Respiratory diseases of pigs are easily contagious, which affects pig breeding efficiency. Cough is one of the obvious symptoms of respiratory diseases. An algorithm based on deep neural network was proposed to accurately identify pig coughs. Log_filter bank (logFBank) and Mel frequency cepstral coefficents (MFCC) were extracted respectively after spectral subtraction denoising and double threshold endpoint detection of the sound signal. Then the two kinds of extracted features and their first and second order differences were used as inputs to the convolutional neural networks (CNNs) and the deep feed forward sequence memory neural networks (DFSMN) for multi-classification training. The effects of the different features and different iteration times on the effectiveness of the model were compared. Except the accuracy of cough recognition, the recognition effects of other pig sounds, such as sneezing, which was easily confused with cough were also analyzed. The experimental resulst showed that when the number of training rounds reached 200, the CNNs model with MFCC as feature had a good effect. The recognition precision of cough on test set was 97%, the cough recognition recall rate was 96%, the F1-score was 98%, and accuracy reached 96.71%. It was showed that the model was effective and feasible, and can provide technical support for pig cough recognition in pig welfare breeding.