Research On Source Localization Based On Active Deep Learning

Underwater source localization has good prospects for development and plays a significant role in many fields,such as resource development,industry and military.Matching field processing calculates matching field vector through the phase and amplitude of acoustic field,leading to the technology of underwater source detection with high precision.However,matching field processing is very sensitive to the mismatch between the established model and the true environment,resulting in limitation in some practical applications.As a research highlight and a main developing trend in the field of artificial intelligence in recent years,deep learning can perform underwater source localization without modeling by learning features from the data.Therefore,underwater source localization based on deep learning is a research highlight nowadays.At present,there are many shortcomings in the application of underwater source localization based on deep learning,such as wasting the ability of abstract learning,learning fewer features from the data and the real-complex conversion which may cause the reversal meaning of the data.This thesis proposes a deep complex residual network for underwater source localization.Based on the traditional deep residual network,this network uses components for plural processing.This network can apply the characteristics of underwater acoustic signals more effectively,make use of the ability of abstract learning and learn more features from the data.In addition,this network will use circle loss function to maximize the similarity between data in one class and minimize the similarity between classes,which will make the training more efficient and the convergence state more stable.Circle loss function can improve the separability of the feature space as well.Based on the experimental results,compared with deep residual network,deep complex residual network has better resolution of underwater acoustic signals and robustness to noise.At the same time,circle loss function can improve the resolution of the network in a noisy environment.With the increasing marine economic competitions,the public underwater source localization data is scarce,and the acquisition process of underwater source localization data consumes too much time and money.A large amount of data will make the learning process longer and need more machine as well.In real life,the network need to adapt to the complicated and volatile marine environment quickly.This thesis proposes an active learning framework for underwater source localization based on deep complex residual network.By combining the ideas of active learning and transfer learning with deep complex residual network,this framework solves the above problems.Based on the experimental results,the active learning framework proposed by this thesis can reduce required training data,reduce training time and cost,and can adapt to different environments quickly.At the same time,compared with other active learning methods,the active learning query strategy proposed in this thesis has more significant effect in most iterations.