Research On Identity Recognition Algorithm Based On Pulse Wave

As an important manifestation of human physiological signals,pulse wave can reflect the individual differences between people and has many advantages.It has gained widespread attention as a novel identification method.Based on the MIMIC dataset and self-built multi-channel pulse dataset,this thesis studies the algorithm application of deep learning on pulse wave identification.The main research work was carried out as follows:The thesis studied the preprocessing methods for the filtering of data noise and the extraction of effective fragments in pulse wave signals.We built a wrist pulse dataset by collecting synchronous pulse data from the cun-guan-chi positions of healthy individuals using the HK-2000 C traditional Chinese pulse diagnosis instrument.Furthermore,we constructed the single-channel and multi-channel pulse sample sets.This article first conducts experimental comparisons on the benchmark pulse dataset using multiple deep one-dimensional benchmark network models.Regarding the higher recognition performance of the Dense Net model on the MIMIC dataset,its basic structure is analyzed and improved,and a one-dimensional pulse wave identity recognition network based on feature reuse was proposed.Through experimentation,An accuracy rate of 97.35% and an EER value of 1.04% were achieved on the benchmark dataset,and the recognition effect of the model under different signal lengths and network parameters was compared and analyzed.The optimal accuracy of98.55% and EER value of 0.69% are achieved on the "guan" channel of the self-built dataset,which validates the superior performance of the network model.To address the feature fusion problem of multi-channel pulse signal,a graph convolutional network was applied to establish correlations among multiple channels’ data.Taking each channel and its differential waveform as the research objects,highdimensional features were respectively extracted as graph nodes using the aforementioned pre-trained network.An adaptive graph generation layer was proposed to establish the topology connection relationship between nodes,and the Chebyshev graph convolutional neural network was used for identity recognition.A 99.75%accuracy rate and 0.11% EER value were achieved on the self-built multi-channel dataset.Through ablation experiments on single-channel and graph convolution modules,the effectiveness of multi-channel feature fusion was verified.Experimental analysis was conducted on the noise immunity of the network under long and short cycles.