With the singleness of the training sample, the phenomenon of overfitting occurs in the deep neural network when used for specific emitter identification (SEI), which in turn affects the accuracy. In this paper, a deep convolutional neural network (CNN) structure based on PID algorithm is proposed to alleviate the problem. The structure builds a feedback loop between the output layer and the input layer of the traditional CNN, transforms the error rate of output layer into the probability of dividing the training set data by using PID algorithm, and inhibits the overfitting by optimizing the composition of training set data. The average recognition rate of the network reaches 92.59%when the method is applied to the recognition of ultrashort wave radio. The variance of the recognition rate is about 1/3 of that of the traditional algorithm, and the training time is reduced by about 35 min, obviously the performance of this method is better than that of the traditional neural network. Experimental results show that the algorithm can enhance the robustness of the deep network and effectively suppress the overfitting phenomenon.