PLFNets: Interpretable Complex-Valued Parameterized Learnable Filters for Computationally Efficient RF Classification

Abstract

Radio frequency (RF) sensing applications such as RF waveform classification and human activity recognition (HAR) demand real-time processing capabilities. Current state-of-the-art techniques often require a two-stage process for classification: first, computing a time-frequency (TF) transform, and then applying machine learning (ML) using the TF domain as the input for classification. This process hinders the opportunities for real-time classification. Consequently, there is a growing interest in direct classification from raw IQ-RF data streams. Applying existing deep learning (DL) techniques directly to the raw IQ radar data has shown limited accuracy for various applications. To address this, this article proposes to learn the parameters of structured functions as filterbanks within complex-valued (CV) neural network architectures. The initial layer of the proposed architecture features CV parameterized learnable filters (PLFs) that directly work on the raw data and generate frequency-related features based on the structured function of the filter. This work presents four different PLFs: Sinc, Gaussian, Gammatone, and Ricker functions, which demonstrate different types of frequency-domain bandpass filtering to show their effectiveness in RF data classification directly from raw IQ radar data. Learning structured filters also enhances interpretability and understanding of the network. The proposed approach was tested on both experimental and synthetic datasets for sign and modulation recognition. The PLF-based models achieved an average of 47% improvement in classification accuracy compared with a 1-D convolutional neural network (CNN) on raw RF data and an average 7% improvement over CNNs with real-valued learnable filters for the experimental dataset. It also matched the accuracy of a 2-D CNN applied to micro-Doppler ( $\mu $ D) spectrograms while reducing computational latency by around 75%. These results demonstrate the potential of the proposed model for a range of RF sensing applications with enhanced accuracy and computational efficiency.