针对存在严重符号间串扰及轻度非线性畸变的数字信道，提出了两种判决反馈递归神经网络均衡器结构，将传统的线性信道判决反馈结构巧妙地融入递归神经网络中，在自适应训练时用期待信号代替判决反馈信号，并对学习步长自适应调节进行了研究．实验结果表明：在自适应参数数目相同的情况下，该均衡器较传统的递归神经网络均衡器具有更好的特性． For the digital channels with severe inter-symbol crosstalk and mild nonlinear distortion, two decision feedback recurrent neural network equalizer structures are proposed. The traditional linear channel decision feedback structure is intelligently integrated into the recurrent neural network. In the adaptive training The expected signal is used to replace the decision feedback signal, and the adaptive adjustment of learning step is studied. The experimental results show that the equalizer has better performance than the traditional recurrent neural network equalizer with the same number of adaptive parameters.