Most large telescopes adopt the design scheme of segmented mirror. In order to obtain high-quality imaging effect, it is necessary to control the piston and tip-tilt errors of segmented telescope system. Compared with traditional detection methods, the error detection method based on neural networks has some advantages, but it is limited to detecting only a single type of error. This paper proposes a method for synchronous detection of piston and tip-tilt errors based on a multi-convolutional neural network. By setting a mask with a sparse sub-pupils configuration at the exit pupil, the sub-waves reflected by the segmented mirrors generate interference-diffraction phenomena, thereby constructing a dataset containing rich piston and tip-tilt errors information. The design includes coarse measurement and fine measurement networks to meet the requirements of large-range and high-precision synchronous detection. Results demonstrate that the method achieves nanometer-level detection of piston errors within the coherent length of the input light source and submilliarcsecond detection of tip-tilt errors within a range of 10 μrad. The method exhibits robust resistance to 40 dB CCD noise, a tolerance of 0. 05 λ RMS (λ0 = 600 nm) for surface shape errors, and portability to six-mirror systems. Additionally, the method has simple optical path, convenient operation and practical significance.