WiFi Sensing on the Edge: Signal Processing Techniques and Challenges for Real-World Systems

Hernandez, Steven; Bulut, Eyuphan

doi:10.1109/comst.2022.3209144

Cited by 31 publications

(14 citation statements)

References 142 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Performing machine learning model inference on-board low power edge devices has recently achieved greater attention for tasks such as device-free wireless sensing [7], computer vision [8], and numerous other tasks [9]. At the core of edge model inference is achieving model compression for use on low power and low resourced devices.…”

Section: Related Workmentioning

confidence: 99%

Sharing Low Rank Conformer Weights for Tiny Always-On Ambient Speech Recognition Models

Hernandez¹,

Zhao²,

Ding³

et al. 2023

Preprint

View full text Add to dashboard Cite

Continued improvements in machine learning techniques offer exciting new opportunities through the use of larger models and larger training datasets. However, there is a growing need to offer these new capabilities on-board low-powered devices such as smartphones, wearables and other embedded environments where only low memory is available. Towards this, we consider methods to reduce the model size of Conformer-based speech recognition models which typically require models with greater than 100M parameters down to just 5M parameters while minimizing impact on model quality. Such a model allows us to achieve always-on ambient speech recognition on edge devices with low-memory neural processors. We propose model weight reuse at different levels within our model architecture: (i) repeating full conformer block layers, (ii) sharing specific conformer modules across layers, (iii) sharing sub-components per conformer module, and (iv) sharing decomposed sub-component weights after low-rank decomposition. By sharing weights at different levels of our model, we can retain the full model in-memory while increasing the number of virtual transformations applied to the input. Through a series of ablation studies and evaluations, we find that with weight sharing and a low-rank architecture, we can achieve a WER of 2.84 and 2.94 for Librispeech dev-clean and test-clean respectively with a 5M parameter model.

show abstract

Section: Related Workmentioning

confidence: 99%

Sharing Low Rank Conformer Weights for Tiny Always-On Ambient Speech Recognition Models

Hernandez¹,

Zhao²,

Ding³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…The fact that ESP32 is an IoT platform that can be found in many daily used IoT devices, the low cost as compared to the other competitors and its compatibility with TFLite for TinyML implementation all make it a good candidate that fulfils the conditions for real-time applications with wide adoption. A comprehensive survey on the use of ESP32 for DFWS-enabled edge can be found in [17].…”

Section: A Csi Extraction Toolsmentioning

confidence: 99%

“…A windowed moving average filter can be used to smooth the time series signal of the CSI retracted amplitude. Also, the Discrete Wavelet Transform (DWT) has been used widely to remove unwanted white Gaussian noise by calculating the highfrequency energy of the CSI [4], [17]. Median filter [4], Butterworth filter with a certain cutoff frequency to remove high-frequency noises [79] are also some other filtering techniques.…”

Section: Challenges Of Dfws On the Edgementioning

confidence: 99%

“…Median filter [4], Butterworth filter with a certain cutoff frequency to remove high-frequency noises [79] are also some other filtering techniques. Also, the Hampel filter is a good candidate for denoising the signal, especially when retaining the original signal is important after the transformation [17], [80]. Some CSI data analysis also requires training on the spectrogram of the sub-carriers CSIs.…”

Section: Challenges Of Dfws On the Edgementioning

confidence: 99%

“…Feature extraction techniques may result in simpler recognition algorithms and in some cases, we do not need complex recognition techniques like DNNs. Statistical analysis of the CSI measurements (i.e., amplitude, phase, temporal differences and power spectral density) such as mean, median, variance, correlations, kurtosis, and other moments may contain sufficient data to build the recognition decisions and hence the raw features are transformed into aggregated single or few features that are rich of information [1], [17], [22].…”

Section: Challenges Of Dfws On the Edgementioning

confidence: 99%

See 2 more Smart Citations

Channel State Information based Device Free Wireless Sensing for IoT Devices Employing TinyML

Hayajneh

Aldalahmeh

Zaidi

et al. 2022

2022 4th IEEE Middle East and North Africa COMMunications Conference (MENACOMM)

View full text Add to dashboard Cite

The channel state information (CSI) of the sub-carriers employed in orthogonal frequency division multiplexing (OFDM) systems has been employed traditionally for channel equalisation. However, the CSI intrinsically is a signature of the operational RF environment and can serve as a proxy for certain activities in the operational environment. For instance, the CSI gets influenced by scatterers and therefore can be an indicator of how many scatterers or if there are mobile scatterers etc. The mapping between the activities whose signature CSI encodes and the raw data is not deterministic. Nevertheless, machine learning (ML) based approaches can provide a reliable classification for patterns of life. Most of these approaches have only been implemented in lab environments. This is mainly because the hardware requirements for capturing CSI, processing it and performing signal-processing algorithms are too complex to be implemented in commercial devices. The increased proliferation of IoT sensors and the development of edge-based ML capabilities using the TinyML framework opens up possibilities for the implementation of these techniques at scale on commercial devices. Using RF signature instead of more invasive methods e.g. cameras or wearable devices provide ease of deployment, intrinsic privacy and better usability. The design space of device-free wireless sensing (DFWS) is complex and involves device, firmware and ML considerations. In this article, we present a comprehensive overview and key considerations for the implementation of such solutions. We also demonstrate the viability of these approaches using a simple case study.

show abstract

Data Augmentation Techniques for Cross-Domain WiFi CSI-Based Human Activity Recognition

Strohmayer,

Kampel

2024

IFIP Advances in Information and Communication Technology

View full text Add to dashboard Cite

WiFi Sensing on the Edge: Signal Processing Techniques and Challenges for Real-World Systems

Cited by 31 publications

References 142 publications

Sharing Low Rank Conformer Weights for Tiny Always-On Ambient Speech Recognition Models

Sharing Low Rank Conformer Weights for Tiny Always-On Ambient Speech Recognition Models

Channel State Information based Device Free Wireless Sensing for IoT Devices Employing TinyML

Data Augmentation Techniques for Cross-Domain WiFi CSI-Based Human Activity Recognition

Contact Info

Product

Resources

About