呼吸机波形图像可以反映机械通气时的呼吸模式。现有的呼吸回路监测和管理技术存在实时监测困难，过度依赖医护人员经验等局限性。为了实现机械通气时实时呼吸模式监测和提高异常呼吸模式的识别准确性，通过引入实时图像采集设备和基于深度学习的图像识别技术，将临床中采集到的真实图像通过基于NanoDet-Plus模型的感兴趣区域提取、图像预处理和基于LAMP(Layer-adaptive Sparsity For The Magnitude-Based Pruning)重要性评估策略和MagnitudePruner剪枝策略剪枝、CWD(Channel-wise Distillation)和BCKD(Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection)蒸馏后的YOLOv8n-Prune模型进行波形图像识别，算法实现了对临床环境中常见异常情况，包括积水、漏气、分泌物累积、人机不同步等多种异常呼吸模式的[1]检测，模型在规模得到压缩的情况下，mAP50达到0.990，mAP50:95达到0.879，通过类激活映射实验，证明了算法具有一定的可解释性。结合实时图像采集设备，算法可做到对异常呼吸模式的实时监测，在临床实践中具有良好的应用价值。

Ventilator waveform images can reflect the breathing pattern during mechanical ventilation. Existing respiratory circuit monitoring and management techniques have limitations such as difficulty in real-time monitoring and over-reliance on healthcare personnel experience. In order to realize real-time respiratory pattern monitoring during mechanical ventilation and improve the recognition accuracy of abnormal respiratory patterns, by introducing real-time image acquisition equipment and deep learning-based image recognition technology, real images acquired in the clinic are processed through region of interest extraction based on the NanoDet-Plus model, image preprocessing, and LAMP (Layer-adaptive Sparsity For The Magnitude-Based Pruning) strategy and MagnitudePruner pruning strategy pruning, CWD (Channel-wise Distillation) and BCKD (Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection). for Distillation in Dense Object Detection) distillation of the YOLOv8n-Prune model for waveform image recognition, the algorithm achieves the detection of a variety of abnormal respiratory patterns in the common abnormalities in clinical environments, including water retention, air leakage, secretion accumulation, and patient-ventilator asynchrony, etc., and the model, in the case of the scale getting compressed, mAP50 reaches 0.990 and mAP50:95 reaches 0.879. The algorithm is shown to be interpretable through class activation mapping experiments. Combined with real-time image acquisition equipment, the algorithm can achieve real-time monitoring of abnormal breathing patterns, which has good application value in clinical practice.