Research On Acoustic Feature Extraction And Classification Model For Early Screening Of Children With Autism

The autism in Young children is a pervasive developmental disorder,They are evident not only in verbal deficiencies,poor social skills but also in stereotyped,repetitive behaviors.There are three aspects to describe in detail: the one is,very little active communication,and not answering the questions;the second is,they are not interested in the surrounding people;The third is simple way of behavior,it is difficult to make changes quickly after being stimulated.Before the age of 3,autism is a high incidence period in children,most of which are boys.The causes mainly focus on family inheritance,dysplasia during pregnancy,and underdeveloped brain functional areas.According to relevant statistics,the number of children with autism is increasing at a rate of 200 thousand per year globally,and in China,children with autism have been ranked as the first mentally disabled diseases.It has gradually become a global research topic.At present,the common auxiliary diagnostic methods include observation of brain with CT scan,EEG,EOG,Applied Behavior Analysis(ABA),and multi-scale comprehensive assessment,etc.The most diagnosis method is still based on long-term clinical observation and a variety of scales.Due to the lack of effective diagnosis methods,which leads to a long diagnosis cycle,low accuracy of diagnosis,and easy to miss the best intervention period.Sound is one of the most common carriers of information transmission in human life.Due to the differences in the structure of each one's vocal organs,the sound information emitted by different people will be slightly different.Therefore,through audio analysis of the collected sounds,you can find some specific information only belongs to himself,which is convenient for studying and summarizing the rules.In order to find an objective and true diagnosis method,this thesis starts from the acoustic perspective of autistic children,collects as many audio data as possible,By using machine learning and audio signal analysis to extract the most representative acoustic feature and screen the most suitable classification model.To build an audio analysis and recognition framework that may provide an acoustic detection method for the early clinical diagnosis of autism children.The main contents of the work are as follows:(1)Collection and production of sound data for autistic children.Due to the impact of complex noise environment and the lack of understanding of autistic children,the design of the collection paradigm is unreasonable,In addition,children do not cooperate results in the poor quality of data collected in the early stage,which seriously hinders the development of the project.Later,we cooperated with the doctors in the children's Department of Anhui Medical University to contact and pacify the families of patients.We used the improved recording paradigm to collect sound data.After simple guidance and training,the medical staff can collect data independently,solving the problem that we can't often collect new data because of the long distance.Through the joint efforts of both sides,It laid a foundation for establishing a sound database for autistic children in the future.(2)Research on audio signal front-end preprocessing algorithm.Through the analysis of the audio signal and Gaussian noise of autistic children,the audio denoising and endpoint detection algorithms suitable for the database in this thesis are determined.In terms of audio denoising,the experiments show that the multi window spectral subtraction algorithm can not only suppress the environmental noise,but also eliminate the "music noise" produced in the process of de-noising,and shows good robustness under different signal-to-noise ratios.In terms of endpoint detection,the five endpoint detection algorithms are compared and analyzed,and by comparing the error with the actual calibration position,it is finally concluded that the uniform subband spectral entropy based on multi-window spectral subtraction has better stability and accuracy.(3)Research on support vector machine model based on optimization algorithm.First introduce the concept of confusion matrix,and then use the four performance evaluation indicators of F1-Score,AUC,Accuracy,Time to evaluate the SVM model based on three optimization algorithms of Grid,GA and PSO.The experimental results show that the Grid has the shortest search time,the PSO has the highest recognition accuracy,and the GA is the worst of the three.Finally,the four acoustic characteristics of PLP,MFCC,LPCC and DWTMFCC were compared with each other in terms of accuracy and robustness by using the Grid and PSO.Finally,it was concluded that under the ideal environment,both PLP and MFCC had high accuracy,while under the low SNR,PLP had the best stable performance,followed by DWTMFCC.(4)Research on recognition model based on convolutional neural network.Due to the one-dimensional nature of the audio signal,this thesis transforms the 2D-convolution model to establish a 1D-convolution model that is more suitable for audio detection.From the experimental results,it can be seen that the performance of the 1D-convolution model are better than the traditional 2D-convolution model both in accuracy,convergence speed and noise resistance(5)SVM-based audio detection system for autistic children.On the basis of summarizing and researching audio signal front-end preprocessing,acoustic feature extraction and model optimization algorithms,a support vector machine based on Matlab2012 a platform is implemented.This system for autistic children,mainly integrating audio acquisition and recording,front-end preprocessing,model training,feature extraction and classification.In addition,The actual test results show that the system can achieve a good recognition effect.