Research And Application Of Activity Recognition Based On Sensor Data And Deep Learning

With the progress of science and technology and the rapid development of sensor technology,sensor-based human activity recognition(HAR)is widely used in intelligent life,health monitoring,health care,smart city and other fields,and plays an important role.In recent years,with the rise of various types of intelligent mobile devices and wearable devices,due to its low cost,small size,easy to collect user data,so the way of activity recognition is becoming more and more simple,through the use of sensor signals to monitor the physiological status of users,provide motion state assessment and activity advice,for the elderly or patients with mental illness,can detect their abnormal activities,and timely rescue work.The application of activity identification in daily life can not only reduce the cost of care,but also reduce the human and financial expenditure of the government and society in social security such as pension.The application of HAR in personal health assistance has an immeasurable application prospect for promoting national health and promoting the construction of smart cities.From the need to wear multiple complex monitoring equipment to carry a single intelligent device,but the reduction of monitoring equipment also leads to higher requirements for the accuracy of activity recognition.At present,most domestic research uses specific data collection methods to identify specified activities,and improves classification accuracy by specific division of initial data.However,the single structure of the model itself lacks accuracy and robustness for different activity data sets.Aiming at the problems of single structure,poor sensitivity to data features,insufficient feature extraction ability and low accuracy of activity recognition in traditional activity recognition models,the convolutional neural network(CNN)is used to automatically extract data features.The CNN depth models with different structural types are reduced by convolution and applied to activity recognition research.The structure of the model is modified and the convolution of each layer is batch normalized to improve the complexity and generalization ability of the comparison model.The results of each model are compared and analyzed on the activity recognition dataset with different features.It is found that compared with the traditional activity recognition model,the CNN depth model convolution dimension reduction is applied to the field of activity recognition,which can effectively improve the accuracy of model recognition.After dimension reduction,some models can effectively improve the accuracy of activity recognition.Due to some problems such as too many model parameters and too complex structure,practical application still needs to be considered.In order to solve the above problems,a multi-channel CNN model with attention mechanism is proposed by comparing the characteristics of various CNN models.The multi-channel method is used to broaden the width of the network,increase the adaptability of the network to the scale of the convolution kernel and construct the CNN model.The different scale information of the data is extracted by using multiple convolution kernels.Finally,the feature fusion is carried out to deal with more and richer features.In the network structure of the same depth,the number of parameters can be reduced by adopting multi-channel modular structure.The convolution module is stacked,the convolution layer is paralleled,and the width of the parameters and the model module is adjusted to obtain more data features.The accuracy of the algorithm is improved by constructing an identity map for each module.In order to further improve the accuracy of the model for activity recognition and classification,the soft attention mechanism model is combined with the CNN model proposed in this paper to improve the attention of the model to the important features of the spatial domain and channel domain,and strengthen the expression of information features.The weight coefficients of different feature dimensions are calculated,and then the coefficients are multiplied by the corresponding elements to form a new feature.The experimental comparison shows that the accuracy and F1 value of the multi-channel CNN network model based on the attention mechanism under different sensor data are improved.Finally,in order to verify the improvement effect of soft attention mechanism on CNN network model,several soft attention mechanism models are embedded into the selected CNN model,and comparative experiments are carried out.The human activity recognition analysis and prediction system is designed and implemented.The sensor data are extracted and visualized.By processing the deep learning network model file,the sensor data results are analyzed and the activity is predicted.