Deep Feature Enhancement And Model Optimization Method For Activity Recognition

With the continuous development of pervasive computing,activity recognition has aroused people's extensive attention.In the field of activity recognition,the deep learning model has achieved good application results,but there are still some restrictive bottlenecks: sensor data is not intuitive.Sexuality leads to the inability to understand the depth features extracted by the model,so it is impossible to combine the sensor data and its features to improve the recognition accuracy of the model.The deep learning model usually performs batch learning in a static environment,and cannot identify new ones in the dynamic environment according to new requirements.Activity;the deep learning model requires more storage and computing resources and is not easy to calculate on the terminal device,so the model needs to be compressed.In view of the above bottlenecks,this paper studies the deep feature enhancement and model optimization methods for activity recognition.In view of the difficulty in understanding the depth characteristics of sensor data,this paper understands and analyzes the depth characteristics of sensor data extracted by deep learning model.In the experiment,it is found that there are redundant features in the depth features extracted by neurons,and different activities correspond.Different distinctive features.Therefore,based on the salient features of different activities,this paper proposes a method of manually extracting salient features,and uses the extracted salient features to obtain better recognition results in the UCIHAR dataset.Since the feature dimensions in some application scenarios are too large or too small,it is difficult to extract significant features by manual analysis.Therefore,it is necessary to study how to automatically extract the salient features of sensor data.In order to solve this problem,this paper proposes automatic extraction and enhancement of significant features.The feature model improves the recognition accuracy of the model by using the enhanced features,and is evaluated in five public activity identification data sets(UCIHAR,USHAR,OPPORTUNITY,WISDM,PAMAP2),and the recognition accuracy is high on the three data sets.For the best results in other research work,the recognition accuracy on the two data sets is basically the same as the best results in other studies.In order to enable the model to learn more new activities according to the new requirements,this paper analyzes the incremental learning in the field of activity recognition,and proposes an incremental deep learning space-time model based on the spatio-temporal features of sensor data.There are three optimization strategies for this model:(1)selecting the old sample data according to the data recognition difficulty score;(2)using the distillation loss reduction model before and after the model update;(3)extracting the spatial characteristics and time series characteristics of the sensor data Identification.The experimental results show that the recognition effect of the depth-space model is better than the convolution model based on the spatial characteristics of the sensor data,and it is also superior to the other research methods.The deep learning model contains a large number of parameters,which help the model to obtain higher recognition accuracy and also occupy more storage and computing resources.In order to reduce the parameters in the model and compress the model,this paper combines the depth of sensor data.The understanding of the characteristics,the importance of neurons is measured,and the convolutional channel pruning method based on the important scores of neurons is proposed.The results show that the method effectively reduces the parameters in the convolutional layer,and at the same time The recognition accuracy has little effect.In addition to channel pruning of the convolutional layer,the paper also prunes the neurons in the fully connected layer based on the weight of the neurons in the fully connected layer.Without affecting the accuracy of model recognition,the model as a whole can achieve 90% pruning ratio.