An Explainable Deep Learning Framework for Morphology-Aware Coal Surface Characterization and Intelligent Coal Processing

Coal surface morphology plays a pivotal role in advanced coal processing operations, encompassing comminution, beneficiation, flotation kinetics, and intelligent process optimization. However, conventional characterization approaches are inherently labor-intensive and susceptible to inter-operator subjectivity, hindering their integration into autonomous industrial monitoring systems. To surmount these challenges, this study proposes an explainable morphology-aware coal surface characterization framework that synergizes unsupervised pseudo-label generation, targeted data augmentation, and explainable artificial intelligence. Initially, a high-dimensional feature set comprising grayscale intensity statistics, entropy, edge density, and Gray-Level Co-occurrence Matrix descriptors was extracted from 454 coal surface images. Subsequently, Principal Component Analysis and K-means clustering were implemented to identify intrinsic morphology-driven structural patterns, generating robust pseudo-labels without manual annotation. These pseudo-classes were utilized to train and benchmark multiple transfer learning architectures, including EfficientNetB0, VGG16, and MobileNetV3. Experimental results demonstrated that MobileNetV3 achieved superior classification efficacy under a strict leakage-safe evaluation configuration, exceeding 90% across key performance metrics while offering significantly lower computational complexity—ideal for edge-computing deployment. Furthermore, Grad-CAM-based interpretability analysis validated that the models focused on physically significant morphological features, such as fracture boundaries and heterogeneous texture transitions. These findings indicate that the proposed framework provides a robust, computationally efficient, and interpretable decision-support tool for smart beneficiation and intelligent industrial coal processing environments.