Research On Classification Of Unbalanced Time Series Data Based On Metric Learning

In recent years,the related research of deep learning algorithms has developed rapidly,and significant research results have been achieved in many fields such as computer vision,natural language processing,and speech recognition.With the rapid increase of sample data sets and the improvement of GPU performance,the performance of deep models has been rapidly improved,and people have gradually realized that the size of the data set has a significant impact on the performance of deep learning models.However,large-scale data sets often have sample imbalances,that is,there are significant differences in the number of data samples between different categories.This imbalance has a certain degree of restriction on the performance of deep network models,and even becomes a model performance improvement.Bottleneck.In the face of complex problems,the deep network model encounters some unbalanced data and its excessively large unbalanced proportions.It is difficult to solve by traditional model optimization methods and data processing methods.If there are other historical relationships between data samples(For example,timing,spatial relationship),then the problem will be more complicated.Aiming at the task of tropical cyclone(TC)intensity estimation,our paper studies the temporal correlation and imbalance of tropical cyclone image samples.The main work of this paper is as follows:1.Construct the tropical cyclone training data set of satellite images.The original data in this paper comes from Japan’s Himawari-8.After data analysis and annotation,a total of 17244 samples from 2005 to 2019 are constructed,and the distribution statistics and analysis are carried out.It is found that the tropical cyclone image differences between typhoons with similar intensities are very small,and the distribution is unbalanced;in addition,the same tropical cyclone Changes and develops at different times with the constraint of adjacent tropical cyclone.2.According to the above problems,we propose a method of tropical cyclone intensity estimation via Siamese Network.In this method,Siamese network is used to extract the image-to-intensity features of tropical cyclone at adjacent time.The tropical cyclone feature distance is constrained in the high-level feature embedding space to improve the data imbalance influence and the temporal correlation.Experiments shown that our method improve the accuracy of typhoon intensity estimation.3.Furthermore,a tropical cyclone intensity estimation method based on triple loss is proposed,which uses Triplet Loss to better constrain the distance relationship between samples in feature space,so that the distance between tropical cyclone samples of the same category is more compact,while the distance between different samples is more distant.