【摘要】 针对周界安防中实际采集的振动信号类别存在样本量非均衡分布的问题，提出一种基于DCL-Net与异构AdaBoost集成学习模型的信号识别方法。该方法以深度可分离卷积（DSC）为基础构建特征提取器，引入卷积块注意力机制（CBAM）增强特征判别能力，并结合长短期记忆网络（LSTM）捕捉时序信号的长期依赖关系，构建DCL-Net网络结构。进一步将DSC,DSC-LSTM和DCL-Net三种网络作为弱分类器，改进AdaBoost集成学习算法进行迭代训练与加权投票，形成异构强分类器。为验证该方法的有效性，搭建了基于分布式声波传感的周界安防系统，模拟了攀爬、敲击、步行及无入侵事件。实验结果表明，所提方法在非均衡数据集上识别准确率达到94.00%，相较于同构集成模型、CNN-LSTM及SqueezeNet等对比模型分别提升了5.03%,8.26%,8.46%，验证了其在周界安防非均衡信号识别中的有效性与鲁棒性。更多还原

【Abstract】 A signal-recognition method that combines DCL-Net and a heterogeneous AdaBoost ensemble model is proposed to address the imbalanced sample distribution in vibration signals from perimeter-security applications. A feature extractor based on depth-wise separable convolution(DSC) is constructed. It is augmented with a convolutional block attention module to enhance feature discrimination and integrated with a long short-term memory(LSTM) network to capture long-term dependencies in time-series signals, thus forming the DCL-Net architecture. An improved AdaBoost algorithm iteratively trains three weak classifiers, i. e., DSC, DSC-LSTM, and DCL-Net, by combining them via weighted voting to form a strong heterogeneous classifier. A perimeter-security system based on distributed acoustic sensing is established to simulate climbing, knocking, walking, and no-intrusion events. Experimental results show that the proposed method achieves 94. 00% recognition accuracy on an imbalanced dataset, thereby outperforDming homogeneous ensemble models CNN-LSTM and SqueezeNet by 5. 03%, 8. 26%, and 8. 46%, respectively. Thus, its effectiveness and robustness are validated.更多还原