Research On Classifying Heartbeat Signals Of Hatching Eggs Based On MifaNet

Avian influenza viruses are easy to mutate,spread rapidly.Vaccination can effectively prevent avian influenza.Vaccines are generally prepared via the hatching eggs culture method.That is,the inactivated strain is injected into healthy hatching eggs for propagation,then collecting,inactivation and purification.The preparation process must be carried out in a sterile environment,it is crucial for producing pure vaccines to remove the dead embryos in time.Hatching eggs activity identification is primarily done manually,which is inefficient and requires a large labor force.Therefore,it is of great significance for the vaccine manufacturing industry to design an automated classification method.The presence of periodic heartbeats is one of the most distinctive characteristics of living embryos.This thesis takes the heartbeat signals of hatching eggs from 9 to13 days-old as the research object and proposes a classification network Mifa Net based on multi-scale information fusion.For the local information and global information in the heartbeat signals,LANet and GANet are established respectively in Mifa Net.For the LANet,the dilated convolution is used to build the network structure,which can extract the locally significant features of signals more comprehensively using increasing the model receptive field.At the same time,the channel attention mechanism is embedded in different stages of network to suppress low-level information and noise adaptively,which enhanced the feature expression ability of the LANet.For the GANet,self-attention mechanism is used to learn bidirectional dependencies in signals to capture periodic heartbeat features.In addition,this thesis proposes a multi-head fusion attention module MHFA,which calibrates the local detail features according to the importance degree obtained from calculating the attention score of the corresponding position between local information and global information.The MHFA module screens and fuses multi-scale information,and outputs the detection results of egg embryonic activity through the classifier.A series of experiments conducted on our data set showed that Mifa Net has a more stable accuracy,which was up to 99.88%.Meanwhile,the precision,recall and F1 score reached 0.9952,0.9973 and 0.9963 respectively,which further verified the effectiveness of Mifa Net.