Research Of User Activities Recognition Model Based On Sensor Data

About a decade ago Wireless Sensor Network(WSN)technologies and applications led to the introduction of Body Sensor Networks(BSNs): a particular type of WSN applied to human health.Since their inception,BSNs promised disruptive changes in several aspects of our daily life.At technological level,a BSN comprises wireless wearable physiological sensors applied to the human body(by means of skin electrodes,elastic straps,or even using smart fabrics)to enable,at low cost,continuous and real-time non-invasive monitoring.Indeed,in the last few years,its diffusion increased enormously with the introduction at mass industrial level of smart wearable devices(particularly smart watches and bracelets)that are able to capture several parameters such as body accelerations,angular velocity.However,since many BSN applications require sophisticated signal processing techniques and algorithms,their design and implementation remain a challenging task still today.In order to overcome the difficulty of this task,this thesis proposes a model based on sensor data from smartphones to recognize human activities.This thesis is divided into the following research points for detailed research:(1)Collecting multiple datasets.A total of three datasets are used in this thesis,each of which has a different size,and the methods and equipment used to collect the data are also different,in order to improve the robustness of the model proposed in this thesis.(2)Preprocess unprocessed sensor data.Since the important evaluation of user activity recognition is mainly accuracy,it is necessary to reduce the influence of other factors,normalize the data to unify the data format,and perform other pre-processing operations before input.(3)Aiming at the singleness for existing identification methods,inspired by the direction of sequence data visualization,the time series data based on sensor is converted into image data by the GAF imaging algorithm.Among them,according to the special situation in user activities recognition problem,an upgraded GAF imaging algorithm is proposed.This makes it possible to effectively handle special cases while guaranteeing the performance of the original imaging algorithm.(4)For multi-source heterogeneity of sensor data,this thesis use multiple sensor data(such as simultaneous acceleration and angular velocity data)for coordination training.In order to meet this training requirement,a fused convolutional neural network,a deep fusion residual network,is proposed so that the hidden information in heterogeneous sensor data can be fully mined.(5)The proposed model is compared with various machine learning methods(traditional machine learning and deep learning methods),and the experimental results are analyzed.The accuracy rate of the activity recognition methods proposed in this paper is significantly higher than other methods,which can be effective in a variety of situations.To recognition the activity of smartphone users,it has good universality.