Analog spatiotemporal feature extraction for cognitive radio-frequency sensing with integrated photonics

Area of Science:

• Photonics
• Radio-Frequency (RF) Engineering
• Cognitive Sensing

Background:

• Analog Feature Extraction (AFE) offers low-latency and efficiency for cognitive sensing by reducing data sparsity.
• Broadband RF applications face challenges with AFE due to electronic circuitry limitations in bandwidth and programmability.

Purpose of the Study:

• To introduce a photonics-based scheme for analog domain spatiotemporal feature extraction from broadband RF signals.
• To develop a trainable photonic feature extractor using integrated photonic circuits inspired by convolutional neural networks.
• To propose a digital-analog-hybrid transfer learning method for efficient and cost-effective training.

Main Methods:

• Implementation of a convolutional neural network-inspired feature extractor on integrated photonic circuits.
• Processing of multi-antenna RF signals to extract temporal and spatial features in the analog domain.
• Utilizing a digital-analog-hybrid transfer learning approach for training the photonic system.

Main Results:

• Demonstrated a photonic analog feature extractor for radar target recognition with a 4-GHz instantaneous bandwidth.
• Achieved a 97.5% target recognition accuracy, maintaining high performance.
• Reduced the required analog-to-digital converter (ADC) sampling rate to 1/4 of the Nyquist rate.

Conclusions:

• The proposed photonics-based AFE scheme effectively processes broadband RF signals.
• This approach enables efficient cognitive RF sensing, applicable to autonomous driving, robotics, and smart factories.
• The trainable photonic extractor and hybrid learning method offer a promising path for advanced RF signal processing.