Research On Double Deep Learning Algorithm For Parkinson’s Disease Speech

Parkinson’s Disease(PD)is a degenerative nervous system disease,seriously harmful,and currently there is no effective treatment in international internationally,and can only diagnose timely treatment as soon as possible.Compared to routine detection methods such as biochemistry,imaging and measters,phonology has the advantages of lossless,low price,convenient,non-contact,etc.,can effectively detect the split barriers of PD,so research a highly efficient PD voice The identification algorithm is of great significance.However,there are still some problems in the current research of PD speech recognition methods.For example:(1)Although the existing PD speech deep learning methods have good non-linear feature extraction capabilities,the quality of their deep features is not high with a small sample size.(2)The existing PD speech learning methods are mostly limited to the original sample space,and the deep information of the structure between samples is not mined.These problems limit the improvement of PD speech recognition accuracy.In response to these problems,this thesis starts with two aspects of feature learning and sample learning,and carries out the following research work:(1)Aiming at the problem of mining deep structure information between samples,this paper proposes a PD voice deep sample learning algorithm based on iterative mean clustering,constructs a deep layered sample learning space,and realizes deep reconstruction of PD voice samples and mining The internal relationship between samples,through sample learning to improve the use efficiency of sample information,obtaining high-quality samples will help improve the performance of the classifier.(2)Aiming at the problem of PD voice deep feature extraction,this paper designs two PD voice deep feature learning algorithms-multi-layer self-sparse coding and deep residual network to realize the deep feature mining of PD voice features and improve the depth and original features.The complementarity of the features helps to improve the feature fusion performance;and the feature learning mechanism of L1 regularized feature selection is introduced to balance the PD voice feature representation ability and compactness.(3)Based on the bagging integrated learning model,the deep sample learning algorithm and deep feature learning are combined to construct a PD voice double deep learning integrated model to improve the PD voice small sample learning ability and the generalization ability of the algorithm.This research provides a new solution for improving the performance of PD speech recognition,and provides new research ideas for sample mining and feature extraction in PD speech recognition.It has a certain reference value and promotes the research progress of PD speech recognition theory and methods.