In high-pressure transportation, characterizing the leakage status of coal dust is an effective means to reduce potential safety hazards in the optimization of energy structures, and it is also conducive to disaster prevention and safety management. Given the existing methods, manual inspection of leakage points requires high measurement skills, entails significant maintenance costs, and is time-consuming and challenging. Therefore, a synergetic network structure based on an instance segmentation, integrated with multiregression models, is proposed. This model is used to study the detailed characteristics of complex coal particles and estimate coal dust parameters, providing a practical means for onsite environmental assessment. First, a cascade mechanism of ghost convolution and a depthwise split attention module is added to the backbone network to reduce the number of network parameters and improve the channel correlation of coal dust images. Second, the multiscale feature pyramid network structure is introduced to increase low-level feature information in coal dust images and enhance attention to small particle characteristics of coal dust. Moreover, the head structure of the segmentation branch is optimized via the parameter-free attention module to improve mask precision. Finally, the optimized elastic network fusion model is used to estimate multiple regression coal dust parameters. The experimental results show that the proposed model outperforms the other models in terms of segmentation accuracy, the intersection ratio, and the recall ratio. The average error in the mass distribution characterization is less than ±10 %, which meets the theoretical expectations. An ideal balance is achieved between computational speed and segmentation accuracy.

中文翻译：

---

通过有效的基于图像的深度学习模型估计煤尘参数


在高压输送中，表征煤尘的泄漏状态是优化能源结构中降低安全隐患的有效手段，也有利于灾害预防和安全管理。鉴于现有方法，手动检查泄漏点需要很高的测量技能，需要大量的维护成本，并且非常耗时且具有挑战性。因此，提出了一种基于实例分割的协同网络结构，并与多回归模型集成。该模型用于研究复杂煤颗粒的详细特征和估计煤尘参数，为现场环境评估提供实用手段。首先，在骨干网络中加入幽灵卷积的级联机制和深度分离注意力模块，以减少网络参数的数量，提高煤尘图像的通道相关性;其次，引入多尺度特征金字塔网络结构，增加煤尘图像中的低级特征信息，增强对煤尘小颗粒特征的关注;此外，通过无参数注意力模块优化了分割分支的头部结构，以提高掩码精度。最后，采用优化的弹性网络融合模型估计多元回归煤尘参数。实验结果表明，所提出的模型在分割精度、交集率和召回率方面优于其他模型。质量分布表征的平均误差小于 ±10 %，符合理论预期。在计算速度和分割精度之间实现了理想的平衡。