Noise-Robust Piston Error Detection in Segmented Mirrors via Wavelet-Driven Deep Multichannel Features

To address the challenges of piston error detection in segmented mirror systems under noisy environments, this article proposes a wavelet-driven deep multichannel feature extraction method (WTCNN). To the best of our knowledge, this is the first application of a multichannel convolutional neural network (CNN) method combined with wavelet transform in piston error detection. The method first employs wavelet transform to decompose single-channel diffraction images into four components. These components are then processed by a modified CNN (ResNet-18) to extract multichannel features and predict the corresponding piston error value. Simulation results demonstrate that within a single wavelength range, the proposed method significantly outperforms traditional CNN-based methods in terms of detection accuracy and noise robustness. Specifically, under a noise intensity of 20 dB, the proposed WTCNN achieves a root mean square error (RMSE) of $0.002535\lambda $ , representing approximately a 7.2-fold improvement in detection accuracy compared to traditional CNN methods (RMSE $= 0.020703\lambda $ ). Furthermore, experiments under wavefront aberrations and varying noise levels (15–40 dB) reveal the strong generalization capability of the proposed method: it maintains high detection accuracy even when the noise level deviates from the training environment, whereas traditional methods suffer a significant decline in performance. This study not only achieves substantial improvements in the accuracy and noise resistance of piston error detection but also provides a theoretical foundation and practical insights for addressing wavefront aberrations in experimental environments.