Underwater Target Recognition Technology Base On Deep Learning

Target decting,tracking,recognition and localization are several important functions of sonar system,playing an important role in battlefield situation control and underwater combat.Along with the informatization and intelligent advancement of naval equipment,modern sea warfare is turning into a combination of multiple combat platform and warship formation,which means a much more complex underwater situation.In other words,the underwater target recognition performance means the fighting capacity and surviving probability of surface warships and warship formations.Meanwhile,the artificial intelligence,especially machine learning and deep learning technology,has been booming in recent years,which may provide several new methods to underwater target recognition.Thus research on underwater target features,based on big data,pattern recognition and machine learning(deep learning),is a new trend of underwater target recognition technology.Firstly,the technique background of this research topic is analyzed in detail,especially the underwater target recognition technology,the artificial intelligence technology as well.In the second part of this thesis,related theories of this article are proposed.The navigation noise is selected as the research target,and two auditory features,the Mel Frequency Cepstrum Coefficient and Gammatone Cepstral Coeffiscient,are introduced to underwater target recognition.Together with the power spectrum density feature,the auditory features are primarily proved to be available in recognition.And the fundamental theories of machine learning,pattern recognition are introduced,and the hold-out method is selected to evaluate the performance of recognition models.In the third part of this thesis,several recognition models,based on K-Nearst Neighbor and Support Vector Machine are established and trained with MFCC and GTCC sample set.The experiment result shows that the MFCC and GTCC features are effective when used in underwater target recognition.In the fourth part of this thesis,a new method named as time-sequence method for deeplearning algorithm is proposed,which means to creat time-sequence samples of higher dimention for deeplearning algorithm,by stacking target samples along the timeline.And the power spectrum density(marked as PSD)sample set,time-sequence Mel frequency cepstrum coefficient(marked as TSMFCC)sample set,time-sequence Gammatone cepstral coefficient(marked as TSGTCC)sample set,time-sequence power spectrum density(marked as TSPSD)sample set,are analysised with two deeplearning algorithm,CNN and DBN,and the experiment result shows that the time-sequence method with deeplearning algorithm is effective in underwater target recognition.In the fifth part of this thesis,several experiments,related to the engineering requirements,are designed for the recognition models of the third and fourth part of this thesis.Recognition performance of various models is tested,in unknown underwater target type,different navigation environment and signal noise ratio.In the sixth part of this thesis,a fast recognition method to underwater target is designed,according to the experiment results in the fifth part of this thesis.Meanwhile,an intelligent learning algorithm based intelligent recognition system for underwater target is proposed,and the pretrain module,together with the automatic recognition module is designed separatly.Furthermore,the graphical user interface of the intelligent recognition system is designed,and an experiment to recognitze a new kind of underwater target is conducted,the recognition speed is tested as well.As a combination of different features and different learning algorithms,the intelligent recognition system is believed to be competent to unknown underwater target,different navigation environment and signal noise ratio,meanwhile has many special advantages in processing speed and man-machine interaction.