Research On Human Activity Recognition Based On Multi-feature Weighted Ensemble

The popularity of micro-sensing smart devices allows people to record daily activity data anytime and anywhere.Human activity recognition has gradually become a research hotspot and is widely used in medical rehabilitation,smart homes and so on.At present,deep learning and traditional machine learning are often used for human activity recognition.Although the former can automatically extract features,they are often easy to overfit on small and medium-scale activity data set.The traditional machine learning methods do not require a large amount of samples,but the features to be extracted manually,and the features largely determine the recognition effect.Therefore,this paper studies the recognition of human activity on small and medium-scale data set.The main work and contributions are as follows:(1)An improved ensemble learning method(MFWE)based on Bagging is proposed,which uses Multi-view instead of Bootstrap sample to ensure the diversity of the base learners,making it more suitable for small and medium-scale data set.At the same time,considering the performance difference of different learners,a reasonable weighted voting ensemble strategy is proposed to obtain the final prediction result.(2)The MFWE method is used for human activity recognition,and a multi-feature weighted ensemble activity recognition algorithm(MFWE-HAR)is designed,which uses three types of global features including time domain features,frequency domain features and other features and two types of local features.MFWE-HAR algorithm construct a KNN classifier for each global feature,use 1D-CNN and Attention-LSTM to extract local features and classify,and obtain the final result through a weighted ensemble strategy of the MFWE.Thus,the entire model has a stronger feature representation ability,and complements the advantages of traditional machine learning methods and deep learning algorithms,and improves the recognition performance on small and medium-scale activity data set.(3)Eight types of activity data of four testers were collected through wearable sensors,and construct a sample set used in the experiment by data segmentation.(4)The effectiveness of the MFWE method and the MFWE-HAR algorithm is proved through five confirmatory experiments.(5)The influences of different feature ensemble,different K value in KNN and unbalanced distribution of data samples on the recognition performance are analyzed through three exploratory experiments.In conclusion,this paper proposes the MFWE method,designs the MFWE-HAR algorithm,and conducts experiments on the collected human activity data set.These experimental results show that the MFWE method is more suitable for small and mediumscale data sets than the traditional Bagging method.The MFWE-HAR algorithm uses a variety of features,and complements the advantages of traditional machine learning methods and deep learning algorithms,it has stronger feature representation ability and better recognition performance compared with single feature classification algorithms and other ensemble learning algorithms for human activity recognition.