Research And Implementation Of Heart Sound Signal Oriented Anomaly Detection

The number of patients with cardiovascular diseases worldwide is increasing rapidly,while medical resources are increasingly scarce.The detection of abnormal heart sounds,as the most direct means to detect cardiovascular diseases,is drawing the attention of researchers around the world.A growing number of detection methods have been applied to the field of heart sound anomaly detection,but there is still room to promote the precision of the detection model.Simultaneously,in practical application scenarios,in order to reduce the computational overhead and easy deployment,it is necessary to compress and optimize the model.Consequently,so as to settling these two problems,this thesis designs a heart sound anomaly detection model based on partition attention and proposes a three-stage model optimization strategy.The following are the research contents of this thesis:First of all,the heart sound anomaly detection model based on the partition attention(PANet)is constructed.Unlike the traditional heart sound anomaly detection method,which uses the time dimension feature of heart sound signal,this thesis uses the image generated by the transformation of the high dimension feature of heart sound signal as the input of the model,and then uses the convolution neural network as the feature extractor.The experimental results show that converting the heart sound signal into bispectrum can effectively represent the coupling relationship between the frequency components of the heart sound signal.Without data enhancement,the accuracy rate is 92.08% on the Physio Net 2016 dataset.In order to enable the network to obtain more feature information from the input,the partition attention module(PA)is designed according to the visual characteristics of the bispectrum,the Fusion Ghost module is designed based on the ghost module.The results indicate that this model can attain 97.89% accuracy,96.96%sensitivity and 98.85% specificity on Physio Net 2016 dataset.Secondly,a three-stage model optimization strategy is proposed,which is followed by structural optimization,knowledge distillation and model pruning to solve the problem that large-scale models cannot be deployed to equipment with insufficient resources in all aspects in the actual scene.The experiment indicates that,on the basis of the small difference between the accuracy of the optimized model and the original one,the parameter quantity and calculation quantity are decreased by 91.1% and 41.22%respectively.In the end,a great quantity of contrast experiments and ablation experiments were carried out in the Physio Net 2016 dataset to test the advantages of the designed model,and the heart sound anomaly detection system was devised and completed with the proposed model as the core component.The results indicate that compared with other models,the model PANet proposed in this thesis has certain advantages in model accuracy,specificity and sensitivity.The compressed and optimized model can be easily deployed and applied,which proves that the heart sound anomaly detection model and model optimization strategy proposed in this thesis are effective exploration in the field of heart sound anomaly detection.