在火电厂的生产过程中，煤尘通过高压运输进入焚烧室进行加热燃烧，在这个过程中，煤尘容易从管道薄弱处泄露，对火电厂的运行产生巨大的危害。传统的检测方法能够较为精确的检测泄露煤尘的参数，但是其硬件设备昂贵，且由于需要检测人员到现场进行采集，会有一定的危险性，针对以上问题，本文在利用图像的手段对泄露煤尘参数进行检测。

首先，针对目前没有开源的煤尘数据集的问题，本文搭建煤尘图像采集平台，构建煤尘图像数据集。在火电厂生产环境中放置煤尘采集装置，获得不同粒度范围的原始煤尘颗粒样本，将颗粒样本放置到图像采集平台的显微镜载物区，通过电脑实时采集图像样本；利用图像预处理手段和标注软件扩充并制作煤尘图像数据集。

其次，针对现有方法对轮廓不规则的颗粒和粘连区域进行分割具有精度不高的问题。本文提出一种基于Mask R-CNN的改进实例分割模型，通过向主干网络添加Ghost卷积与Spilt-Attention模块的级联结构，增加不同通道间煤尘颗粒的关注，同时减小了网络的参数量；在特征融合结构内添加一条自底向上路径，减少小目标颗粒的语义特征信息损失；在分割支路部分增加SimAM无参数，在不增加参数量的前提下，增加了掩膜的精度；最后，应用深度可分离卷积，平衡改进模块带来的计算负担。实验结果表明，改进的Mask R-CNN模型在不同煤尘粒径范围内的平均交并比和召回率均优于其他分割模型，在速度和精度之间实现了较好的权衡，验证了本文模型进行颗粒分割的有效性。

然后，针对煤尘图像中颗粒厚度信息不明确、难以提取的问题，建立一种煤尘颗粒厚度回归模型。通过与其他煤尘颗粒参数的比较，模型选取颗粒最佳外接矩形的长径、最佳外接矩形的短径、颗粒的投影面积这三个参数作为输入量。在预测煤尘颗粒厚度之前，使用帕森优化算法对GBDT与弹性网络的融合机器学习模型进行超参数寻优，确保模型的性能最佳，之后通过模型的输入量预测煤尘颗粒的厚度参数，实验表明本文提出的方法能够很好的预测煤尘颗粒的厚度。

最后，通过本文所提出的改进的Mask R-CNN分割算法，以及煤尘颗粒厚度预测模型，进行颗粒特征参数提取实验。首先，给出多种粒径的定义方法，经比较选取最佳外接矩形的长径、最佳外接矩形的短径、颗粒的投影面积作为本文颗粒粒径参数测量方法，并给出煤尘质量分布的定义。通过煤尘的质量公式，得到煤尘颗粒的质量模型，对火电厂四个特定环境进行误差实验，结果表明该模型在不同场景下的样本进行测试时均取得较好表现，最大误差小于10%，并且对煤尘颗粒的质量浓度与颗粒质量分布进行表征。

During the production process in thermal power plants, coal dust is transported under high pressure into the incineration chamber for heating and combustion. In this process, coal dust is prone to leakage from weak points in the pipelines, posing significant hazards to the operation of the power plant. Traditional detection methods can accurately measure the parameters of leaked coal dust, but they require expensive hardware equipment. Moreover, since inspection personnel need to collect samples on-site, there is a certain degree of risk involved. To address these issues, this paper proposes the use of image-based techniques to detect the parameters of leaked coal dust.

First, to address the lack of coal dust particle image datasets, a coal dust particle and image acquisition platform was set up to construct a coal dust image dataset. Particle collection boxes were placed in the production environment of coal preparation plants to obtain coal dust particle samples of different size ranges and in their original form through mechanical sieving and collection plates. These particle samples were then placed in the microscope loading area of the image acquisition platform, and image samples were collected in real-time via a computer. Image enhancement algorithms and annotation software were utilized to expand and create the coal dust image dataset.

