The egg embryo culture method is commonly used for the preparation of avian influenza vaccines. The rapid and accurate classification of hatching eggs into active and early removal of dead embryos from live embryos can effectively avoid bacterial or mycobacterial contamination due to embryo death and it is of great importance for the improvement of hatching efficiency. Currently, the heartbeat signal of chicken embryos is mainly used as the basis for distinguishing dead embryos from live embryos. However, after 96 h of avian influenza virus injection, the heartbeat signal of live egg embryos is between that of ordinary live embryos and dead embryos, which is easily confused with dead embryos. This type of data is called chicken embryo heartbeat confusion signal, and is added to the data set as a separate category. The original dual classification of dead embryos and live embryos was changed to a triple classification of dead embryos, ordinary live embryos and 96 hour live embryos. An absolute average value normalization preprocessing method was proposed based on confusing heartbeat signals of hatching eggs, to enhance the original data features and improve the classifiability of the data. A shallow dual branch network structure residual fully temporal convolutional with transformer network (RFTNet) with residual structure was proposed based on temporal convolutional network (TCN) and transformer for global features and detail features. The experimental results showed that the three-classification absolute average value normalization preprocessing method and RFTNet two-branch network proposed demonstrated good performance in the classification task of hatching eggs confusion dataset with a detection accuracy of 99.75%. In addition, the three evaluation indexes of detection accuracy, recall rate and F1 score reached 99.75%, 99.74% and 99.7%, respectively, further verifying the effectiveness of the method.