Research On Underwater Terrain Aided Navigation Technology Based On Deep Learning

The prerequisite for underwater vehicles to accomplish various underwater tasks is to use internal or external sensors to obtain their real-time position information in the underwater environment,which is also known as underwater localization and navigation technology.Underwater localization and navigation is a technology that provides the attitude,speed,position and other information of the underwater vehicle,which is the prerequisite for the underwater vehicle to be able to successfully complete various underwater tasks,and is also the focus and difficulty of its research.In short distance voyage,inertial navigation system can provide more accurate localization information for underwater vehicles,but with the increase of voyage distance and the accumulation of time,the drift error generated by inertial navigation will gradually increase,thus deviating from the original target position.For underwater vehicles that need to accomplish long range and long distance,it is necessary to correct the drift error generated by inertial navigation through absolute localization,so as to obtain more accurate position information.Due to the serious attenuation of electromagnetic waves underwater,GNSS cannot be used for absolute localization of underwater submersibles.At the same time,the complicated and changeable underwater environment has posed a great challenge to the underwater localization and navigation method,which is mainly based on inertial navigation,and a single localization and navigation technology will inevitably be unable to meet the future needs of long range,multi-targeting and high-precision.The future navigation system of underwater intelligent vehicle will inevitably be a combination of navigation methods integrating a variety of technologies.Among them,the terrain-aided navigation technology combining the basic inertial navigation technology and terrain matching technology can make the underwater robot well realize the purpose of highprecision,strong concealment,and all-ocean navigation in the passive state.Aiming at the current problems faced in underwater localization and navigation,this paper proposes an underwater terrain-assisted navigation method based on deep learning,which constructs a deep learning model for extracting global style and local detail terrain features and a self-distillation contrast learning framework for model training,and realizes high-precision terrain matching at the level of abstract features.Using the extracted terrain abstract features from the same deep learning model,highly reliable fitness regions were selected from the pre-acquired terrain maps,and the fusion processing of underwater terrain matching and underwater terrain fitness analysis was realized.The main research content and innovative results of the thesis are as follows:(1)A deep learning model for extracting terrain features is investigated.Different from the traditional manual extraction of underwater terrain features,a data-driven approach is adopted,and the ideas of self-attention mechanism and convolutional neural network are combined and introduced into terrain matching,and a global style and local detail feature extraction network is proposed.Automatic extraction of local and global terrain features is realized,and abstract local detail and global style feature representations are obtained,which improves the accuracy of underwater terrain matching.Aiming at the problems of the difficulty in labeling terrain data samples and the insufficient number of terrain samples used for training,a self-distillation contrast learning model is constructed by combining the ideas of knowledge distillation and contrast learning,which achieves end-to-end learning of underwater terrain features without the use of sample labels and negative samples.(2)A deep learning-based underwater terrain matching method is investigated,which realizes high-accuracy terrain matching in the dimension of abstract features by comparing the abstract representations between the real-time terrain data and the terrain data to be matched,rather than the terrain elevation values themselves.In order to resist measurement errors and rotational variations present in underwater terrain measurements,enhancement methods for underwater terrain data are proposed,including randomly adding zero-mean Gaussian noise and randomly flipping the terrain data up and down and left and right.(3)Meanwhile,a nonlinear supervised terrain suitability classification method based on deep learning is proposed,which uses a nonlinear network to fit a terrain suitability classifier and realizes automatic classification discrimination of terrain area suitability performance.A fusion processing method is proposed to realize terrain matching and terrain suitability analysis in the dimension of abstract features,and the feature extraction network is used to abstractly characterize terrain features,which not only can realize highprecision terrain matching,but also can use the matching results directly for terrain suitability analysis.In this paper,the effectiveness of the proposed method is verified by relevant ablation experiments and performance tests.A large number of tests have been conducted using different types of datasets and under different experimental conditions,and the results show that the proposed method has good robustness and generalization ability.From the results of the comparative and simulation experiments,it can be seen that the proposed method can realize high-precision terrain matching and high-reliability terrain fitness analysis.The results of the study can provide useful reference and guidance for related tasks such as terrain matching and terrain suitability analysis.