EI Compendex Source List（2020年1月）
EI Compendex Source List（2019年5月）
EI Compendex Source List（2018年9月）
EI Compendex Source List（2018年5月）
EI Compendex Source List（2018年1月）
中国科学引文数据库来源期刊列
CSSCI(2017-2018)及扩展期刊目录
2017年4月7日EI检索目录（最新）
2017年3月EI检索目录
最新公布北大中文核心期刊目录
SCI期刊（含影响因子）
EI Compendex Source List
论文范文
1. Introduction Deep learning is a new area of machine learning research aimed at establishing a neural network that simulates human brain analysis and learning. The concept of deep learning was proposed by Hinton in 2006 [1]. In recent years, deep learning has drawn wide attention in the field of pattern analysis and has gradually become the mainstream method in the fields of image analysis and recognition and speech recognition [2]. In recent years, some people apply it to the speech signal denoising [3] and the dereverberation problem [4]. Convolutional Neural Network (CNN) is one of the core methods in depth learning theory. CNN can be classified as deep neural network (DNN) but belongs to supervised learning method. Y. LeCun proposed CNN is the first real multilayer structure learning algorithm, which uses space relative relationship to reduce the number of parameters to improve training performance, and has achieved success in handwriting recognition. Professor Huang Guangbin from Nanyang Technological University first proposed extreme learning machine in 2004. Extreme learning machine (ELM) has been proposed for training single hidden layer feed forward neural networks (SLFNs). Compared to traditional FNN learning methods, ELM is remarkably efficient and tends to reach a global optimum. In this paper, CNN and ELM methods are introduced into underwater acoustic target classification and recognition, and a underwater target recognition method based on depth learning is proposed. In view of the prominent performance of convolutional neural network in speech recognition and its frequent usage in speech feature extraction, the convolutional neural network is used to extract the features of underwater ship’s sound signal, and the corresponding network model and parameter setting method are given. This method is compared with Support Vector Machine (SVM) [5] and k-nearest neighbors (KNN) [6], Hilbert-Huang Transform (HHT) [7], and Mel frequency cepstral coefficients (MFCC) [8] methods. The experimental results show that underwater target recognition based on depth learning has a higher recognition rate than the traditional method. Then ELM fed with the CNN features is used as the classifier to conduct an excellent classification. Experiments on the actual data set of civil ships obtained 93.04% recognition rate; compared to the traditional Mel frequency cepstral coefficients and Hilbert-Huang feature, recognition rate greatly improved. 2. Related Work The current development of passive sonar high-precision underwater target automatic identification method to prevent all types of water targets in the raid is to strengthen the urgent task of the modern war system. A series of theoretical methods and technical means involved in the automatic identification of underwater targets can be applied not only to national defense equipment research but also to marine resources exploration, marine animal research, speech recognition, traffic noise recognition, machine fault diagnosis, and clinical medical diagnosis field. In the 1990s, researchers of all countries applied artificial neural network into underwater target recognition system. The methods such as power spectrum estimation, short-time Fourier transform, wavelet transform, Hilbert-Huang transform, fractal, limit cycle, and chaos [8–11] failed to fully consider the structure features of sound signal and the features extracted by such methods have prominent problems, such as worse robustness and low recognition rate. In 2006, a paper published on Science, the world top level academic journal, by Geoffrey Hinton, machine learning master and professor of University of Toronto and his student Ruslan, aroused the development upsurge [12] of deep learning in research field and application field. Since then, many researchers began working on deep learning research. In 2009, Andrew Y. Ng, etc. [13] extracted once again the features of spectrogram by using convolutional deep belief networks, and all of the results derived from use of extracted features to multiple voice recognition tasks are superior to the ones of the system which recognizes by taking Mel frequency cepstrum coefficient directly as a feature. In 2011, Hinton used restricted Boltzmann machine [14] to learn the waveform of original voice signal to obtain the distinguishable advanced features. The experiment shows that such method is better than the traditional Mel frequency spectrum feature in performance. In 2012, four major international scientific research institutions summed up the progress made by deep learning in sound recognition task [15] and pointed out in the paper that the experiments show the effect of deep neural network is better than the one of traditional Gaussian mixed model. In 2013, Brian Kingsbury, researcher of IBM, and several other persons [16] took the logarithmic Mel filter coefficient as the input of deep convolution network and further extracted the “original” features (Mel filter coefficient), and the experiment shows that the recognition rate has a relative increase of 13–30% compared to the traditional Gaussian mixed model and has a relative increase of 4–12% compared to deep neural networks. The research of Palaz et al. [17] shows that the system of using original voice signal as the input of convolutional neural networks to estimate phoneme conditional probability has achieved similar or better results than the TIMIT phoneme recognition task of the standard hybrid type HMM/ANN system. In 2014, Palaz et al. [17] put forward the convolutional neural networks of restriction right sharing mechanism and once again achieved amazing results. In 2015, three artificial intelligent masters, LeCun et al. [18], jointly published an overview titled “deep learning” on Nature, giving a comprehensive introduction on the theory of deep learning. Nowadays, deep learning has become the research hot spot of the world. Considering the outstanding performance of convolutional neural networks in voice recognition, this article uses convolutional neural networks to conduct feature extraction and recognition to underwater ship voice signal. 
|
|
|
