Previous Projects

In recent years, real-time radio-frequency (RF) signal classification has become essential across diverse applications such as radar-based human activity recognition (HAR), RF waveform classification, and wireless communications. Traditionally, RF classification relies on a two-step approach involving computationally expensive transformations (like spectrogram generation), followed by classification using deep neural networks (DNNs). However, these methods face challenges such as high computational latency, limited interpretability, and inefficiency in processing complex-valued RF signals directly.

To overcome these limitations, we introduce PLFNets, a novel deep learning framework specifically designed for direct classification of complex-valued raw IQ radar data, using parameterized learnable filters (PLFs).

Key Contributions:

• Parameterized Learnable Filters (PLFs):

  • Developed a generalized and flexible filter-learning framework utilizing structured functions including:

    • Sinc

    • Gaussian

    • Gammatone

    • Ricker Wavelets

  • Each filter function is parameterized by physically meaningful attributes such as center frequency and bandwidth, facilitating interpretability.

• Complex-Valued Neural Network Architecture:

  • Introduced a fully complex-valued CNN structure that directly processes raw IQ radar signals.

  • Integrated PLFs as the initial layer, significantly reducing computational complexity and latency by directly learning from the signal’s raw spectral features.

• Enhanced Interpretability:

  • Learned filter parameters provide clear physical interpretations (e.g., identifying critical frequency bands), enabling better understanding and trust in the model’s decisions.

Performance Evaluation and Results:

• Datasets:

  • Evaluated extensively using both synthetic RF modulation waveforms (LFM, Barker, Frank, Costas, Rectangular) and a real-world American Sign Language (ASL) radar dataset.

• Accuracy and Efficiency:

  • Achieved an average of 47% improvement in accuracy over conventional 1D CNNs on raw RF data.

  • Delivered approximately 7% improvement over CNNs using real-valued learnable filters.

  • Matched or exceeded the accuracy of computationally intensive 2D CNN models applied to micro-Doppler spectrograms, while reducing computational latency by around 75%.

• Computational Advantages:

  • Demonstrated significantly lower training and inference times, making PLFNets highly suitable for real-time and near real-time RF sensing applications.

Impact and Applications:

The proposed PLFNets framework sets a new standard for computational efficiency, interpretability, and accuracy in radar-based applications. It is especially valuable for:

• Radar-based Human Activity Recognition (HAR)

• RF waveform modulation recognition

• Real-time RF sensor deployments in smart environments

• Edge computing platforms for immediate classification tasks

Future Directions:

Further research aims to extend the PLFNets framework to:

• Real-time embedded system implementations.

• Address complex multi-signal scenarios.

• Expand application domains, including joint radar-communication systems and advanced wireless communications.

Through PLFNets, the integration of deep learning and radar signal processing reaches a new level, offering an innovative pathway towards efficient, real-time, and interpretable RF signal analysis.

The PLFNet Architecture

This project showcases a novel deep learning framework designed to reconstruct high-resolution time-frequency representations from radar signals. These signatures are crucial for applications like human activity recognition (HAR). Here's a breakdown of the key ideas:

Motivation and Challenges:
Traditional methods, such as the Short-Time Fourier Transform (STFT), require a trade-off between time and frequency resolution. They are also highly sensitive to noise and demand extensive parameter tuning, which can limit their effectiveness in practical, real-world scenarios.

The HRSpecNet Approach:
HRSpecNet tackles these challenges with a novel, three-stage deep neural network architecture that directly converts 1D complex radar signals into high-quality 2D time-frequency representations:

• Autoencoder Module:
  This part of the network works to denoise the input signal. By learning a compact representation, it helps to remove interference and improve the overall signal quality.

• Convolutional STFT Module:
  Instead of using a fixed window length like the traditional STFT, this module uses learnable convolutional filters. It creates multiple intermediate frequency domain representations, adapting to various signal characteristics.

• U-Net Module:
  Finally, the U-Net fuses the features from the previous block to produce a high-resolution spectrogram. Its design ensures that subtle details—such as the intersection of frequency components—are preserved, even in noisy conditions.
  
  

Performance and Validation:
The HRSpecNet framework was evaluated on both synthetic signals and real-world data (including a challenging dataset of American Sign Language gestures). The results demonstrated:

• Improved Resolution and Noise Robustness:
  HRSpecNet produces clearer, sparser, and more accurate time-frequency representations compared to traditional STFT and other deep learning approaches.

• Enhanced Classification Accuracy:
  When the generated micro-Doppler signatures were used for classifying human activities, HRSpecNet achieved higher accuracy, improving performance by over 3% relative to state-of-the-art methods.

• Computational Efficiency:
  Despite its advanced capabilities, HRSpecNet maintains high computational efficiency, making it a practical choice for real-time applications.

Impact:
By effectively addressing the limitations of conventional techniques, HRSpecNet opens up new possibilities in radar-based HAR applications—from improved safety in automotive systems to enhanced performance in healthcare monitoring and security systems. In summary, HRSpecNet represents a significant step forward in leveraging deep learning for radar signal processing, providing robust, high-resolution micro-Doppler signatures that enable more accurate and efficient human activity recognition.

This study fills a research gap by directly comparing RF sensors for real-time human motion recognition in applications like human-computer interfaces and smart environments, using datasets collected from a 77 GHz mmWave FMCW TI Radar and a Raspberry Pi 3B+ equipped with Nexmon firmware for Wi-Fi. Spectrograms are generated from both systems under the same scenarios, and the performance is evaluated in a 7-class human activity recognition (HAR) task. The results show that radar outperforms Wi-Fi by 32.7% in accuracy, underscoring the superiority of radar technology in HAR while also highlighting Wi-Fi’s potential for indoor activity monitoring. Additionally, both the datasets and the associated code have been made publicly accessible to support further research in this emerging field.