Whispers of the Void: An AI-Powered Radio Signal Classification System for Searching Extraterrestrial Civilizations

This article introduces the SETI-inspired machine learning pipeline 'Whispers of the Void', whose core goal is to use AI technology to automatically distinguish Earth-based Radio Frequency Interference (RFI) from potential extraterrestrial technical signals. It addresses the core challenge of signal filtering in SETI research and pushes forward a new chapter in the search for extraterrestrial intelligent life. This system combines radio astronomy and deep learning, and improves signal screening efficiency through a human-machine collaboration model.

Since ancient times, humans have questioned the loneliness of the universe. In the 20th century, radio astronomy provided tools for SETI. Radio telescopes can capture faint electromagnetic signals from the universe, but more than 90% of them are Earth-based interference (satellite communications, radar, broadcasting, etc.). How to filter extraterrestrial signals from massive data is the core problem of SETI.

The 'Whispers of the Void' system focuses on classifying two types of signals:

• Technical signals: Narrowband carriers, modulation patterns, Doppler shift, spatial positioning consistency;
• RFI: Continuous interference, pulsed interference, drifting signals, broadband interference, accounting for over 90% of observation time.

Data Preprocessing and Feature Engineering

Raw dynamic spectrograms undergo normalization, denoising, and segmentation, then features such as bandwidth, duration, and drift rate are extracted.

Deep Learning Model

A CNN or CNN+LSTM hybrid architecture is used: CNN learns time-frequency image features, LSTM captures time-series dynamics; preprocessed data is input, and signal classification probabilities are output.

Training Data

Real radio observation data is used with professional annotations (RFI/candidate signals, subcategories, credibility). Data augmentation (translation, offset, scaling) is applied to improve generalization ability.

Real-Time Data Stream Processing

1. Buffer slicing → 2. Candidate signal detection →3. Feature extraction →4. Classification inference →5. Post-processing filtering →6. Alarm recording.

Human-Machine Collaboration

AI pre-screens candidate signals, and astronomers conduct manual review; signals that pass the initial screening trigger joint observations by multiple telescopes to verify repeatability and spatial positioning.

• Class imbalance: The ratio of RFI to candidate signals reaches millions to one, solved using resampling, cost-sensitive learning, anomaly detection, and ensemble methods;
• Open-set recognition: Need to identify unknown signal types;
• Interpretability: Use attention mechanism visualization and SHAP value analysis to reveal the basis of model decisions.

Scientific Significance

• Understand the cosmic distribution of technological civilizations;
• Promote data-driven scientific discovery models;
• Extend the boundaries of human curiosity.

Future Outlook

• Integrate multi-wavelength multi-messenger analysis of optics, infrared, and gravitational waves;
• Real-time verification via global telescope networks;
• Expand search scale by combining with citizen science platforms.