Research On Activity Recognition Of Smartphone Sensor Based On Extreme Learning Machine

Activity identification is a typical time series classification problem,which is exactly to discover the human activities generating the continuous part of the sensor data flow.Activity identification can be widely applied in many areas such as human-computer interaction,health care,education,teleconference,sports and so on.Currently,the smart phone embedded with a variety of sensors has become the ideal platform for identification and monitoring.And Extreme Learning Machine has been a more advanced research method in this field after years of research.From perspectives of supervised learning,semi-supervised learning and unsupervised learning,this paper studies ELM and its improved algorithm for further optimization of the ELM.In addition,with the help of three mobile phone datasets,we can verify the feasibility and effectiveness of ELM algorithm in sensor activity identification.Firstly,this paper studies the ELM supervisor learning model(H-ELM),and combines the H-ELM with the deep reconstruction models(DRMs).In the process of initializing the DRM template parameters by H-ELM,the author adopts a new method which is to combine the weight matrix and the bias matrix to simplify the stochastic process steps.The experimental results show that the accuracy of the method can be as high as 99% for the sensor dataset.Secondly,the ELM semi-supervised learning model SS-ELM is studied,the results show that the proposed method can improve the accuracy of the activity recognition,and the training time is shortened about 10% by the combination of PCA and SS-ELM algorithm,which is to reduce the redundancy between sample data and improves the activity recognition effect.Finally,as for the unsupervised learning algorithm US-ELM of ELM,this paper mainly studies the algorithm principle and the solution process.The results show that US-ELM can play a role in clustering the sample data.On the whole,ELM has a high correct rate of recognition in the field of active identification of smartphone sensors,which has a strong potential in common fields.