Sound Events-based Modeling Of The Behavior Of The Elderly Living Alone

The prevalence of senile dementia is increasing year by year,and elderly people who live alone are more likely to develop the disease due to lack of communication with the outside world.Therefore,the study on the behavior and abnormal detection of the elderly living alone is helpful to prevent Alzheimer’s disease.At present,the most of health monitoring for the elderly is based on video monitoring and location information,which infringes on personal privacy.The event detection technology based on sound signal is relatively mature.Therefore,this dissertation proposes to study the behavior of the elderly living alone based on sound events.The main research work in this dissertation is as follows:（1）Research on sound signal acquisition,preprocessing and feature extraction.In the home environment,11 typical simple sound events and 1 complex sound event data are collected.Adopting digital filter,resample,frame windowing and endpoint detection to preprocess the sound signal.Five features such as STFT,LPCC,Log-mel,MFCC and GFCC are extracted.（2）Research on simple sound event classification method.For five different features,this dissertation adopts the event classification methods based on SVM,KNN,LSTM and1 D CNN,and carries out the ergodic combination experiment,so as to select the relatively optimal combination of features and classification methods.The experimental results show that the optimal combination based on MFCC feature and LSTM event classification has relative higher recognition rate and lower time cost is based on MFCC feature and LSTM event classification.（3）Research on state-sequence modeling of complex sound events.Taking the washing intelligent washing machine as the research object,the whole process of the washing machine is divided into four states: water intake,soaking,washing and dehydration.The SVM classification method based on decision tree is used to classify the state.The state-sequence is constructed according to the time sequence and duration of the state.The experimental results show that when the feature is Log-mel and the decision tree is soaking/washing/water intake/dehydration,the recognition rate reaches the highest.After using the adjacent minimum feature distance method to correct the error state in the reconstructed state-sequence,the number of segments is corrected correctly,and there is only error in the positioning of the beginning and end of the segments.（4）Research on modeling method of behavior of elderly living alone based on sound events.In this dissertation,a modeling method of compatible serial and parallel events based on event space-time matrix is proposed for the first time.Based on the classification of simple sound events and the state-sequence of complex sound events,the behavior-sequence within the period is obtained by serial parallel space-time modeling method.According to the DTW minimum distance method,the center of the event space-time matrix,which is the reference of the behavior routine,is calculated.The abnormal behavior is detected according to the distance threshold.Taking a life schedule as an example,the behavior of the elderly living alone is modeled.The feasibility of the model is verified.The research work done in this dissertation is oriented to the prevention of Alzheimer’s disease,which is worthy of long-term and in-depth research.The work done in this dissertation has a certain reference value for peers.Figure [67] Table [20] Reference [80]...