受采场地质条件变化、泵站压力波动及自动跟机系统误差等因素影响，液压支架在自动跟机移架过程中存在丢架情况，人工丢架监测严重影响工作面自动跟机效率。而基于传感器和感知信息的液压支架实时丢架状态监测方法的稳定性和可靠性较差。针对上述问题，提出一种液压支架丢架状态视觉自动检测方法。首先采用YOLOv8对实时获取的工作面监控视频图像进行工作面目标区域划分，通过充分学习工作面图像内部特征准确获取液压支架底座及推杆的轮廓信息与位置信息，分析不同液压支架底座及推杆的位置信息，确定监控视频图像中的支架号；然后提取相邻液压支架最小底座区域局部图像，利用融合多尺度特征信息的ResNet50卷积网络对底座局部图像进行特征提取，获取图像多尺度融合特征信息，再将特征信息映射到类别空间，获取不同液压支架状态的概率分布，根据概率判断液压支架正常移架或丢架状态，结合支架号信息确定处于丢架状态的液压支架。实验结果表明：基于监控视频的工作面目标区域平均分割精度为0.98，准确实现目标区域结构化提取；支架号自动识别准确率为98.78%，为液压支架丢架状态检测提供准确的支架号信息；工作面液压支架丢架状态视觉自动检测的平均准确率达99.17%，单帧图像处理时间为36 ms，满足采煤工作面AI视频监控系统检测丢架状态的实时性与可靠性需求。

Due to changes in mining field geological conditions, fluctuations in pump station pressure, and errors in automated following system, hydraulic supports may experience a loss state during the automatic frame movement process. Manual monitoring of support loss significantly affects the efficiency of automated following at the working face. However, real-time loss state monitoring methods for hydraulic supports based on sensors and perception information often lack stability and reliability. To address these issues, a visual automatic detection method for hydraulic support loss state was proposed. First, YOLOv8 was used to segment key areas of the working face from real-time monitoring video images. By thoroughly analyzing the internal features of the working face images, the contours and location information of hydraulic support bases and push rods were accurately obtained. The location information of different hydraulic support bases and push rods was analyzed to determine the number of each hydraulic support in the monitoring video. Then, the local image of the smallest base region between adjacent hydraulic supports was extracted. A ResNet50 convolutional network, incorporating multi-scale feature fusion, was employed to extract features from the local image of the base, obtaining multi-scale fusion feature information of the image. This feature information was then mapped to a classification space to derive the probability distribution of different hydraulic support states. The support's normal movement or loss state was determined based on these probabilities, and the support number information was combined to identify the hydraulic support experiencing a loss state. Experimental results showed that the average segmentation accuracy of key areas at the working face using monitoring video was 0.98, achieving structured extraction of target regions. The automatic identification accuracy for support numbers was 98.78%, providing precise support number information for hydraulic support loss state detection. The average detection accuracy of the visual automatic detection method for hydraulic support loss state at the working face was 99.17%, and the single frame image processing time is 36 ms, meeting the real-time and reliability requirements for detecting support loss states in coal mining face AI video monitoring systems.