Research On Activity Recognition And Transfer Learning Method Based On GAN

In the context of big data,the lack of data remains a common and serious problem.Because the diversity of transactions cannot be fully covered in a limited training set,or the related scenes rarely appear in real life,it will lead to the lack of sample data for model.There are many solutions to this problem.The main solution of this paper is to oversample the missing data,witch enriching the sample.The task of this paper is based on the human activity recognition task with sensor data,and puts forward solution focusing on two problems caused by the lack of data in this field.The first problem is unbalanced data in the same domain.There is a problem of imbalance in the amount of data in the activity category in the training set.The second problem is the domain adaptation problem in transfer learning focusing on people in the field of activity recognition.The both of problems contain the problem of lacking data.In the first part,the paper conducted experiments in three public data sets.First of all,the problem description and the analysis of the confusion matrix are found.When the classifier is trained in the unbalanced data,the model is likely to have a classification error for a small number of categories in the test.The solution proposes in this paper is to use the confrontation generation network to oversample the minority data.The sensor data analysis then notes that the sensor data has the characteristics of a time window.According to the characteristics of the data,the model called BAGAN-SSIM is proposed a model for generating a small amount of data.In the second part,this paper is aimed at the domain adaptation problem in migration learning.The necessity of migration learning between people in the field of activity identification is expounded,and a model of confrontation migration network is proposed for this problem.In the process of data analysis,it is found that the feature distribution of the confusing categories of the original domain and the target domain data classification will be very similar.According to this point,the requirement of opening the class spacing is proposed.The cosFace function is used as the loss function of the model through experimental comparison.In the final part,the confrontation model is transformed into an online learning model.The accuracy of the model in the online case has decreased.Experiments show that the accuracy of the model after migration is reduced with the type of input tags,and the accuracy of the model remains above 75%.