煤泥水高效处理是煤炭洗选加工的关键环节，为解决现有煤泥水沉降处理过程中药剂添加不准确的问题，提出了一种利用视频分帧处理算法和YOLO算法实现煤泥水澄清层浊度识别的方法：采用分帧处理算法将采集到的浊度为100～1 000 NTU的煤泥水样本视频分解成900张图片，结合labelimg数据标注工具，按照9∶1的比例将图片数据集随机划分为训练集和验证集；随后使用YOLOv5算法的开源YOLOv5m.pt预权重训练模型对训练集进行训练，通过设置交并比、置信度和训练轮数，以损失函数、精确度和召回率为评价指标，并根据检测结果的置信度大小来评估模型目标分类的性能。结果表明：所提出的识别算法在训练过程中损失率持续下降并稳定至0.001，不同浊度下的识别精确度从0.1左右持续升高并稳定在0.98，召回率从0.2左右持续升高并稳定在0.99，最优置信度可高达0.98。该方法对不同浊度范围煤泥水都具有良好的识别效果，具有一定的适应性和泛化能力；同时，相同浊度下的识别精确度相对稳定，具有较好的鲁棒性，这表明该方法具有较高的准确性和稳定性。研究成果可为煤泥水动态检测提供新思路，进而有助于促进煤泥水处理领域的智能化进展。

Efficient treatment of coal slurry water is a key link in coal cleaning and processing process. The problem currently encountered in treatment of coal slurry water is that addition of agents cannot be accurately controlled during the slurry settling process. To tackle the problem, the method for recognition of turbidity of clarified layer of coal slurry water using video frame-splitting processing and YOLO algorithms is proposed. For testing the effectiveness of the proposed method, videos of the slurry water samples with a turbidity in a range of 100~1 000 NTU are splitted into 900 images, the dataset of which are then divided randomly into training and validation sets according to a ratio of 9∶1 using frame-splitting algorithm and labeling data annotation tool. Following that, the open-source YOLOv5m.pt preweighted training model of YOLOv5 algorithm is used to train the training set. Through setting IOU, confidence level and number of training rounds, and with loss function, accuracy and recall rate as evaluation indicators, the model classification performance is assessed based on confidence level of the detected results. As indicated by study results, the loss ratio of the recognition algorithm keeps decreasing at first and then stay stabilized at 0.001 during the training process; the accuracy of recognition for samples at different level of turbidity tends to rise continuously from about 0.1 to 0.98 while the recall rate is seen to go up to continuously to 0.99, with an optimum confidence level being as high as 0.98; the algorithms used demonstrate a remarkably good effect for recognition different turbidities and certain adaptability and generalization capabilities as well; and for detection of the same turbidity, the accuracy is relatively stable with a good robustness, well demonstrating the accuracy and stability of the algorithms. The research-derived achievements provide new ideas for dynamic detection of turbidity of coal slurry water and a valuable reference for making intelligent transformation of coal slurry water treatment system.