataset previously introduced in [33]. This dataset comprises FM signals collected over three consecutive days from various fixed locations, covering both indoor and outdoor environments. The signals were captured at a center frequency of 97.5 MHz with a bandwidth of 4 MHz and a sampling rate of 5 Msps.
Key Insights and Ideas
- WPD's Superiority for FM Signals: The experiments reveal that WPD, compared to STFT and EMD, provides a more robust and accurate time-frequency representation for FM signals, especially in challenging outdoor environments. This is attributed to its ability to capture finer details and adapt to signal characteristics.
- Knowledge Distillation for Efficiency: The study strongly validates the effectiveness of knowledge distillation. The student model, trained with distilled knowledge, significantly outperforms a student model trained without it, demonstrating that efficiency can be achieved without sacrificing substantial accuracy.
- Lightweight Models for Edge Deployment: The research highlights the critical need for computationally efficient models in modern positioning systems. WK-Pnet's student model, with its reduced FLOPs and parameters, achieves dramatically lower inference latency, making it practical for real-time applications on smartphones and other edge devices.
- Haar Wavelet Basis is Optimal for Indoor: While different wavelet bases were tested, the Haar wavelet basis proved to be the simplest and most effective for indoor positioning accuracy and stability, offering the best balance of performance and computational simplicity.
Key Results
- Accuracy: WK-Pnet achieves high positioning accuracy, with indoor MDE as low as 0.0269m and outdoor MDE around 0.7829m (using the Teacher Model). The Student Model also demonstrates competitive accuracy.
- Efficiency: WK-Pnet significantly reduces computational load and inference latency. Compared to FM-Pnet, WK-Pnet reduces FLOPs by approximately 95.9% and parameters by about 99.3%. Inference latency is reduced by 90.5%.
- Time-Frequency Transformation Impact: WPD consistently outperformed STFT and EMD in terms of positioning accuracy and stability, particularly in outdoor scenarios.
- Wavelet Basis Impact: The Haar wavelet basis yielded the best indoor performance, while the Daubechies wavelet showed the best outdoor results. The Haar basis was chosen for its simplicity and computational efficiency.
- Knowledge Distillation Impact: The student model with knowledge distillation showed improved accuracy and stability compared to a student model without it, though slightly lower than the teacher model.
Limitations and Future Work
- Limitations: While WK-Pnet shows excellent performance, the paper notes that the Student Model's positioning stability outdoors is slightly lower than that of the Teacher Model. The performance of both methods can decline over consecutive days, suggesting potential environmental drift or signal variations.
- Future Work: The authors suggest exploring the combination of FM signals with other signal sources to further enhance positioning accuracy. They also propose investigating few-shot and self-supervised learning approaches to reduce the reliance on extensive labeled data and labeling costs.
Summary updated