Abstract
Machine vision was utilized in this study to accurately classify the low concentration slurry. Orthogonal experiment L(9)(3(4)) indicated that the optimal coal slurry collection images were achieved with exposure value of 10, slurry layer thickness of 7 cm, and light intensity of 5 × 10(4) lux. Subsequently, a new low concentration classification model was systematically developed, encompassing aspects such as original image acquisition, data augmentation, dataset partitioning, classification algorithm design, and model evaluation. DCGAN was employed for image generation, achieving favorable outcomes with generator learning rate set at 5 × 10(- 5), discriminator at 1 × 10(- 6), and iteration number at 2000. At the point, the maximum SSIM similarity reached 0.9381, and the pHash similarity was 0.9375. Results from subsequent CNN model training, with 200 iterations, the accuracy on training and validation sets was demonstrated over 95% for coal slurry concentration prediction. Further evaluation using recall, precision, and F1-score revealed CNN network model metrics: maximum recall 1.000, minimum 0.800; maximum precision 1.000, minimum 0.700; and highest F1 score 1.000, lowest 0.778. Additionally, the accuracy of this model on the test set reached as high as 94%. The findings indicated the excellent performance in low concentration detection of coal slurry throughout this study.