Abstract: To mitigate the thickness effect of the Dual Energy X-Ray Transmission (DE-XRT) system and achieve accurate identification and density prediction of coal and gangue within a wide thickness range, this study thoroughly summarizes the imaging principle of DE-XRT and fuses depth cameras with DE-XRT at the data level. Based on the size of coal gangue and regional similarity, the DE-XRT images and thickness images were partitioned. By expanding single pixels into multi-pixel sets, the two types of information were utilized more accurately and efficiently. Furthermore, using the X-ray and thickness information of each partition as the basis, fused information is calculated via the X-ray energy attenuation formula, and then aggregated into block matrices. Multidimensional heterogeneous features were extracted from these block matrices and the DE-XRT images. A robust pre-identification model was then constructed by combining the Relief feature selection algorithm with a Support Vector Machine (SVM) classifier optimized by a GA (Genetic Algorithm). Additionally, via transformation calculations of block matrices and statistical analysis of large-scale samples, the fuzzy intervals of coal at various density levels were determined, and a density prediction model based on these intervals was developed. This model considers the impact of using a single formula to calculate coals of different density grades. By establishing fuzzy intervals, the problem of coal density prediction is decomposed into multiple density-grade prediction problems. Meanwhile, the degree of deviation between the target to be predicted and the interval lower limit is used to characterize its proximity to the lower-limit density grade. Experimental results demonstrated that, on a dataset with coal densities ranging from 1.3 to 1.8 g/cm³, gangue densities greater than 1.8 g/cm³, and average thicknesses spanning 5 to 100 mm, the pre-identification model achieved an Pre of 97.522% and an F₁ of 0.962. Compared with existing X-ray algorithms, grayscale texture methods, and deep learning algorithms, the Pre and F₁ were improved by at least 6.433% and 2.888%, respectively. The density prediction model exhibited a mean error not exceeding 5.882%. Specifically, 46.993 of targets had prediction errors below 4%, and 93.233% had errors below 10%. The proposed models effectively reduce the impact of the thickness effect on the sorting system, enhance the accuracy of coal-gangue identification and density prediction across a broad thickness range, and provide a theoretical foundation for developing intelligent photoelectric coal-gangue sorting technologies.