Second, addressing the issue of low accuracy and time-consuming segmentation of irregularly shaped particles and adhesion areas by existing methods, this paper proposes an improved instance segmentation model based on Mask R-CNN. By adding a cascaded structure of Ghost convolution and Split-Attention modules to the backbone network, the model enhances the focus on coal dust particles across different channels while reducing the number of network parameters. An additional bottom-up path is incorporated into the feature fusion structure to mitigate the loss of semantic feature information for small target particles. In the segmentation branch, a parameter-free SimAM module is added to improve mask accuracy without increasing the parameter count. Finally, depthwise separable convolution is applied to balance the computational burden introduced by the improved modules. Experimental results demonstrate that the improved Mask R-CNN model outperforms other segmentation models in terms of average intersection over union and recall across different coal dust particle size ranges, achieving a good balance between speed and accuracy. This verifies the effectiveness of the proposed model for particle segmentation.

Next, to address the issue of unclear and difficult-to-extract thickness information of particles in coal dust images, a regression model for coal dust particle thickness is established. The model selects three features as input: the major axis length of the particle's best-fit bounding rectangle, the minor axis length of the best-fit bounding rectangle, and the projected area of the particle. These selected parameters are validated using the distance correlation coefficient method. After comprehensively comparing various machine learning regression models, a fusion model combining Gradient Boosting Decision Trees and Elastic Net is used to calculate the particle thickness information, and the Particle Swarm Optimization algorithm is employed to achieve automatic hyperparameter tuning of the model. Experimental results show that the proposed model can quickly calculate the thickness information of coal dust particles with only a few features.

Finally, experiments for extracting particulate characteristic parameters were conducted using the improved Mask R-CNN segmentation algorithm proposed in this paper and the coal dust particle thickness prediction model. Firstly, various definitions of particle size were presented. After comparison, the long axis and short axis of the best-fit bounding rectangle, as well as the projected area of the particle, were selected as the measurement methods for particle size parameters in this study. Additionally, the definition of coal dust mass distribution was provided. By utilizing the mass formula for coal dust, a mass model for coal dust particles was derived. Error experiments were carried out in four specific environments of thermal power plants. The results indicated that the model performed well when tested with samples from different scenarios, with the maximum error being less than 10%. Furthermore, the model was able to characterize both the mass concentration and mass distribution of coal dust particles.

[1] 王永臻, 杨丽丽, 门相勇, 等. 我国能源行业高质量发展面临的挑战与对策[J]. 中国矿业, 2025, 34(01): 70-78.

[2] 刘春花. 试论火电厂煤尘污染防治的方法与措施[J]. 科技风, 2013, (07): 1671-7341. 2013. 07. 213.

[3] 宁业敏, 陆斌, 梁裕庆. 火电厂输煤系统粉煤尘污染治理试验研究[J]. 广西电力技术, 1995, (04) 1671-8380. 1995.04.002.

[4] 陈清华, 许曾生, 王小润, 等. 基于滤膜称重法的自动化粉尘质量浓度检测装置的研究[J]. 煤炭学报, 2024, 49(07). BY23.1219.

[5] 王胜南, 李敏艳. 基于ECT的气固两相流颗粒浓度在线无损检测[J]. 电子测量与仪器学报, 2021, 35(12). B2104247.

[6] 贺振怀, 王杰. 基于β射线吸收法的粉尘浓度测量技术[J]. 煤矿安全, 2019, 50(07). 2019. 07. 018.

[7] 杜永潇, 孙晓立, 杨军, 等. 基于压电原理的装配式结构套筒灌浆饱满度检测方法研究[J]. 铁道科学与工程学报, 2024, 21(02). 43-1423.

[8] 彭朝阳, 李宏波, 周建明, 等. 2种快速光学法与魏氏法红细胞沉降率检测结果一致性分析[J]. 检验医学, 2024,39(06): 602-607.

