To address the limitation of error performance in Massive MIMO detection algorithms under high-order modulation scenarios, a novel detection network is proposed. This network is constructed based on a projected gradient descent framework that approximates the maximum likelihood solution through an iterative structure, which is then implemented using a neural network. In each network module, parameters are first learned through a neural network, followed by a nonlinear transformation using a designed normalized multi-segement activation function to enhance the network′s mapping capability under high-order modulation. Finally, a denoiser is employed to eliminate estimation errors and channel noise. Furthermore, to tackle the issue of accuracy degradation with increased network depth, residual connections are introduced between network modules. Simulation results show that when the number of transmitting and receiving antennas of the system is 64×32 and the signal-to-noise is 16 dB, the bit error rate of the proposed detection network is close to 10-4, and the bit error rate is reduced by an order of magnitude compared with other detection algorithms, showing the bit error performance close to the optimal detection algorithm, and has good robustness.