Machine Learning in Sensors: Feature Learning from Complex Spectra in mmWave Radar Signals

Feature learning from complex spectra in millimeter-wave (mmWave) Radar is an increasingly significant and challenging field with substantial implications for various applications including autonomous vehicles, surveillance systems, industrial systems, and human-computer interactions. This paper articulates a robust argument for the application of Complex-Valued Restricted Boltzmann Machines (CV-RBMs) as an effective technique for extracting meaningful features from complex, high-dimensional spectral data. Traditional methods often struggle to fully capture the inherent structural intricacies within such complex-valued data, which frequently exhibits local correlations and hierarchical relationships. CV-RBMs, by leveraging the inherent complex nature of spectral information, excel in uncovering robust and discriminative features crucial for a multitude of downstream applications. The proposed CV-RBM model demonstrably excels in feature learning from complex-valued data, achieving good accuracy in terms of reconstruction error. Measurement results of our extensive experiments have shown a significant advancement in data representation, allowing for the encoding of complex spectral data into a highly representative format. The reduction in dimensionality not only streamlines computational processes but also facilitates the extraction of highly informative features, ultimately leading to enhanced performance in various tasks such as classification, clustering, and regression.