Analysis Of Crying Cause By Baby Cryingalgorithm

In our daily life,people's pace of life is gradually accelerating,for inexperienced newlyweds,it is undoubtedly a great challenge for them to take care of their newborn babies efficiently.According to the relevant research,the baby's crying can indicate the baby's certain needs or problems,such as hunger,pain,lethargy,discomfort,cold or heat,and so on.For a baby's crying,a small number of experienced parents may be able to understand the exact meaning of crying,but it is obviously impossible for most people.With the development of artificial intelligence and voice technology in recent years,it is possible to solve this problem.The research goal of this thesis is to study a variety of algorithms that can automatically identify the causes of baby crying,so that parents can more easily understand the specific meaning of baby crying,so as to help babies grow up more healthily.The following is a summary of the research content of this thesis:First,since most of the existing infant crying data sets at home and abroad are not public,and the types of infant cries collected are different from each other,it is necessary to establish their own infant crying data sets for related research work.Through the study of Dunstan Baby Language(DBL)and other theories,we mainly collected five types of baby crying data on Youtube and other websites,and established our own baby cry data set after preprocessing and other steps.According to the theory proposed by DBL,the infant crying cause recognition algorithm studied in this thesis is mainly divided into five categories: hunger,sleepy,burp,gassy and discomfort.Secondly,this thesis studies a codebook-based algorithm for identifying the cause of infant crying,which uses the MFCC of the characteristic parameters.It is mainly divided into two stages,one is the codebook production stage,this stage mainly uses the KMEANS clustering algorithm,the other is the recognition stage,this stage mainly uses the distance function and the KNN nearest neighbor algorithm.And again,it is divided into four dimensions: frame length,frame overlap,the number of clustering and the number of neighbors,and a series of accuracy tests are carried out,and a model parameter with high precision is obtained.Compared with the traditional codebook algorithm,the average recognition rate has been improved by about 10%.Thirdly,this thesis studies a baby crying cause recognition algorithm based on neural network,which also uses MFCC feature parameters.The neural network model is composed of three hidden layers.in the experiment,the activation function is compared with Relu activation function,tanh activation function and sigmoid activation function.Then the advantages and disadvantages of different activation functions are summarized and analyzed,the changes of accuracy are compared and the previous theoretical analys is is verified.Finally,a comparative experiment is carried out with the traditional LVQ n eural network.The results show that the highest recognition accuracy obtained by using this neural network structure has been improved by 4%.Fourth,this thesis studies a deep learning infant crying cause identification algorithm based on MFCC feature parameters.This algorithm also uses MFCC feature parameters,and the deep learning model is changed on the basis of WaveNet network model.And the principles of causal void convolution,residual connection and other main structures in the model structure are explained in detail.And part of the structure graphics of the whole network model after modification are given.Finally,a comparative experiment is carried out between the model and the traditional convolution neural network,and their advantages and disadvantages are analyzed and summarized,which proves that the network has strong speech learning ability.Fifth,this thesis studies a deep learning baby crying cause identification algorithm based on spectrogram,which uses the spectrogram as the input feature parameter,and explains in detail the concept and principle of the spectrogram and the advantages of using it.The selected deep learning model is obtained by changing the Inception-v3 model.And two improved ways of Inception-v3 network structure are explained in detail,and the training method used in the research is pre-training.Then the steps of the algorithm are introduced in detail,and experiments are carried out on this basis,and the results show that the accuracy of the algorithm is greatly improved compared with the previous three algorithms.Finally,a comparative experiment is carried out with the traditional convolution neural network.the superiority of the algorithm in recognition rate is verified.