基于改进残差网络的罗氏沼虾发声信号分类方法

Acoustic signal classification methods of Macrobrachium rosenbergii based on improved residual network

摘要:

目的 水产养殖中虾类行为的精准识别对饲料投喂优化和疾病预防管理具有重要意义。针对传统光学监测方法在复杂养殖环境中的局限性，从被动声学监测角度，针对虾类在复杂养殖环境中传统光学监测方法的局限性。方法 本研究结合被动声学技术获取罗氏沼虾的不同行为发声信息，提出了一种基于深度学习的罗氏沼虾行为分类方法。通过采集摄食、移动及打斗三种行为的发声信号，将其转换为Mel频谱图作为数据集，并比较了CNN、ResNet18和VGG16神经网络模型分类效果。结果 ResNet18的识别准确率 (97.67%)优于VGG16和CNN；在引入批量归一化 (Batch Normalization, BN)算法后，BN-ResNet18的识别准确率提升至99.00%，较原始ResNet18提高了1.33%。此外，BN-ResNet18在14.0~44.1 kHz频段内表现出最优的分类性能，进一步证明了残差连接与BN模块的协同优化能够有效提升模型性能。结论 BN-ResNet18在复杂行为发声信号特征建模分类中展现出较高的准确性和稳健性。本研究为基于虾类行为发声信号的监测识别提供了技术支持，对水产养殖的智能化研发具有潜在应用价值。

Abstract:

The precise identification of shrimp behavior in aquaculture is of great significance for optimizing feeding and disease prevention. In view of the limitations of traditional optical monitoring methods in complex aquaculture environments, with integrated passive acoustic technology, this research acquires the acoustic information associated with different behaviors of the Macrobrachium rosenbergii and proposes a deep learning-based method for behavior recognition in M. rosenbergii. The acoustic signals of three behaviors (i.e., feeding, moving, and fighting) were collected and converted into Mel spectrograms as the dataset. Then the classification effects of CNN, ResNet18, and VGG16 neural network models were compared. The results showed that ResNet18 in terms of recognition accuracy (97.67%) outperforms VGG16 and CNN. After introducing the Batch Normalization (BN) algorithm, the recognition accuracy of BN-ResNet18 increased to 99.00%, representing a 1.33% enhancement relative to the baseline ResNet18 model. In addition, BN-ResNet18 showed the best classification performance in the 14.0-44.1 kHz frequency band, which further proved that the synergistic optimization of residual connection and BN module could effectively enhance model performance. BN-ResNet18 demonstrates high accuracy and robustness in feature classification of complex behavioral acoustic signals. This study provides technical support for intelligent recognition based on the acoustic signals of shrimp behaviors and has potential application value in the refined management of aquaculture.