Research On Few-Shot Lightweight Underwater Acoustic Target Recognition Technology Based On Deep Neural Network Model

Underwater acoustic target recognition technology is an important technical means to build the strategy of “maritime power” and realize marine security.Studying underwater acoustic target recognition technology deeply has important theoretical and applied value for national economy and national defense construction.In this background,aiming at the problems and challenges faced in the practical application of underwater acoustic target recognition,such as few open data sources,difficulty in constructing data sets,inapplicability of conventional deep neural network models,difficulty in deploying and running models due to limited equipment space,resources and power consumption and lack of a large number of labeled data to train the model,this dissertation conducts the research on few-shot lightweight underwater acoustic target recognition technology based on deep neural network model.Based on underwater acoustic target recognition data collection and dataset construction,according to the idea of "model design and construction"-"effect improvement and optimization"-"specific scene breakthrough",underwater acoustic target recognition technology based on deep neural network model,underwater acoustic target recognition technology based on joint deep neural network model and underwater acoustic target recognition technology under few-shot condition are explored and studied.The main work and achievements of this dissertation include:1.The dataset of underwater acoustic target recognition is constructed to provide necessary data support for subsequent technical research.For the problem of limited open data source in the area of underwater acoustic target recognition and the difficulty in dataset construction,this dissertation carries out the basic work of underwater acoustic target recognition data collection and data set construction.The situation of open acoustic data source is investigated and analyzed.For underwater acoustic target recognition,the ONC dataset is constructed which covers 4 categories and a total of 300000 samples over a period of 17 months.In addition,the factors affecting model recognition performance in the dataset are qualitatively analyzed and compared,which provides ideas and directions for the improvement and optimization of the dataset.2.A Multiscale Residual Deep Neural Network(MSRDN)and its Lightweight MSRDN are proposed for underwater acoustic target recognition to provide basic model support for subsequent technical research.Aiming at the problems such as the inapplicability of conventional deep neural network in underwater acoustic target recognition and the difficulty of deploying and running the model when space,resources and energy consumption are limited,a Multiscale Residual Unit(MSRU)is designed,and MSRDN is proposed based on it.On ONC dataset,the recognition accuracy of MSRDN reaches 83.15%,which is better than other comparison methods.Comparing the training process and results of two generative adversarial networks modified by MSRU,the effectiveness and characteristics of MSRU are demonstrated.In addition,based on the empirical results of network deep search and the lightweight design idea of deep neural network,Lightweight MSRDN is proposed.When the accuracy is reduced by only 0.02%,the number of parameters is reduced by 42.64M(78.41%)and the floating point operations is reduced by 4.999G(78.16%).The efficiency of the model has been significantly improved.3.A Naive Joint Model based on naive association and an optimized Joint Model based on synchronous deep mutual learning are proposed for underwater acoustic target recognition,which effectively combine multi-modal information to further improve the recognition effect of the model and provide efficient model improvement for subsequent technical research.In order to further improve the effectiveness of underwater acoustic target recognition,a Naive Joint Model is proposed based on Lightweight MSRDN and other lightweight models based on the analysis and demonstration of the existence of complementary space and the availability of the combined model.The Joint Model is proposed by further utilizing synchronous learning mechanism to improve the joint learning method,which can realize efficient mutual learning among branch models while saving about 11.54%-16.27% model training time.On the ONC dataset and Deep Ship dataset,the recognition accuracy of Joint Model reaches 85.20% and 79.50%,respectively,which is superior to other comparison methods.The results of ablation experiments show that the improved effect of the joint model is not a unique accidental phenomenon of a particular model combination,and this joint learning mode can be extended to different model combinations to effectively improve the underwater acoustic target recognition effect.4.A learning framework for underwater acoustic target recognition model with few samples and a semi-supervised fine-tuning training method based on similarity of depth features are proposed.Combined with necessary data support and efficient model improvement,the problem of underwater acoustic target recognition under small sample conditions is broken.To solve the problem of lack of labeled data in underwater acoustic target recognition,a four-stage model learning framework for underwater acoustic target recognition is proposed,and a semi-supervised fine-tuning training method based on depth feature similarity is proposed for the semi-supervised fine-tuning training stage in the learning framework.A set of small sample datasets with different amounts of labeled data are constructed,and the performance baselines of four underwater acoustic target recognition models are established based on these datasets.Compared with the baseline,using the proposed framework effectively improves the recognition effect of the four models.Especially for the Joint Model,the recognition accuracy has increased by 2.04%to 12.14% compared with the baseline.Only 10 percent of the labeled data is needed to exceed the model performance on the full dataset,effectively reducing the dependence of model on the number of labeled samples.The problem of lack of labeled samples in underwater acoustic target recognition is alleviated.