Research On Snoring Detection And OSAHS Recognition Based On Spike Neural Network

Snoring is a common phenomenon affecting human sleep quality.It also an early symptom of many high-incidence sleep diseases.Obstructive Sleep Apnea-Hypopnea Syndrome(OSAHS)is one of sleep diseases.Due to the prevalence of sleep disorders,it is an important research fields to realize the local detection of diseases at home.Many studies have completed snoring recognition by neural networks.However,traditional neural networks have a large number of parameters to be calculated,and applying them to embedded devices to complete local detection is not easy.Based on this,this thesis uses the spiking neural network for snoring detection and OSAHS recognition.It converts real number calculations into pulse calculations,which reduces the consumption of computing resources and achieves good results in classification tasks.It provides a scheme for large-scale OSAHS home detection.The main work and achievements of this thesis are as follows:(1)The original sleep sound data of 108 subjects were recorded,including 26 subjects without the disease,33 patients with mild disease,35 patients with moderate disease and 14 with severe disease.The sound data was processed by filtering and noise reduction,endpoint detection,and intercepting snoring events.Then two real datasets were constructed,including snoring-non-snoring dataset and OSAHS-related snoring dataset,which provided a data basis for subsequent snoring detection and OSAHS recognition.In addition,multiple acoustic features were extracted from the sound data as input features for neural networks.(2)The theory of Long Short-Term Memory Spiking Neural Networks(LSTM-SNN)based on Long Short-Term Memory was studied.LSTM-SNN first performed spiking encoding on the input features.Then LSTM-SNN unit used two threshold activation functions to replace the sigmoid and tanh functions so that the network used spiking for information propagation.Finally it used the backpropagation algorithm based on gradient substitution to complete the parameter update.A single-layer LSTM-SNN network model was designed for snoring detection,and Mel frequency cepstrum coefficients were extracted as input features.Multiple sets of comparative experiments were designed to compare the influence of different model parameters on the classification results,and threshold coding and key point coding were discussed.The classification effect of the two encoding methods was analyzed based on the classification results of different gradient surrogate function models.Ultimately,the single-layer LSTM-SNN network achieved an accuracy rate of93.4% on the snoring-non-snoring test set.Compared with the LSTM network with the same structure,the parameters were reduced by 27% when the accuracy rate decreased by 1.2%.(3)An OSAHS recognition model based on multi-feature input was designed to extract four features of the snoring signal,Mel frequency cepstral coefficient,filter bank coefficient,linear prediction coefficient and short-term average energy.After learning each feature based on a two-layer LSTM-SNN setup base network,the learning results were connected to complete the classification.Aiming at the range of AHI with different prevalence levels of OSAHS,three binary classification experiments were completed on whether the AHI index was greater than 5,15,and 30,and the accuracy rates of the test set reached 89.3%,90.5%and 91.3%,respectively.Compared with the LSTM network with the same structure,the parameters were reduced by 29.9%.