Multi-Band Wi-Fi Neural Dynamic Fusion

Sorachi Kato, Pu Wang, Toshiaki Koike-Akino, Takuya Fujihashi, Hassan Mansour, Petros T. BoufounosIEEE Trans. Wireless Communications 2024

Abstract

Wi-Fi channel measurements across different bands, e.g., sub-7-GHz and 60-GHz bands, are asynchronous due to the uncoordinated nature of distinct standards protocols, e.g., 802.11ac/ax/be and 802.11ad/ay. Multi-band Wi-Fi fusion has been considered before on a frame-to-frame basis for simple classification tasks, which does not require fine-time-scale alignment. In contrast, this paper considers asynchronous sequence-to-sequence fusion between sub-7-GHz channel state information (CSI) and 60-GHz beam signal-to-noise-ratio (SNR)s for more challenging tasks such as continuous coordinate estimation. To handle the timing disparity between asynchronous multi-band Wi-Fi channel measurements, this paper proposes a multi-band neural dynamic fusion (NDF) framework. This framework uses separate encoders to embed the multi-band Wi-Fi measurement sequences to separate initial latent conditions. Using a continuous-time ordinary differential equation (ODE) modeling, these initial latent conditions are propagated to respective latent states of the multi-band channel measurements at the same time instances for a latent alignment and a post-ODE fusion, and at their original time instances for measurement reconstruction. We derive a customized loss function based on the variational evidence lower bound (ELBO) that balances between the multi-band measurement reconstruction and continuous coordinate estimation. We evaluate the NDF framework using an in-house multi-band Wi-Fi testbed and demonstrate substantial performance improvements over a comprehensive list of single-band and multi-band baseline methods.

Introduction

Integrated Sensing and Communication (ISAC) is a key pillar of future wireless networks, aiming to repurpose Wi-Fi signals for environmental perception. Modern Wi-Fi devices often operate across multiple frequency bands, such as the sub-7 GHz (802.11ac/ax/be) and 60 GHz (802.11ad/ay) bands. These bands offer complementary sensing capabilities: lower frequencies provide rich multipath information via Channel State Information (CSI), while millimeter-wave (mmWave) bands offer stable beam Signal-to-Noise Ratio (SNR) measurements.

Comparison of static frame-to-frame fusion versus continuous sequence-to-sequence estimation.

A major hurdle in multi-band sensing is the inherent asynchrony between measurements. Because different bands operate under distinct protocols and clocks, their sampling rates are mismatched and irregular. Previous work largely focused on "frame-to-frame" fusion for static classification tasks. However, as shown in Figure 1, continuous estimation tasks like trajectory tracking require a "sequence-to-sequence" approach that can handle timing disparities without loss of information.

The Neural Dynamic Fusion (NDF) Framework

To address the asynchrony problem, we propose the Neural Dynamic Fusion (NDF) framework. Unlike standard Recurrent Neural Networks (RNNs) that operate on discrete time steps, NDF models the evolution of latent states as a continuous-time process using Neural ODEs.

The Multi-Band Neural Dynamic Fusion (NDF) architecture.

The architecture, detailed in Figure 2, consists of three primary components:

  1. Asynchronous Encoders: Separate encoders process the CSI and beam SNR sequences. These encoders utilize an ODE solver to "back-propagate" the sequence to a common initial time $t_0$, producing initial latent conditions $z_0^c$ and $z_0^b$.
  2. Latent Dynamic ODEs: These initial conditions are evolved forward in time using a learnable ODE function. This allows the system to query the latent state of any band at any arbitrary timestamp. By querying both bands at the same set of prediction times, the framework achieves latent alignment in a continuous space.
  3. Fusion and Decoding: The aligned latent states are fused and passed to a Coordinate Decoder to reconstruct the $(x, y)$ trajectory. To ensure the latent space remains physically meaningful, auxiliary decoders reconstruct the original CSI and beam SNR measurements.

The entire system is trained using a customized loss function based on the Variational Evidence Lower Bound (ELBO), which balances tracking accuracy with signal reconstruction.

Experimental Evaluation and Results

We evaluated the NDF framework using an in-house multi-band testbed featuring a TurtleBot mobile robot. Ground truth positioning was provided by a high-precision LiDAR scanner. The dataset comprised over 43,000 CSI frames and 9,500 beam SNR frames, reflecting the natural sparsity of mmWave measurements.

The NDF framework achieved a mean localization error of 14.8 cm, a substantial improvement over baseline methods. For comparison, the best-performing multi-band RNN baseline (RNN-Decay) yielded an error of 68.5 cm, while single-band CSI-only methods exceeded 1 meter in error.

Key insights from the evaluation include:

  • Robustness to Sparsity: NDF effectively bridges the gap between high-rate CSI and low-rate beam SNR, leveraging the stability of the latter to refine the resolution of the former.
  • Latent Space Continuity: Visualization of the latent space via t-SNE confirmed that NDF learns a smooth, spatially consistent representation of the environment, unlike discrete models which often produce fragmented clusters.
  • Generalization: The model demonstrated strong performance in scenarios predicting future trajectories and generalizing to unseen spatial regions.

Conclusion

The Multi-Band Neural Dynamic Fusion framework provides a robust solution for asynchronous Wi-Fi sensing. By shifting from discrete-time processing to continuous-time latent dynamics, NDF overcomes the fundamental challenge of protocol-induced asynchrony, paving the way for high-precision ISAC applications.

Summary updated

Abstract

Wi-Fi channel measurements across different bands, e.g., sub-7-GHz and 60-GHz bands, are asynchronous due to the uncoordinated nature of distinct standards protocols, e.g., 802.11ac/ax/be and 802.11ad/ay. Multi-band Wi-Fi fusion has been considered before on a frame-to-frame basis for simple classification tasks, which does not require fine-time-scale alignment. In contrast, this paper considers asynchronous sequence-to-sequence fusion between sub-7-GHz channel state information (CSI) and 60-GHz beam signal-to-noise-ratio~(SNR)s for more challenging tasks such as continuous coordinate estimation. To handle the timing disparity between asynchronous multi-band Wi-Fi channel measurements, this paper proposes a multi-band neural dynamic fusion (NDF) framework. This framework uses separate encoders to embed the multi-band Wi-Fi measurement sequences to separate initial latent conditions. Using a continuous-time ordinary differential equation (ODE) modeling, these initial latent conditions are propagated to respective latent states of the multi-band channel measurements at the same time instances for a latent alignment and a post-ODE fusion, and at their original time instances for measurement reconstruction. We derive a customized loss function based on the variational evidence lower bound (ELBO) that balances between the multi-band measurement reconstruction and continuous coordinate estimation. We evaluate the NDF framework using an in-house multi-band Wi-Fi testbed and demonstrate substantial performance improvements over a comprehensive list of single-band and multi-band baseline methods.
BibTeX
@article{kato2024multi,
  author = {Sorachi Kato and Pu Wang and Toshiaki Koike-Akino and Takuya Fujihashi and Hassan Mansour and Petros T. Boufounos},
  title = {{Multi-Band Wi-Fi Neural Dynamic Fusion}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2024},
}
Figure