[9] 陈茗, 胡边, 李靖. 静电感应法在线测量煤尘速度[J]. 电工技术, 2023, (02). 2023. 02. 017.

[10] 陈劲, 陈晓东, 赵辉, 等. 基于红外热成像法和超声波法的钢管混凝土无损检测技术的试验研究与应用[J]. 建筑结构学报, 2021,42(S2): 444-453.

[11] G. Zheng, Y. Yan, Y. Hu, W. Zhang, L. Yang and L. Li, "Mass-Flow-Rate Measurement of Pneumatically Conveyed Particles Through Acoustic Emission Detection and Electrostatic Sensing," in IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1-13, 2021, Art no. 9502413.

[12] M. Hajizadehmotlagh, D. Fahimi, A. Singhal and I. Paprotny, "Wearable Resonator-Based Respirable Dust Monitor for Underground Coal Mines," in IEEE Sensors Journal, vol. 23, no. 7, pp. 6680-6687, 1 April1, 2023.

[13] Z. Zhao, D. Li, G. Liu, F. Wu and J. Sui, "An Improved Dust-Concentration Measurement Algorithm Based on Multifeature Fusion of β-Ray Intensity Fluctuations," in IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 9, pp. 6420-6428, Sept. 2020.

[14] Zhang H, Nie W, Liang Y, Chen J, Peng H. Development and performance detection of higher precision optical sensor for coal dust concentration measurement based on Mie scattering theory. Opt Lasers Eng. 2021;144:106642.

[15] 孙毅,李长杨,毛亚郎等.基于光散射斑图成像特征的微细颗粒原位粒径检测方法[J].中国机械工程,2023,34(16):2001-2008.

[16] 梁良, 唐守锋, 童敏明, 等. 基于THz时域混沌特征的煤尘细度检测方法的研究[J]. 光谱学与光谱分析, 2019,39(05):1392-1397.

[17] V. Semenov et al., "Open-Air Miniature Fine Dust Sensor," in IEEE Sensors Journal, vol. 22, no. 6, pp. 5616-5627, 15 March15, 2022.

[18] F. Pedersini, "Improving a Commodity Dust Sensor to Enable Particle Size Analysis," in IEEE Transactions on Instrumentation and Measurement, vol. 68, no. 1, pp. 177-188, Jan. 2019.

[19] 马军伟, 杨付岭. 四电极静电传感器测量煤尘质量浓度的检测方法[J]. 矿山机械, 2024,52(03): 50-55. 2024. 03. 010.

[20] 梁红, 王凤箫. 一种改进的后散射型光电煤尘传感器[J]. 中国传媒大学学报(自然科学版), 2010, 17(03) :33-36.

[21] 魏明生, 童敏明, 梁良,等. 矿井煤尘粒度和浓度实时在线检测系统实验研究[J].煤矿安全, 2016,47(05): 30-33.

[22] 童敏明, 魏明生, 郝继飞. 基于改进同步迭代算法的矿井煤尘检测[J]. 自动化与仪表, 2012,27(02):6-8+52.

[23] 马立修, 谭博学, 姜静, 等. 基于双D光纤传感器的煤尘细度检测研究[J]. 煤炭学报, 2012,37(08): 1369-1372.

[24] H. Zhang et al., "Deep Multimodel Cascade Method Based on CNN and Random Forest for Pharmaceutical Particle Detection," in IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 9, pp. 7028-7042, Sept. 2020.

[25] 姬厚展, 高正阳, 李永华, 等. 在线检测中不同形状煤尘颗粒的流动特性[J]. 中国粉体技术, 2023,29(01):

[26] 李海滨, 孙远, 张文明, 等. 基于YOLOv4-tiny的溜筒卸料煤尘检测方法[J]. 光电工程, 2021, 48(06): 73-86.

[27] 李轩, 杨舟, 陶新宇, 等. 基于Mask R-CNN结合边缘分割的颗粒物图像检测[J]. 应用光学, 2023, 44(01):93-103.

