Synthetic aperture is the mainstream structure of current astronomical telescopes. However, after the synthetic aperture telescope is deployed in orbit, there will remain tilt and piston error between adjacent segments, which will sharply deteriorate the imaging quality of the optical system. The traditional piston error detection method based on dispersed fringe sensor has the question that it is difficult to detect the piston error within one wavelength, and the detection accuracy is restricted by the detection range. The method in this paper constructs multiple monochromatic light channels by opening windows in different areas on the dispersed fringe pattern, calculating and obtaining the feature value in each window to form a feature vector. Then, the convolutional neural network is introduced to distinguish the feature vector to detect piston error. Among them, the training set construction method adopted in this paper only needs raw data in one wavelength to construct a training set covering the entire detection range. Through simulation, the method proposed in this paper achieves the detection range of [-208λ, 208λ] (λ=720nm), and regardless of the presence of noise, the root mean square value of the detection error does not exceed 17.7nm (0.027λmin, λmin=660nm). © 2021 SPIE