Research On Detection Technology Of Dangerous Sound Based On Deep Learning

Nowadays,people's lives are full of various sounds,some of which are reflections of dangerous events and can warn people of danger,like explosions,gunshots,screams,etc.Therefore,the dangerous sounds detection has potential application value.In recent years,more and more experts and scholars have studied the dangerous sounds detection,which has gradually become an important part of audio signal processing.At present,most of the research focuses on artificial selection features and traditional machine learning methods,like MFCC features,support vector machines,gaussian mixture model,etc.These methods face the problem that difficult to select good features and solve complex classification.Although there are some studies currently trying to use deep learning methods,these network models have a single structure,a simple hierarchical structure,and poor portability.In this paper,we have studied the detection method of dangerous sounds based on deep learning,designed a variety of network models,compared with the baseline system using traditional methods,and gradually improved the accuracy of detection of dangerous sounds.The main work contents and innovations of this article are as follows:This article expounds the basic theoretical knowledge of deep learning,designs a framework for detection of dangerous sound based on features and classifiers,and builds a baseline system based on traditional MFCC features and GMM classifiers.The system was developed and tested on the unified dangerous sound data set in this paper,and the detection accuracy rates of the training and testing groups were 77% and 68%,respectively.This method can preliminarily be qualified for the task of detecting dangerous sounds and use it as a control group for the later detection system.This paper designs and builds a system about the dangerous sounds detection based on deep learning model,selects 64-dimensional log-Mel spectrum as input features,and designs DNN model,CNN model and hybrid model as classifiers.The DNN model uses three hidden layers,its activation function uses the ReLu,and the Dropout layer is added to prevent overfitting during the training process.The CNN model is a modification of vgg-16.In order to prevent overfitting,besides adding Dropout layer,BN mechanism is also used in the network.Considering the timing of sound data,the hybrid model is mainly implemented by adding CNN and RNN,in which the CNN part is the modification of the CNN model,and the RNN part uses ordinary RNN and LSTM.After the system was developed and tested on the data set of this paper,detection results were obtained.The accuracy of the system which using DNN,CNN,C-RNN,and C-LSTM models were 74.5%,86.2%,90.0%,and 91.6 %.Comparing the three deep learning models,the hybrid model has higher detection accuracy than the single DNN and CNN models.However,compared with C-RNN,the detection effect of C-LSTM has little improvement.