放煤量的智能监测技术是智能化综放开采发展的方向之一。针对厚煤层综放开采混合介质下落过程煤矸智能识别现有方法存在适用范围窄、误判率高等问题，提出一种基于实时区域卷积神经网络(Faster R-CNN)改进的CBAM Faster R-CNN煤矸混合放出状态分析识别方法。该方法以综放开采工作面煤矸石下落状态为背景，将注意力机制算法(CBAM)融入ResNet50特征提取网络，研究了注意力机制融合于ResNet50主干特征提取网络的融合优化方法，确定了以煤矸石下落状态检测为目标的最佳融合位置，增加了提取煤与矸石权重信息的能力，使得特征提取过程重点关注煤矸运动状态参量；构建了粉尘环境下综放开采煤矸混合放出状态试验平台，对高速摄像机所采集的煤矸原始运动图像进行暗通道去雾与模糊集增强预处理操作，分析识别不同工况下静态图像，定量研究了模型权重优化以及粉尘环境预处理对煤矸混合放出状态识别的影响。研究结果表明：基于CBAM Faster R-CNN模型的预处理图像识别精确率较原始图像提升了8.84%，召回率提升了6.68%；模型优化效果方面，CBAM Faster R-CNN模型平均查准率和召回率分别为82.63%，86.53%，高于经典模型的74.69%，79.93%；F1-score值较经典模型提升了7%。因此，基于“预处理+CBAM Faster R-CNN”的图像识别方法为实现粉尘环境下综放开采放煤量的精准辨识提供了可靠的理论支撑。

Intelligent monitoring of coal release is one of crucial directions for an intelligent fully mechanized caving ina top⁃coal caving mining operation. The current methods have some problems including a narrow scope ofapplication, high error rate, and hardly fulfillment of the intelligent identification for the mixed media of thick coalseam in the caving process. An improved CBAM Faster R-CNN coal⁃gangue mixed release state identification methodbased on real⁃time regional convolutional neural network (Faster R-CNN) was presented in this paper. The methodintegrates the attention mechanism algorithm (CBAM) into the ResNet50 feature extraction network in the context of the coal⁃gangue falling state at a comprehensive top⁃coal caving face, investigates the fusion optimization methodof the attention mechanism fusion in the ResNet50 backbone feature extraction network, determines the best fusion po⁃sition with the coal⁃gangue falling state detection as the target, increases the ability to extract the weight informationof coal⁃gangue, and makes the feature extraction process focus on the coal⁃gangue movement state parameters. A no⁃vel testing platform of the coal⁃gangue masses was constructed considering in⁃situ dusty environment. All initial imagesbeing captured by high⁃speed camera were totally pre⁃processed with dark channel defogging and fuzzy set enhance⁃ment. The model weighting optimization was quantitatively studied with the analysis of static images under differentworking conditions. Also, the identification influence on the mixed and release state adopting both the model weightingoptimization and dust setting preprocess with the analysis of vast static images under different working conditions wasquantitatively revealed. The relevant results show that the accuracy of preprocess image recognition based on CBAMFaster R-CNN is improved by 8.84% and the recall rate is improved by 6.68% being compared with the initial ima⁃ges. Besides, the average check rate and recall rate of CBAM Faster R -CNN are 82.63% and 86.53% beinghigher than classical model with respectively 74.69% and 79.93% in terms of the model optimization effect. F1⁃scorevalue is also improved by 7%. Therefore, the image recognition with hybrid preprocessing and CBAM Faster R-CNNprovides a theoretical support for the accurate identification of the coal⁃gangue masses in a fully mechanizedtop⁃coal caving operation considering in⁃situ dusty setting.