TABASCAL II: Removing Multi-Satellite Interference from Point-Source Radio Astronomy Observations

ABSTRACT

"In the first tabascal paper we showed how to calibrate in the presence of Radio Frequency Interference (RFI) sources by simultaneously isolating the trajectories and signals of the RFI sources. Here we show that we can accurately remove RFI from simulated MeerKAT radio interferometry target data, for a single frequency channel, corrupted by up to 9 simultaneous satellites with average RFI amplitudes varying from weak to very strong (1 – 10 cubed Jy). Additionally, tabascal also manages to leverage the RFI SNR to phase calibrate the astronomical signal. tabascal effectively performs a suitably phased up fringe filter for each RFI source which allows essentially perfect removal of RFI across all RFI strengths. As a result, tabascal reaches image noises equivalent to the uncorrupted, no-RFI, case. For larger RFI amplitudes, the resulting image noise is 10-100x smaller than those from traditional RFI flagging methods such as AOFlagger. Consequently, point-source science with tabascal almost matches the no-RFI case with near perfect completeness for all RFI amplitudes. In contrast the completeness of AOFlagger and idealised 3σ flagging drops below 40% for strong RFI amplitudes where recovered flux errors are 10x-100x worse than those from tabascal. Finally we highlight that tabascal works for both static and varying astronomical sources."

The following improvements detail how the methodological architecture of tabascal II can be generalized to enhance various AI systems, moving beyond its original application in radio astronomy to solve complex signal separation and reconstruction problems across multiple domains.

System Improvement: Generalizing the core Bayesian Forward Model (Section 4.2) into a Dynamic Interference Mitigation Module (DIMM) for hyperspectral and radar data processing. This module replaces traditional fixed-threshold flagging with a probabilistic framework that simultaneously models the signal, environmental noise, and interference sources.

Specific Implementation: The system utilizes a Multi-Component Gaussian Process (GP) structure, where the input signal is decomposed into:

1. Target Signal Component: Modeled using an SE covariance function tuned to expected spatio-temporal variability (analogous to V RFI in tabascal).

2. Interference/Noise Component: Modeled via a trajectory-based GP (analogous to the RFI satellite movement), accounting for time-varying environmental biases (e.g., cloud motion, atmospheric distortion).

3. System Response Component: Modeled using a time-correlated gain/calibration function (analogous to G p in tabascal) derived from pre-calibration data.

What the Improved AI System Can Do: It can reconstruct high-fidelity images by seeing through dynamic interference, achieving noise profiles statistically consistent with the theoretical ideal sensor performance, even when the interference is highly correlated with the environmental conditions.

System Improvement: Implementing a Stochastic Signal Separation Engine (SSEE) for high-bandwidth, multi-path communication systems where RFI or signal decoherence is frequent and complex.

Specific Implementation: The SSEE utilizes the fringe rate filtering concept (Section 4.1 & 4.2) within its Fourier domain processing to decouple desired data streams from time-varying interference. Instead of relying on simple power thresholds, it uses a Bayesian estimator that determines the likelihood of a complex signal by jointly fitting:

1. The intended data stream (modeled by a stationary power spectrum P 0).

2. Known interferers (modeled via TLE/trajectory-based GPs).

3. System latency and phase drift.

What the Improved AI System Can Do: It achieves reliable decoding and artifact removal in high-noise environments, maintaining bit error rates comparable to noise-free conditions, even when traditional signal processing methods fail due to severe time-domain correlation issues.

System Improvement: Developing a Probabilistic Artifact Correction Framework (PACF) for time-series medical scans where patient movement or detector drift acts as complex, time-varying interference.

Specific Implementation: The PACF applies the joint optimization principle of tabascal to solve for three unknowns simultaneously:

1. The true underlying physiological signal.

2. The continuous, non-linear motion parameters (motion model).

3. The detector/system gain variations.

This is executed using a scalable MAP estimation approach (Section 4.3) applied to the measured data, allowing for simultaneous reconstruction of the cleanest possible image while quantifying the uncertainty introduced by movement.

What the Improved AI System Can Do: It provides high-resolution medical images that are statistically superior to standard motion-compensated scans, enabling accurate diagnoses even in patients with significant involuntary movement, without introducing artificial smoothing or signal loss.

System Improvement: Creating a Time-Varying Interference Predictor (TVIP) for large industrial sensor arrays (e.g., vibration monitoring, factory floor acoustics).

Specific Implementation: The TVIP uses the GP interpolation mechanism (Section 4.2.2) to model the baseline relationship between different sensor inputs and their expected noise/interference profiles. It continuously predicts when a specific interference event will occur based on current environmental data (e.g, nearby vehicle traffic) and then proactively models the resulting signal distortion before it corrupt the primary data stream.

What the Improved AI System Can Do: It allows for predictive maintenance by identifying and compensating for RFI/noise sources before they saturate sensors, ensuring operational data streams remain clean and reliable.

In the first TABASCAL paper we showed how to calibrate in the presence of Radio Frequency Interference (RFI) sources by simultaneously isolating the trajectories and signals of the RFI sources. Here we show that we can accurately remove RFI from simulated MeerKAT radio interferometry target data, for a single frequency channel, corrupted by up to 9 simultaneous satellites with average RFI amplitudes varying from weak to very strong (1 - 1000 Jy). Additionally, TABASCAL also manages to leverage the RFI signal-to-noise to phase calibrate the recovered astronomical signal. TABASCAL effectively performs a suitably phased up fringe filter for each RFI source which allows essentially perfect removal of RFI across all strengths. As a result, TABASCAL reaches image noises equivalent to the uncorrupted, no-RFI, case. Consequently, point-source science with TABASCAL almost matches the no-RFI case with near perfect completeness for all RFI amplitudes. In contrast the completeness of AOFlagger and idealised 3 sigma flagging drops below 40% for strong RFI amplitudes where recovered flux errors are about 10x-100x worse than those from TABASCAL. Finally we highlight that TABASCAL works for both static and varying astronomical sources.