[28] 王征, 张赫林, 李冬艳.特征压缩激活作用下U-Net网络的煤尘颗粒特征提取[J]. 煤炭学报, 2021, 46(09):3056-3065.

[29] 张伟, 隋青美. 基于小波变异本质粒子群和模糊熵的图像分割[J]. 光电子.激光, 2010,21(08):1264-1268.

[30] Xiang Yang, ZHAO Yindi, DONG Jihong. Remote sensing image mining area change detection based on improved UNet siamese network[J]. Journal of China Coal Society, 2019, (12).

[31] YAN Ran, LIAO Jideng, WU Xiaoyong, et al. Researchon classification method of sand and gravel aggregatebased on convolutional neural network[J]. Laser & OptoelectronicsProgress，2021，58（20）：211-218.

[32] Z. Wang, Z. Ji, X. Liu, J. Zhang and S. Yang, "An Efficient Segmentation Model With Multipath Attention Mechanism Enabling Particle Size Characterization of Coal Dust," in IEEE Transactions on Industrial Informatics, vol. 20, no. 4, pp. 6313-6324, April 2024

[33] ZHANG Fang, WU Yue, XIAO Zhitao, et al. Nanoparticlesegmentation based on U-net convolutional neuralnetwork[J]. Laser & Optoelectronics Progress， 2019，56（6）：137-143.

[34] 王宇, 陈婧, 王高. 基于改进的模糊C均值聚类算法的颗粒种子图像分割方法[J]. 中北大学学报(自然科学版), 2018,39(02):177-182.

[35] 谢涛. 基于深度学习的微细粒矿物识别研究[D]. 中国矿业大学, 2020.DOI:10.27623/d.cnki.gzkyu. 2020.001335.

[36] 徐江川, 金国强, 朱天奕, 等.基于深度学习U-Net模型的石块图像分割算法[J]. 工业控制计算机, 2018,31(04):98-99+102.

[37] Zheng Wang, Xu Zheng, Dongyan Li, Helin Zhang, Yi Yang, Hongguang Pan,A VGGNet-like approach for classifying and segmenting coal dust particles with overlapping regions,Computers in Industry,Volume 132,2021,103506.

[38] ZHU Daqing, CAO Guo. Particle size detection of sandstone images based on full convolutional network[J].Computer and Modernization，2020（7）：111-116.

[39] SHEN Zhichao. Research on detection algorithm of granular cropsbased on YOLO[D]. Harbin: Harbin Institute of Technology, 2021

[40] LI Xialin. Paticle defect detection based on machine vision[D]. Chengdu: University of Electronic Science andTechnology of China, 2019.

[41] Brendan J. Florio, Phillip D. Fawell, Michael Small,The use of the perimeter-area method to calculate the fractal dimension of aggregates,Powder Technology,Volume 343,2019,Pages 551-559,ISSN 0032-5910.

[42] WANG, ZHENG, LIU, XUFEI, JI, ZHAOXIANG, et al. Effective estimation model of coal dust characterization parameters with image sensing[J]. Measurement Science & Technology,2023,34(12 Pt.1):125008.1-125008.5.

[43] T. Andersson, M.J. Thurley, J.E. Carlson, A machine vision system for estimation of size distributions by weight of limestone particles, Miner. Eng. 25 (2012) 38–46.

[44] C.H. Hsieh, K.Y. Chen, M.Y. Jiang, J.J. Liaw, J. Shin, Estimation of PM2.5Concentration Based on Support Vector Regression With Improved Dark Channel Prior and High Frequency Information in Images, IEEE Access. 10 (2022) 48486–48498.

[45] Y. Jing, S. Guo, F. Chen, X. Wang and K. Li, "Dynamic Differential Pricing of High-Speed Railway Based on Improved GBDT Train Classification and Bootstrap Time Node Determination," in IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 9, pp. 16854-16866, Sept. 2022.

[46] 孟晗, 刘俊杰. 生物洁净室动态环境下多变量耦合的颗粒物浓度预测模型[J]. 科学通报, 2024,69(07):866-877.

