Publications

Particle size analysis of rock fragments plays a crucial role in mining engineering. However, traditional non-contact image-based methods typically rely solely on RGB images, which are highly sensitive to illumination changes, shadow interference, and fragment texture. This reliance limits the accuracy and generalization capability of conventional approaches and often necessitates retraining when applied across different scenes. To address these issues, this study introduces normal maps and provides a detailed analysis of their advantages in representing rock fragment features. Furthermore, we propose a multimodal instance segmentation framework named Adaptive Feature Recombination Network (AFRNet). AFRNet incorporates a modality effectiveness perception mechanism to adaptively guide the fusion process while suppressing interference from unreliable modalities. In addition, it employs a multi-scale attention fusion module to fully exploit and utilize the strength of each modality. This study systematically compares three fusion strategies—data-level, feature-level, and decision-level—and conducts experiments under various modality combinations. Experimental results demonstrate that incorporating normal maps significantly improves segmentation accuracy and enhances model robustness in degraded environments such as low illumination and shadow interference. Moreover, the model trained in a laboratory environment is directly transferred, without retraining, to a practical particle size analysis task at an actual mining site in Nanjing. The resulting particle size distribution curves exhibit a deviation of less than 10% compared with manually labeled results, validating the proposed method’s zero-cost transferability and engineering applicability.

Borehole breakout (BO) has increasingly been utilized to estimate in-situ stress magnitudes given the importance of the stress field in subsurface activities and the limitations of conventional stress measurement techniques. In this study, a new backpropagation neural network model is developed to estimate both maximum and minimum horizontal stress magnitudes from multiscale BO data. A total of 150 experimental data points from pre-stressed true-triaxial laboratory tests and 44 field data from a mine site in Australia and the literature are collected and employed for model development and validation. Compared to previous studies, the collected data set is significantly enhanced in both quantity and quality. To address discrepancies in stress magnitudes between experimental and field data, the three principal stresses are normalized by borehole wall strength (BWS). Overall, the model achieves mean absolute percentage errors of below 8 per cent for the maximum horizontal stress and below 20 per cent for the minimum horizontal stress, significantly outperforming the previous model developed for this purpose. Furthermore, these error rates fall within the typical error range (10–20 per cent) of conventional stress measurement techniques, indicating the model’s sufficient accuracy for practical applications. Moreover, the effectiveness and generalizability of the model are verified using 166 additional BOs from two mine sites, which are independent of those used in model development. Continuous and detailed stress profiles are established based on these BOs, covering greater depth intervals than the stress measurements from the overcoring method. The results of this study demonstrate that the proposed model can provide reliable and accurate stress estimation, utilizing input parameters that can be readily obtained from borehole geophysical logs.