Application Of Deep Neural Network In Mobile Environment Sensing System

Environmental perception is an act of recognizing the existence of sound in the environment,and it is an important field of artificial intelligence research.Due to the enhancement of the computer,especially after the popularity of mobile applications,through the mobile side to identify the environment in different sources of the scene,academics have started to study from the human voice as well as other basic sources and so on,which the analys from the perspective of source is the main content.Deep learning is the hottest field of artificial intelligence in the current development,huge breakthroughs are generated in speech recognition,image recognition,and natural language processing.With the advancement and development of deep learning technology,there are some very feasible models and algorithms,such as convolution neural network CNN,cyclic neural network RNN,depth neural network DNN and recurrent neural network,how to improve the accuracy of environmental perception recognition in environmental perception by using the depth learning theory is the frontier problem of artificial intelligence research.In view of the above problems,this paper proposes an environment-aware model based on the combination of CNN and depth RNN.Combining the advantages of both,using the sound spectrum transformed from the recorded background environment audio file as input,and then use the structural characteristics of CNN to automatically extract the specific characteristics of the environment from the spectrum diagram,and finally input the feature vector extracted from the CNN to the depth RNN to complete the classification recognition.For the above model,this paper has also done the relevant work:1.In order to reflect the authenticity of the model,a series of environmental audio sample datasets were collected by device recording,and using this data set to train the network structure in this model.In order to verify the effectiveness of the proposed model,this paper mainly compares the traditional MFCC + K neighborhood and MFCC + GMM systems and the model in the same data set.The results show that the model proposed in this paper is an effective environment-aware model.2.In this paper,the network structure of the proposed model is composed of CNN and deep RNN.In order to study the model of the network structure and parameter settings whether can achieve good recognition results for the identification of environmental audio data sets in this paper,three sets of comparison experiments are made for the network structure of this model,including the use of different activation function,different size convolution kernel and abandoned the depth of the RNN and added the full connection layer of the model system to make comparison experiments in the study of audio recognition rate.Through the experimental results on the data set of this paper illustrate the validity of the network structure and parameter setting of the model,and having more advantages than other forms of neural network.Finally,compar with other paper’s depth neural network models,verifing the advantages of the proposed model in the same theoretical model.3.Based on the Tensorflow depth learning framework,the network structure in this model is realized,and copied the training of the network model to the mobile side Android platform,and tested the effectiveness of the system through the sample and the actual ambient audio.