[47] Rousseeuw P J , Driessen K V .A Fast Algorithm for the Minimum Covariance Determinant Estimator[J].Technometrics, 1999, 41(3):212-223.

[48] JIANG Bingyou, ZHANG Yuqian, YU Changfei, et al. Predictionof coal dust particle size after spraying dust reduction inroadway based on orthogonal experiment and regression analysis[J/OL]. Coal Science and Technology,2024-06-13.

[49] ZHOU Gang，NIE Wen，CHENG Weimin，et al. Influence regulations analysis of high-pressure atomization dust-settling to dust particle’s microscopic parameters in fully mechanized caving coal face[J]. Journal of China Coal Society， 2014， 39（10） ： 2053−2059.

[50] 赵政, 李德文, 吴付祥, 等. 基于多传感融合的粉尘质量浓度检测技术[J]. 煤炭学报,2021,46(07):2304-2312.

[51] ZHAO Zheng．Detection technology of metal dust concentration based on Mie scattering method ［J］． Instrument Technique and Sensor，2018( 5) : 108－110，119

[52] Shi L M, Zhang J H, Zhang D, et al. Developing a dust storm detection method combining Support Vector Machine and satellite data in typical dust regions of Asia[J]. Advances in Space Research, 2020, 65: 1263-1278.

[53] 孙国宝, 周继伟．物联网技术在智能油气田井场数字化建设中的应用[J]．信息系统工程，2021，10：16-18．

[54] 孟晨光，刘跃成，甄国涌．机器视觉领域的FPGA非线性双边滤波系统设计[J]．单片机与嵌入式系统应用，2023(7)：57-61．

[55] P. Meiyan, S. Jun, Y. Yuhao, L. Dasheng and Y. Junpeng, "M-FCN based sea-surface weak target detection," in Journal of Systems Engineering and Electronics, vol. 32, no. 5, pp. 1111-1118, Oct. 2021.

[56] J. Wang, C. Pang, X. Zeng and Y. Chen, "Non-Intrusive Load Monitoring Based on Residual U-Net and Conditional Generation Adversarial Networks," in IEEE Access, vol. 11, pp. 77441-77451, 2023.

[57] Y. Hualong and G. Daidou, "Research on Double Encryption of Ghost Imaging by SegNet Deep Neural Network," in IEEE Photonics Technology Letters, vol. 36, no. 10, pp. 669-672, 15 May15, 2024.

[58] X. Wang, B. Fan, Y. Fan, R. Xu, G. Feng and Q. Guan, "Method and Spatiotemporal Analysis for Impervious Surface Extraction Based on an Improved DeepLabV3+ Model," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, pp. 2893-2907, 2025.

[59] D. Xi, Y. Qin, J. Luo, H. Pu and Z. Wang, "Multipath Fusion Mask R-CNN With Double Attention and Its Application Into Gear Pitting Detection," in IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1-11, 2021, Art no. 5006011.

[60] D. Kumar and X. Zhang, "Improving More Instance Segmentation and Better Object Detection in Remote Sensing Imagery Based on Cascade Mask R-CNN," 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, Brussels, Belgium, 2021, pp. 4672-4675.

[61] X. Tang et al., "Image Synthesis and Modified BlendMask Instance Segmentation for Automated Nanoparticle Phenotyping," in IEEE Transactions on Medical Imaging, vol. 42, no. 12, pp. 3665-3677, Dec. 2023.

[62] C. Peng, L. Zheng, Q. Liang, T. Li, J. Wu and X. Cheng, "ST-SOLOv2: Tracing Depth Hoar Layers in Antarctic Ice Sheet From Airborne Radar Echograms With Deep Learning," in IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-14, 2024, Art no. 4303014.

[63] Zhang, Hang et al. “ResNeSt: Split-Attention Networks.” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2020): 2735-2745.

[64] Han, Kai et al. “GhostNet: More Features From Cheap Operations.” 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019): 1577-1586.

