Research On Cross-user Activity Recognition Method Based On Deep Transfer Learning

With the rapid development of artificial intelligence,cloud computing and other technologies,human activity recognition has become a research hotspot.It is widely used in smart home,intelligent medical and smart factory.At the same time,human activity recognition based on wearable sensors has been widely used with the development of micro sensors and intelligent hardware.There are three types of human activities: static activity(standing),dynamic activity(walking)and transitional activity(standing-to-walking).At present,most of the researches focus on the recognition of static and dynamic activities.And there are still some deficiencies in the research of transitional activity which can play an important role in practical applications,so the accurate identification of transitional activity is an important research content.In current activity recognition based on wearable sensor,most of the models are usually trained and tested on the same batch of users' behavior data.But they ignore the behavior differences between people,resulting in poor performance of the trained models when facing new target users.Although training the targeted model for each target user can achieve accurate recognition of cross-user activity,this method is no doubt unrealistic,which not only consumes a lot of cost and energy,but also greatly limits the application of activity recognition system.Therefore,to realize the recognition of user's transitional activity and overcome the problem of user's personalization in activity recognition,this paper proposes a cross-user activity recognition model based on deep transfer learning,which not only realizes the accurate recognition of static,dynamic and transitional activity,but also solves the problem of accuracy degradation caused by cross-user activity recognition.The main research work is as follows(1)In order to improve the accuracy of activity recognition model and recognize accurately human static,dynamic and transitional activity,an activity recognition model based on SDAE and Bi LSTM is proposed.The SDAE model is used to extract high level representations of behavior data and Bi LSTM model is used to mine the dependency before and after the activity time series.What's more,resampling technology is utilized to alleviate the class imbalance caused by the short-term characteristic of the transitional activity.(2)In order to improve the accuracy of cross-user activity recognition,a deep adaptive network model based on SDAE and Bi LSTM is proposed(SDAL-DNM model).The model maps the user's behavior data in the training set and the target user's behavior data to the same feature space.By continuously narrowing the distance between the two user behavior data features,the deep network can be adaptively adjusted to reduce the two kinds of user activity differences.And the performance of the cross-user behavior recognition model can be improved.