Dysphonia Analysis Of Parkinson’s Disease Based On Fractional Spectrogram

Dysphonia is one of the early symptoms of Parkinson’s disease,and the analysis of Parkinson’s dysphonia has become a hot research in the current field,especially the analysis based on the time-frequency description of the spectrogram has received much attention.However,the typical spectrogram has the problems of single perspective and limited information extraction.To solve these problems,this paper proposes a fractional spectrogram method to describe Parkinson’s dysphonia through multi-angle energy information.Based on this,the Fr Swin model based on fractional spectrogram and the method based on fractional attribute topology are proposed to analyze Parkinson’s dysphonia respectively.The details are as follows:Firstly,in order to compensate the limitation of the typical spectrogram to describe the dysphonia signal from a single perspective,the method of fractional spectrogram is proposed.The method increases the rotation factor of angle to transform the dysphonia signal into fractional spectrogram and enhances the ability to obtain richer energy information from different angles.Further,the methodological framework of deep learning and acoustic feature description based on fractional spectrograms is proposed.Secondly,the Fr Swin based on fractional spectrograms is proposed to enhance the interpretability of deep learning.The method takes the fractional spectrograms at different orders as inputs to improve the interpretative meaning of the resulting feature.At the same time,the shared parameters of the Fr Swin pre-trained on Image Net are migrated to the fractional spectrograma using transfer learning,and fine-tuned to obtain the final weight parameters,thus solving the drawback of small data size and enabling the model to learn more knowledge that is beneficial to classification.Finally,the classification labels are calculated based on the feature obtained from the Fr Swin and the weight parameters training with the transfer learning.In the experiments,97.80% is achieved on Database-1 when the order is 0.5,and 98.75% is achieved on Database-2 when the order is 0.9.The experimental results show that Fr Swin not only improves the classification accuracy and interpretability,but also has stable performance.Finally,in order to strengthen the correspondence of energy information in the fractional spectrogram,the acoustic feature description method based on fractional attribute topology is proposed.The method counts the energy information in the fractional spectrogram and obtains the confidence interval of direction values by kernel density estimation.A mapping is established based on the affiliation between the confidence intervals and the directional attributes,which is then transformed into a formal context.Based on the formal context,the fractional attribute topology is established to describe the association between the direction attributes in the sub-region.By analyzing the discrete degree of the fractional attribute topology,the connected component features are extracted for the analysis of Parkinson’s dysphonia.In the experiments,99.57% and 96.38% are achieved on Database-1 and Database-2,when the order is 0.7,respectively.The results show that the proposed features can effectively describe the Parkinson’s dysphonia.