To enhance the recognition performance of short-sample click-like bionic hydroacoustic communication signals in complex underwater environments, a recognition method based on fused multimodal contrastive learning is proposed. Firstly, by incorporating time-frequency and fused feature modalities, discriminative features are directly extracted from fine-grained signal structures, reducing reliance on long-sequence statistical stability. Secondly, the samples of the same class are pulled together and those of different classes are pushed apart to the model to focus on robust features that are insensitive to multipath interference using the contrestive learning mechanism. Finally, the introduction of a text modality provides semantic constraints for feature learning, guiding the model to capture category-related essential attributes. Simulation and lake trial results demonstrate that the proposed method significantly outperforms existing algorithms based on time-delay statistics and statistical-convolutional neural networks under short-sample and multipath hydroacoustic channel conditions. Specifically, the lake trial achieves a recognition accuracy of 88.70% with a short sample length of 0.7 s.
本文采用含有类别信息的描述性文本作为文本模态的输入,基本格式如下:“The sound belongs to {BCCCT/RCT}”。对描述性文本进行字节对编码后,输入文本编码器获得文本模态嵌入向量。文本编码器采用基于Transformer架构的模型[13],其核心包含12层编码器层、8个注意力头以及512维的隐藏层。
JIANGJ J, QIAOF, LIY, et al. Recognition method for the bionic camouflage click communication trains modulated by time delay difference[J]. The Journal of the Acoustical Society of America, 2022,152(1):No.491.
[5]
桂阳.非合作仿生水声通信信号检测及识别方法研究[D].南京:东南大学,2023:1-89.
[6]
YAOQ W, JIANGJ J, YUX L, et al. Recognition method for biomimetic camouflage communication signal imitating cetacean click in underwater multipath channels[J]. Applied Acoustics, 2024,222:No.110039.
[7]
GUZHOVA, RAUEF, HEESJ, et al. Audioclip: extending clip to image, text and audio[C]//Proceedings of the 2022 IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway, USA: IEEE, 2022:976-980.
[8]
JIANGJ J, LIS Y, WANGX Q, et al. Study of the relationship between sound signals and behaviors of a sperm whale during the hunting process[J]. Applied Acoustics, 2021,174:No.107745.
[9]
WATKINSW A. Watkins marine mammal sound database[EB/OL]. (2020-03-09)[2025-10-16].
[10]
WATKINSW A, FRISTRUPK, DAHERM A, et al. Sound database of marine animal vocalizations structure and operations[J].Woods Hole Oceanographic Institution, 1992. DOI:10.1575/1912/854 .
LIUS, QIAOG, ISMAILA. Covert underwater acoustic communication using dolphin sounds[J]. The Journal of the Acoustical Society of America, 2013,133(4):300-306.
[13]
RADFORDA, KIMJ W, HALLACYC,et al. Learning transferable visual models from natural language supervision[DB/OL]. (2021-02-26)[2025-10-16].
[14]
KEDEMB. Spectral analysis and discrimination by zero-crossings[J]. Proceedings of the IEEE, 1986,74(11):1477-1493.
[15]
FLORES-FUENTESW, RIVAS-LOPEZM, SERGIYENKOO, et al. Combined application of power spectrum centroid and support vector machines for measurement improvement in optical scanning systems[J]. Signal Processing, 2014,98:37-51.
[16]
GIANNAKOPOULOST, PIKRAKISA. Introduction to audio analysis: a MATLAB approach[M]. Amsterdam,Netherlands: Elsevier, 2014:233-240.
[17]
KIMH G, MOREAUN, SIKORAT. MPEG‐7 audio and beyond: audio content indexing and retrieval[M]. New York, USA: John Wiley & Sons, 2005:1-285.
[18]
XIEJ, ZHUM Y. Investigation of acoustic and visual features for acoustic scene classification[J]. Expert Systems with Applications, 2019,126:20-29.
[19]
张健.基于深度学习的水下目标识别的研究[D].成都:电子科技大学,2020:1-80.
[20]
PEREIRAL M, SALAZARA, VERGARAL. A comparative analysis of early and late fusion for the multimodal two-class problem[J]. IEEE Access, 2023,11:84283-84300.
[21]
GADZICKIK, KHAMSEHASHARIR, ZETZSCHEC. Early vs late fusion in multimodal convolutional neural networks[C]//Proceedings of the 2020 IEEE 23rd International Conference on Information Fusion. Piscataway, USA: IEEE, 2020. DOI: 10.23919/FUSIDN45008.2020.9190246 .
[22]
GUZHOVA, RAUEF, HEESJ, et al. ESResNet: environmental sound classification based on visual domain models[C]//Proceedings of the 2020 25th International Conference on Pattern Recognition. Piscataway, USA: IEEE, 2021:4933-4940.