Research On Classification Of High-dimensional Small Sample Data Based On Bayesian Deep Learning

With the rapid development of deep learning,important progress has been made in data classification based on a large number of samples.On many data sets,it has surpassed human recognition.However,in reality,the number of samples for many problems is extremely scarce or the scene is difficult to reproduce,and the data used for training is very scarce.Traditional deep learning methods are not effective in this case.And Bayesian learning can use the prior distribution of parameters and a small number of samples to estimate the posterior distribution,and realize machine learning under the condition of small samples.This paper is mainly based on the Bayesian variational auto-encoder network combining Bayesian learning and deep learning.It starts from the two directions of data augmentation and neural network construction,and studies the image classification method under small samples.The main work includes the following aspects:The existing basic data enhancement methods and data generation methods are studied.Bayesian variational auto-encoders and generative adversarial networks are improved.A Bayesian variational generative adversarial network model for small sample data generation is presented.The generator uses a Bayesian variational auto-encoder.The hidden layer sampling process of feature extraction can effectively avoid overfitting problems due to small bands of sample data,and the reconstruction loss function restricts the reconstructed picture to be consistent with the input picture.The problem of generating a model against adversarial networks collapsed.Through different data generation methods tested on the same classification network,the data generated in this paper can be used to train classification networks,and has better performance in data classification scenarios than other data enhancement methods.The existing metric network and transfer learning methods suitable for small sample data classification are studied,and the Bayesian variational auto-encoder is used to improve the twin network.A Bayesian twin variational classification for small sample data classification is established.The self-encoder neural network first uses a pre-trained Bayesian variational auto-encoder to perform dimension reduction processing on the training sample data,extracts the hidden features of the samples,and then matches the two samples in pairs to the similarity matching input twin network Returns the same sample category and accuracy rate as the label of the sample to be tested,and completes the classification of small sample data.Compared with other classification methods on the standard data set,the Bayesian twin variational auto-encoder proposed in this paper has better classification accuracy,and has higher mean and lower standard deviation,and the classification performance is more stable.The data enhancement method and sample classification method given in this paper are applied to transformer partial discharge pattern recognition.Experiments show that this method has higher recognition accuracy than the comparative classification method.The application of complex problems in actual scenarios proves that this sample is small Expansion of data classification methods.