Research On Deep Network Model For Domestic Activity Recognition Based On Audio Data

At present,the research on domestic activity recognition has attracted much attention.To establish an efficient intelligent monitoring system can monitor the sudden behavior of the elderly in real time,which plays a very important role in improving the quality of life of the elderly.For the protection of personal privacy,domestic activity recognition based on audio sensor is becoming a new study of contemporary issues.Compared with the traditional method,the deep learning technology has a huge advantage.However,there are models classification accuracy to be improved,traditional convolutional not effectively distinguish between high and low frequency components of the feature maps,ordinary convolution insensitive to the relative relationship between the local feature problems.This paper will address the above problems,in order to improve the accuracy of domestic activity recognition,a suitable neural network model is proposed.Firstly,in the past,learning ability was improved by stacking the number of convolutions,which brought a test to the calculation,it can also cause over-fitting problems.To address these issue,this paper is designed based on attention convolution recurrent neural networks,used as the main building Xception convolution neural network and the recurrent neural network.In order to further improve classification accuracy,adding time attention mechanisms Experiments show that,the network model proposed in this paper has better recognition accuracy in dataset.Secondly,the traditional convolution operation does not distinguish the high and low frequency information of the feature map,which causes the problem of spatial redundancy.Introduce octave convolution,and use it as the main building of octave convolution recurrent neural network.And added Squeeze-and-Excitation module enhanced convolution channel relationship features.The SpecAugment data enhancement technology and semi supervised learning method are adopted.Experimental results show that the proposed network model is better able to identify the type of domestic activities.Finally,as convolution can only capture local features and relationships,and can’t deal with the relationship between parts and the whole well,attention capsule network is proposed in this paper.Capsule network provides a method similar to human perception system,which can identify the whole simply by identifying its parts.The temporal attention mechanism can increase the attention to the significant part by weighting.Experimental results show that the attention capsule network proposed in this paper has better effect on domestic activity recognition based on audio data than other networks.