Activity Recognition Based On Mixed Attentional Mechanism And Lifelong Learning

In recent years,activity recognition has received increasing attention within a variety of fields such as health,smart healthcare,smart home,and smart driving.Activity recognition is full of significance and challenges within each field and needs to be studied deeply in order to provide good assistance to users.Based on the big data generated by sensors and the rapid development of hardware GPUs,deep learning is increasingly researched in the field of activity recognition.Activities can be represented as time-series data,and the one-dimensional time-series data is converted into a two-dimensional tensor representation of activities by sliding window segmentation,which is then input to CNN for activity recognition.To express the time series data temporally,RNN is used for sequence-to-sequence processing,and its variant LSTM is the main method for activity recognition research.The use of LSTM for sequence processing has two drawbacks;first,it cannot process sequences in parallel;second,it cannot remember longer sequence information well.Through research,the attention mechanism is introduced to explore the relevant temporal context,which effectively solves the shortcomings of LSTM and proposes the Deep Conv Attn model.The conventional feature extraction layer is viewed equivalently for all features and does not pay much attention to the important features.In this paper,based on the problems existing in the feature extraction layer of Deep Conv LSTM model,Deep Conv Attn2.0 model is further proposed.Through analysis,the visual attention mechanism is incorporated into the first and last convolutional layers,which can autonomously learn the weight distribution of the feature map,allowing the model to focus more on important underlying edge features and high-level semantic features,and suppress features that are not important to the model.The experimental results show that the Deep Conv Attn2.0 model improves the F1 score by 3.57% compared with the Deep Conv LSTM model on the Opportunity activity recognition dataset.The first research point of this paper,hybrid attention mechanism allows neural network models to focus more on important features and suppress non-important features during feature extraction and sequence processing,similar to how human vision will focus on a certain region when looking at something;the second research point of this paper,lifelong learning allows neural network models to have a learning mechanism similar to the human brain.At present,most of the lifelong learning research focuses on the design of loss function,which is used to constrain the knowledge of old and new tasks through the loss function.The knowledge distillation loss,on the output layer of the old task,distills more useful knowledge to guide the model to learn the new task by warming up.This method does not consider the importance of hidden layer knowledge.The elastic weight curing loss uses the second-order gradient of the old task parameters to constrain the old and new tasks,which has the disadvantage of being computationally intensive.In this paper,we propose a smoothed distillation guard loss to address the shortcomings of knowledge distillation loss and elastic weight curing loss,and do four sets of experiments to compare them.Based on two subtasks of the Opportunity activity recognition dataset,i.e.,18 classes of gestures old task and 5 classes of locomotion new task,the Deep Conv Attn2.0 model is used for the comparison of multiple sets of experiments.The experimental results show that the F1 scores of the training dataset after mixing the new task data,the new data of the old task and part of the old data of the old task using back-propagation gradient optimization with smoothed distillation guard loss outperform the results of single-task training and testing.