Research On Automatic Recognition Methods Of Heart Sound Diagnosis Information

Heart sound diagnostic information can be used to diagnose the pathological condition of an individual's heart or other parts of the body.During the clinical auscultation,the doctor can only analyze the patient's current heartbeat.It is easy to misjudge due to the patient's mood,body movements and other changes in the way the heart beats.Therefore,long term recorded heart sounds may be more suitable for exact diagnose.Certainly the workload of long-term heart sound signal analysis is undoubtedly huge.Taking into account the excellent characteristics of machine learning and deep learning algorithms,we combine deep learning algorithms,machine learning algorithms and heart sound diagnosis together to study the machine aided processes and methods of eff-icient identifying abnormal heart sounds.The research content includes:First,in order to study the recognition of heart sound signals by the deep learning network,the heart sound signals are converted into two-dimensional images,and the deep learning network is used for training and modeling.Image conversion involves two methods,one is the common spectrogram format,the other is the MFCC(Mel-frequency Cepstrum Coefficient)heat map format.Secondly,through the network construction for the picture data and the adjustment of related parameters,the recognition performance of the shallow convolutional neural network and the deep convolutional neural network for the pictures in these two formats are compared and analyzed.Then,some feature data is extracted from the network,and the extracted feature data is trained,classified,and compared with the machine learning classifier.Then,in order to further improve the recognition ability of the heart sound,multi-features extraction of the heart sound signal is studied.To avoid the performance of the subsequent processing due to the incorrect segmentation of the heart sound signal,the heart sound signal will not be periodically segmented.Since the heart sound signal is a non-stationary signal,the conventional framing operation is performed on the heart sound signal to ensure that each frame signal is approximately a smooth signal for processing.Six characteristic parameters are extracted for each frame signal,which are 18 dimensions in total.The machine learning classifier was used to classify and identify the feature data,and multiple sets of comparison experiments were then carried out.By comparing several evaluation indicators,the model XGBoost becomes the best one in the data set selected in this paper.Finally,according to the data distribution characteristics of the signal features,the features are processed separately,and the MFCC features are pre-classified using the Gaussian mixture model.The pre-classified results are used as the weights of the MFCC feature data,and combined with the non-MFCC features.Through these operations,we achieve the best experimental results based on the data set selected in this paper.