Abstract: Non-contact respiratory state monitoring based on WiFi Channel State Information (CSI) has shown broad application potential in smart healthcare and home health scenarios due to its advantages of high privacy, low cost, and easy deployment. However, respiratory signals are easily affected by environmental interference during data acquisition, resulting in degraded signal quality. Meanwhile, due to limited acquisition conditions and privacy protection, the number of samples is often insufficient, leading to a small-sample problem, and the generated samples usually lack consistency with real samples in key discriminative features, thereby reducing the effectiveness of data augmentation. In addition, the small differences among different respiratory patterns further increase the difficulty of classification and limit the depth of research on multi-state respiratory recognition. To address these challenges, a SiamGAN-CNNBiLSTM–based respiratory state recognition system was proposed. First, a multi-state respiratory dataset consisting of normal breathing, apnea, sneezing, yawning, and coughing was constructed, and a systematic CSI preprocessing pipeline was designed to improve signal quality. Second, a SiamGAN module integrating a Generative Adversarial Network (GAN) and a Siamese Network was developed. Through the adversarial mechanism of the generator and discriminator, samples with higher realism were produced, while a Siamese encoder was introduced to extract 128-dimensional intrinsic features. A mean squared error constraint was further imposed in the feature space to enhance the consistency between generated and real samples in key discriminative characteristics. This strategy strengthened the discriminability and generalization of generated samples, preserved the physical properties of the original signals, and avoided the introduction of task-irrelevant high-frequency noise. Finally, a deep classification network combining a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (BiLSTM) network was constructed to fully exploit the spatiotemporal features of respiratory signals and achieve high-accuracy recognition of multiple respiratory states. Experimental results demonstrated the effectiveness of the proposed SiamGAN-CNNBiLSTM method in an office scenario: compared with the CNN-BiLSTM method, the accuracy increased by 2.2 percentage points, precision by 2.3 percentage points, recall by 2.3 percentage points, and the F1-score by 2.3 percentage points; compared with the Transformer method, the accuracy increased by 8.3 percentage points, precision by 8.2 percentage points, recall by 8.3 percentage points, and the F1-score by 8.3 percentage points. In a corridor scenario, the method achieved an accuracy of 96.7%, while precision, recall, and F1-score all reached 96.8%, only slightly lower than those in the office scenario, further indicating the robustness and stability of the method across different environments.

Key words: WiFi channel state information (CSI), data augmentation, convolutional neural network–bidirectional long short-term memory (CNN-BiLSTM), generative adversarial network (GAN), siamese network, respiratory state recognition

摘要： 基于WiFi信道状态信息（CSI）的非接触式呼吸状态监测因其高隐私性、低成本及易部署等优势，在智慧医疗与居家健康领域展现出广泛应用前景。然而，呼吸状态数据在采集过程中易受环境干扰，导致信号质量下降；同时，受限于采集条件与隐私保护，样本数量有限，常存在小样本问题，且生成样本往往缺乏与真实样本在关键判别特征上的一致性，降低了增强数据的有效性；此外，不同呼吸模式间信号差异较小，进一步增加了分类识别难度，限制了多状态呼吸识别的研究深度。针对上述挑战，本文提出一种SiamGAN-CNNBiLSTM的呼吸状态识别系统。首先，构建包含正常呼吸、呼吸暂停、打喷嚏、打哈欠与咳嗽多状态呼吸数据集，并设计系统化的CSI信号预处理流程以提升信号质量；其次，提出融合生成对抗网络（GAN）与孪生网络（Siamese Network）的SiamGAN模块，通过生成器与判别器的对抗机制生成具有更高真实性的样本；同时，引入孪生编码器提取信号的128维本质特征，并在特征空间施加均方误差约束，以强化生成样本与真实样本在关键判别特征上的一致性。该策略既增强了生成样本的判别性与泛化性，也有效保留了原始信号的物理属性，避免引入分类无关的高频噪声；最后，设计融合卷积神经网络（CNN）与双向长短期记忆网络（BiLSTM）的深度分类网络，充分挖掘呼吸信号的时空特征，实现多种呼吸状态的高精度识别。实验结果表明，所提SiamGAN-CNNBiLSTM方法在办公室场景中表现优异：与CNN-BiLSTM方法相比，准确率提升2.2个百分点，精确率提升2.3个百分点，召回率提升2.3个百分点，F1分数 提升2.3个百分点；与Transformer方法相比，准确率提升8.3个百分点，精确率提升8.2个百分点，召回率提升8.3个百分点，F1分数提升8.3个百分点。在走廊场景中，本方法的准确率达到96.7%，精确率、召回率和F1分数均为96.8%，仅较办公室场景略有下降，进一步表明了方法在不同环境下的鲁棒性与稳定性。

关键词: WiFi CSI, 数据增强, CNN-BiLSTM, 生成对抗网络, 孪生网络, 呼吸状态识别