摘要
对于被动声呐接收到的水声信号,将信号的时域波形转化为时频谱图和梅尔谱图后,可采用神经网络和集成学习的方法,将信号识别转换为图像识别问题。利用多种卷积神经网络对信号谱图进行训练学习,并通过堆叠法(Stacking)将单网络结构作为初级学习器构建多网络集成模型,可进一步提高目标识别准确率。利用DeepShips数据集进行目标识别仿真验证,结果表明,多网络集成模型在四分类数据集上的识别准确率可达100%,能够有效提高被动声呐的目标识别能力,对水下目标智能探测和识别具有参考价值。
For underwater acoustic signals received by passive sonar, the signal recognition can be converted into an image recognition problem by using neural networks and ensemble learning methods after converting the time domain waveform of the signal into a time-frequency spectrum and a mel-spectrogram. This paper uses a variety of convolutional neural networks to train and learn the signal spectrogram, and uses the Stacking method to construct a multi-network integration model using a single network structure as a primary learner, which can further improve the target recognition accuracy. The DeepShips dataset is used for target recognition simulation verification. The results show that the recognition accuracy of the multi-network integration model on the four-classification dataset can reach 100%, which can effectively improve the passive sonar target recognition capability and has reference value for intelligent detection and recognition of underwater targets.
作者
汤航
樊书宏
TANG Hang;FAN Shuhong(The 705 Research Institute of CSSC,Xi'an 710077,China)
出处
《舰船科学技术》
北大核心
2025年第12期111-116,共6页
Ship Science and Technology
关键词
水下目标识别
梅尔语谱图
集成学习
underwater target recognition
mel-spectrogram
ensemble learning
