CNN-based compensation of faulty planar phased-array radiation patterns

Phased-array antennas experience severe radiation pattern degradation when elements fail. Traditional compensation methods, such as genetic algorithms and particle swarm optimization are computationally expensive, requiring 120–300 s of iterative optimization per failure event. This paper proposes a novel convolutional neural network (CNN) framework for rapid, pattern-level compensation of faulty planar phased arrays. The CNN processes a 64 × 64 orthographic 2D radiation pattern image (azimuth–elevation) from a 4 × 4 array with one failed element and one fixed reference element. It directly predicts recovery phase shifts for the 14 remaining elements to restore the original beam pattern. Trained on 8,000 CST-generated patterns with randomized phases, the CNN achieves sub-degree accuracy (Mean Absolute Error (MAE) = 6.1 × 10− 3 (rad) and high fidelity (R² = 0.97). Quantitative and visual results demonstrate effective recovery of main-beam characteristics and sidelobe suppression, closely matching intact-array performance. Crucially, inference takes about 200 ms on a standard GPU—orders of magnitude faster than iterative optimizers, enabling real-time self-healing. In addition, the model reduces average pattern error by 32.1% (± 3.8%) across 120 test cases, with performance ranging from 28.28% to 35.91% improvement in RMSE, demonstrating consistent fault compensation capability, indicating substantial recovery of the degraded radiation pattern. This proof-of-concept demonstrates that CNNs can effectively learn complex pattern-to-phase inverse mappings for faulty arrays, offering a computationally efficient alternative to traditional optimization methods for scenarios where radiation pattern measurements are available. While the current study is limited to single-element failures in simulation, it establishes feasibility for future practical implementations incorporating measurement systems and multi-fault scenarios.