针对综采工作面液压支架护帮板处于未收回异常状态导致采煤机滚筒与护帮板干涉问题，提出一种改进YOLOv5s的采煤机滚筒与液压支架护帮板干涉状态智能识别方法。运用课题组前期提出的基于边界约束和非线性上下文正则化的去雾去尘方法对视频图像进行清晰化处理，提高综采工作面监控视频图像质量；对YOLOv5s模型进行改进，通过将YOLOv5s主干网络中的普通卷积Conv替换为分类效果更佳的Ghost卷积，减少了模型的参数数量，提高了模型识别速度，同时引入坐标注意力机制，提高了模型对护帮板和滚筒特征提取能力，从而提高模型识别精确率。运用软非极大值抑制算法（Soft-NMS）的锚框筛选方法，减少因护帮板重叠而发生漏检问题。针对采煤机滚筒与液压支架护帮板干涉状态判定问题，提出液压支架护帮板与采煤机滚筒锚框重合度的判定方法。运用本文改进YOLOv5s模型与YOLOv5s、YOLOv3-tiny模型进行对比分析，结果表明：本文方法与原模型相比的识别精确率提高了约8.1%，GFLOPs降低1.86倍；mAP@.5达到97.2%、平均识别速度为检测时间为5.9 ms。运用本文方法对煤矿实际综采工作面采煤机滚筒与液压支架护帮板视频图像进行干涉状态识别试验验证，结果表明：对采煤机滚筒与液压支架护帮板干涉状态识别准确率为96%。

For the problem of interference between shearer drum and hydraulic support guard plate, an interference state intelligent recognition method for shearer drum and hydraulic support guard plate of improved YOLOv5s algorithm is proposed. The use of boundary constraint and non-linear contextual regularization based on the group's previous proposed method of defogging and dust removal to clarify the video image, improve the quality of the monitoring video image of the comprehensive mining face. The YOLOv5s model is improved by replacing the ordinary convolutional Conv in the YOLOv5s backbone network with Ghost convolution, the improved algorithm greatly reduces the number of model parameters and improves the model recognition speed. At the same time, the coordinate attention mechanism is introduced to improve the model's ability to extract the features from the guard plate and shearer, and improve the recognition accuracy. The soft non-maximum suppression algorithm (Soft-NMS) anchor frame screening method is used to reduce the problem of missed detection due to overlapping guard plates. For the problem of determining the interference state of shearer drum and hydraulic support guard plate, the method for determining anchor box overlap degree between hydraulic support guard plate and shearer drum. The improved YOLOv5s algorithm is compared with YOLOv5s and YOLOv3-tiny algorithm. The results indicate that compared with the original YOLOv5s model, the recognition accuracy of this method has been improved by about 8.1%, and GFLOPs have been reduced by 1.86 times. mAP@.5 was increased to 97.2%, and the average recognition speed is 169 frames/s. The improved YOLOv5s algorithm is used to validate the interference state recognition effect for video images of shearer drum and hydraulic support guard plate in in actual fully mechanized mining faces, and the results show that the recognition accuracy of interference state between the coal shearer drum and the hydraulic support guard plate is 96%.