摘要
水声目标识别是近年来各国的研发热点,但是由于水声目标难以采集而导致样本数据不足,严重影响了神经网络的识别效率以及自动化识别装备的水平和性能的发挥;为此,提出了一种基于样本扩充网络的水声目标分类模型优化方法,通过搭建掩模重建的样本扩充网络,充分利用无标注数据进行训练,使模型学习到样本的全局高维特征,再生成样本加入后续的识别模型训练中,在两次试验过程中,平均识别准确率从76%提升至80%,最佳识别准确率从88%提升至96%;基于实测数据的实验表明,该方法提升了分类器的准确率、收敛速度以及稳定性。
Underwater acoustic target recognition is a research and development hot spot in many countries in recent years.However,it is difficult to collect underwater acoustic targets,resulting in the sample data insufficient,which seriously affects the recognition efficiency of neural network and the level and performance of automatic recognition equipment.Therefore,an optimization method of underwater acoustic target classification model based on the sample expansion network is proposed.Through building the sample expansion network reconstructed by the mask,the unlabeled data is fully utilized to train the model,learn the global high-dimensional features of the samples,and then generate the samples to be added to the subsequent recognition model training.Based on the results of two experiments,the average accuracy of target classification model improves from 76%to 80%,with its maximum accuracy improving from 88%to 96%.Experimental results show that this method improves the accuracy,convergence speed and stability of the classifier.
作者
张博轩
赵天白
常振兴
蒋翔宇
王少博
ZHANG Boxuan;ZHAO Tianbai;CHANG Zhenxing;JIANG Xiangyu;WANG Shaobo(The 54 th Research Institute,China Electronics Technology Group Corporation,Shijiazhuang 050081,China;State Grid Power Space Technology Co.,Ltd.,Beijing 102213,China;College of Communication and Information Engineering,University of Electronic Science and Technology of China,Chengdu 610731,China)
出处
《计算机测量与控制》
2024年第4期143-150,共8页
Computer Measurement &Control
基金
国家自然科学基金项目(U20B2071)。
关键词
水声目标识别
样本扩充网络
循环对抗生成网络
掩码训练
梅尔倒谱系数
underwater acoustic target recognition
sample expansion network
cycle generative adversarial networks
mask training
mel-scale frequency cepstral coefficients
