Recognition Of Ship Radiated Noise Based On Transfer Learning And Auditory Perception Characteristics

Use the ship radiated noise to complete target recognition tasks is one of the key technologies of underwater acoustic signal processing,it also the has been widely concerned by the academic community.Ship radiated noise is mainly composed of mechanical noise,hydrodynamic noise,and propeller noise.The line spectrum and continuum information contained in its time-frequency domain characteristics can effectively reflect the characteristics of the mechanical structure and propeller structure of different ships,this is the very basis of ship identification.However,the extracted feature quantity of traditional feature extraction method can be quite redundancy,it requires manual processing and calculation of features,and the operation process is also cumbersome and inefficient.Traditional classifiers are too sensitive to sample abnormal points.Also,the recognition rate of complex classification problems under the interference of marine environment and other factors is insufficient.With the development of deep learning,the research of its application on ship identification has gradually increased.Deep learning methods have feature processing capabilities,it can perform batch training on data.Also,deep learning is not sensitive to abnormal points and higher recognition rate performance in complex situations.However,deep learning methods rely heavily on data,and it is difficult to directly use complex network models in the computer field for the current problem of insufficient samples of underwater acoustic data.Specific underwater acoustic recognition problems should be specifically designed for the model.As one of the important deep learning ideas,transfer learning is often used to solve the classification problem of small samples or even unlabeled samples,which is suit for with status of underwater acoustic target recognition.Based on the above-mentioned problems of underwater acoustics field,this article first starts from the principle of ship radiated noise generation,analyzes its components and features in the frequency domain,uses auditory perception features to replace traditional feature extraction methods,this methods can effectively solve the number of traditional time-frequency domain features which have great impact on subsequent recognition and classification.Then use the deep learning network framework to classify the target features,explore the influence of different deep learning method structure design methods and recognition rate when solving the underwater acoustic target recognition problem.At the same time,to solve the problem of too few underwater acoustic data training set samples which cause the recognition rate is unstable,a threshold termination algorithm classifier is proposed to early stop the recognition network training in order to obtain a stable recognition rate.Finally,this paper proposes the use of multi-dimensional feature fusion algorithm combined with transfer learning method to improve the stability of the system.This method uses simulated ship noise data to pre-train the network model and helps training small sample data.It can fully extract the characteristics of ship radiated noise and effectively alleviate the problem of instability in recognition results during the small sample training process.Comparing and verifying methods through simulation data and measured sea trial data,the method proposed in this paper has a more stable recognition rate result than traditional methods when training with small sample data,and obtains a conclusion with certain reference value.