In order to achieve low complexity balancing of nonlinear damage at the receiver of short-range fiber optic data communication systems, we propose an equalization structure named Decision Feedback Neural Network which introduce the Decision Feedback Structure into the Fully Connected Neural Network. The nonlinear distortion is introduced by using a photodetector with a linear working area that does not match the experimental system. The experimental system is built based on a 56 Gbit/s PAM4 with a C-band direct-modulated laser, and we compare the equalization performance of decision feedback neural network with other equalization schemes. Experimental results show that compared with the fully connected neural network, the improved scheme achieves a sensitivity improvement of 2 dB at 20 km transmission, and the equalization performance is close to the convolutional neural network with lower complexity. This paper has great significance for the rate and capacity upgrade of short-distance optical fiber communication system, and can be used as a reference for further scientific research and industrial application.