Research On Multi-scale Structural Neural Network For ECG Identification

In recent years,with the rapid development of the Internet,the problem of information security has become more and more serious,and the requirement for identity recognition systems has been increasing.As a complementary scheme to existing biometrics,ECG signals can only be detected in vivo,validation does not require explicit user actions,and contains cardiac health information.It has a wide range of applications and has attracted much attention.The key point in ECG signal recognition techniques is feature extraction and recognition.Traditional methods for ECG recognition,such as geometric feature-based methods and subspace-based methods,rely on prior knowledge,and some transform domain features have large computational complexity,which makes it difficult to improve classification accuracy.Deep learning techniques driven by big data,especially networks with multi-scale structure,have achieved high performance in ECG signal recognition.The closed-set ECG signal identification application tends to mobile devices,where the individuals to be identified are a fixed set.However,feature extraction backbone networks with complex topologies require significant storage and computational resources.Mobile device resources do not have the conditions to deploy a heavyweight ECG recognition network,and it is necessary to make the network lightweight.However,the closed-set identification algorithm model needs to be retrained when the set of identified individuals changes,which cannot be applied in scenarios with large human mobility.Typically,similarity computation is used to implement open-set identification of the individuals to be identified.In this way,when the extracted features satisfy the characteristics of small intra-class gap and large interclass gap,it is more favorable for accurate recognition.However,the features extracted by traditional neural networks are relatively scattered in the mapping space,and they are discriminative for closed set recognition problems but lack discriminative power for open set recognition problems.In view of the above problems,facing different application scenarios,this paper studies the ECG recognition method based on multi-scale convolutional neural network.The main work is as follows:1.In order to solve the problem that the closed-set ECG identification network is limited to hardware conditions when deployed on the mobile terminal,this paper proposes a closed-set ECG signal identification method based on multi-scale structure neural network.In the training phase,multi-scale features are extracted from the ECG signal,and the multi-scale analysis of the signal fully mines the information under the multi-scale of the signal to improve the recognition performance.In the inference phase,the multi-scale structure is transformed into a single-branch structure with a smaller storage footprint through parameter fusion,which further reduces resource consumption while maintaining model performance.Experimental results show that the proposed closed-set ECG recognition neural network achieves higher recognition accuracy on public datasets and reduces the storage space requirement.2.In order to solve the problem that the extracted features of traditional ECG identification neural network lack discrimination in the open set identification scene,this paper proposes an open-set ECG identification method based on multi-scale structure neural network.Firstly,the single-layer convolution in the Res Net bottleneck block was replaced by a set of convolution groups connected in a hierarchical manner similar to the residual in the feature extraction network structure,so as to realize the multi-scale analysis of the signal and mine the multi-scale features of the signal at a fine-grained level;At the same time,the Additive angular margin loss(Arcface)function is used to normalize the feature vectors and the classification weights,so that the classification depends on the Angle between the feature vectors and the weights on the hypersphere.Then,the method of adding a margin to the Angle interval is used to guide the model to train in the direction of more dense samples from the same class and more scattered samples from different classes to improve the accuracy of identity recognition.Experiments show that the designed method has high recognition accuracy on public datasets.