[65] L. Liang, Y. Zhang, S. Zhang, J. Li, A. Plaza and X. Kang, "Fast Hyperspectral Image Classification Combining Transformers and SimAM-Based CNNs," in IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1-19, 2023, Art no. 5522219.

[66] Z. Huang, L. Huang, Y. Gong, C. Huang and X. Wang, "Mask Scoring R-CNN,"2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 2019, pp. 6402-6411.

[67] Y. Sun et al., "IRDCLNet: Instance Segmentation of Ship Images Based on Interference Reduction and Dynamic Contour Learning in Foggy Scenes," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 9, pp. 6029-6043, Sept. 2022.

[68] W. Zheng et al., "Polygonal Approximation Learning for Convex Object Segmentation in Biomedical Images with Bounding Box Supervision," in IEEE Journal of Biomedical and Health Informatics.

[69] Z. Zhang and C. Jung, "GBDT-MO: Gradient-Boosted Decision Trees for Multiple Outputs," in IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 7, pp. 3156-3167, July 2021.

[70] H. Wang, C. Bian, L. Kong, Y. An, Y. Du and J. Tian, "A Novel Adaptive Parameter Search Elastic Net Method for Fluorescent Molecular Tomography," in IEEE Transactions on Medical Imaging, vol. 40, no. 5, pp. 1484-1498, May 2021.

[71] Zhang Z, Yang J, Su X, et al. Analysis of large particle sizes using a machine vision system[J]. Physicochemical Problems of Mineral Processing, 2013, 49(2): 397-405.

[72] Zhang Z, Liu Y, Hu Q, et al. Multi-information online detection of coal quality based on machine vision[J]. Powder Technology journal, 2020, 374: 250-262.

[73] 刘超, 张爱琳, 李树刚, 等. 基于Pearson特征选择的LSTM工作面瓦斯浓度预测模型及应用[J]. 煤炭科学技术, 2022: 1-9.

[74] 赵慧超, 刘耀东, 刘铭礼, 等. 基于Spearman相关性和燃烧三维数值计算的汽油机碳氢排放分析[J]. 吉林大学学报(工学版), 2023: 1-10.

[75] 孙宇豪, 李国通, 张鸽. 距离相关系数融合GPR模型的卫星异常检测方法[J]. 北京航空航天大学学报, 2021, 47(4): 844-852.

[76] 王震, 李映雪, 吴芳, 等. 冠层光谱红边参数结合随机森林机器学习估算冬小麦叶绿素相对含量[J]. 农业工程学报, 2024, 40(4): 171-182.

[77] 孟琪, 赵鹏, 宦克为, 等. 近红外无创血糖浓度的Label Sensitivity算法和支持向量机回归[J]. 光谱学与光谱分析, 2024, 44(3): 617-624.

[78] R. Kraemer, O. Düzgöl, S. Li and N. Calabretta, "Data-Driven SOA Parameter Discovery and Optimization Using Bayesian Machine Learning With a Parzen Estimator Surrogate," in Journal of Lightwave Technology, vol. 42, no. 2, pp. 721-731, 15 Jan.15, 2024, doi: 10.1109/JLT.2023.3316353.

[79] H. Zhu, H. Liu, Q. Zhou and A. Cui, "A XGBoost-Based Downscaling-Calibration Scheme for Extreme Precipitation Events," in IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1-12, 2023, Art no. 4103512, doi: 10.1109/TGRS.2023.3294266.

[80] Vallebuona G, Arburo K, Casali A. A procedure to estimate weight particle distributions from area measurements[J]. Minerals Engineering, 2003, 16: 323-329.

[81] Zhang Z, Liu Y, Hu Q, et al. Multi-information online detection of coal quality based on machine vision[J]. Powder Technology journal, 2020, 374: 250-262.

[82] Banta L, Cheng K, Zaniewski J. Estimation of limestone particle mass from 2D images[J]. Powder Technology, 2003, 132: 184-189.