Ash content is a key indicator to evaluate coal flotation concentrate quality and adjust flotation process parameters, which could be determined by analyzing froth images. In this research, a multi-scale multi-task neural network (MSTNet) was developed to realize accurate determination of the ash content of industrial coal flotation concentrate by analyzing froth images. Furthermore, transfer learning is used to further improve model accuracy for low-resolution images. Results obtained using industrial data show that MSTNet achieves a higher prediction accuracy while requiring less computations than previous models. It reaches the maximum R2 of 0.9063 with a processing time of 0.0035 seconds per image, while its competitors only reach the maximum R2 of 0.7231 with a processing time of 0.0038 seconds per image. This suggests that MSTNet surpassing its competitors in both accuracy and speed. Furthermore, MSTNet achieves the minimum MAPE of 0.0300, indicating that MSTNet has a mean relative prediction error of ± 3 %. This proves the high prediction accuracy of MSTNet. These results indicate that the proposed MSTNet holds great promise for practical applications. Its practical application will lead to more efficient and intelligent coal production.

中文翻译：

---

多尺度多任务神经网络结合迁移学习准确测定工业浮选煤精矿灰分


灰分是评价浮选选矿质量、调整浮选工艺参数的关键指标，可通过分析泡沫图像来确定。本研究开发了一种多尺度多任务神经网络 （MSTNet），通过分析泡沫图像实现对工业浮选煤精矿灰分的准确测定。此外，迁移学习用于进一步提高低分辨率图像的模型准确性。使用工业数据获得的结果表明，与以前的模型相比，MSTNet 实现了更高的预测精度，同时需要的计算量更少。它达到了 0.9063 的最大 R2，每张图像的处理时间为 0.0035 秒，而其竞争对手仅达到了 0.7231 的最大 R2，每张图像的处理时间为 0.0038 秒。这表明 MSTNet 在准确性和速度上都超过了其竞争对手。此外，MSTNet 的最小 MAPE 为 0.0300，表明 MSTNet 的平均相对预测误差为 ± 3%。这证明了 MSTNet 的高预测精度。这些结果表明，所提出的 MSTNet 在实际应用中具有很大的前景。它的实际应用将导致更高效、更智能的煤炭生产。