| Disk pelletizing is one of the key steps in the modern iron ore pellet production process,with the purpose of aggregating the mixture of iron ore fines and other raw materials into green pellets in a disk pelletizer.The particle size distribution(PSD),fill level and velocity field of the green pellets are the key parameters in the disk pelletizing process,which have significant impact on the product quality,efficiency and the energy consumption of the subsequent process.Therefore,studying the rapid and accurate measurement methods of these parameters is the premise and foundation for realizing the optimal control of the disk pelletizing process,and has important theoretical and engineering application value.Machine vision is an image-based technology for sensing,analyzing and understanding target objects,which has been widely used in process inspection and measurement in a variety of industries.However,the working conditions of the disk pelletizing process are quite complex,such as the high rotation speed of disk,wet and rapidly moving green pellets,and overlapping particles.Under certain conditions,dusts and water steams(haze)may randomly occur during the pelletizing process.These complex factors bring many challenges to the existing machinevision based measuring methods,and make it difficult to measure the above-mentioned key parameters,or the measuring accuracy and speed to be improved to meet the engineering requirement.To deal with the above difficulties,and with the support of the National Natural Science Foundation Project "Research on the on-line detection and automatic control of large-scale pelletizing process based on machine vision"(No.: 61973108)and the Postgraduate Innovation Project of Hunan Province "Research on the key algorithms of image processing of stacked granular objects based on deep learning"(No.: CX20190302),the present work systematically studies the machine vision based methods for accurate and fast measurement of the key parameters in disk pelletizing process,including particle size distribution of green pellets(quality parameter),fill level and particle velocity field(process parameters).The research contents and major contributions are detailed as follows.(1)Considering the problems of uneven illumination,particle overlapping and great scale differences among granules in the captured images,a machine vison method based on lightweight deep leaning network LUnet is proposed to realize the fast and accurate measurement of PSD of green pellets.Firstly,taking U-net as the backbone,channel number reduction and introducing Batch-normalization layer are carried out to establish a lightweight deep learning network LUnet for the fast and accurate image segmentation.According to the image segmentation results by LUnet,a contour detection method based on logic operation is then proposed to obtain the contour information via separating the densely stacked pellets in the image;Finally,the particle size is obtained by ellipse fitting,and the particle size distribution is obtained via the statistics of all the detected particles.In this thesis,a dataset Pellet Images1# is established based on images captured from an industrial disc pelletizer,and extensive comparative experiments are conducted.The results demonstrate that the proposed LUnet achieves better segmentation accuracy than other methods;Meanwhile,The proposed contour detection method is adaptive to handle densely overlapping and irregularly shaped particles,and it is universal to many application fields.Compared with U-net,the AUC and DICE indexes of image segmentation by LUnet are increased by 8% and 12%,respectively,while the measuring error of PSD is decreased by 14%.Moreover,the amount of network parameters is reduced by 52.7%,and the processing speed is accelerated by about 34.5%.By use of the proposed method,the accurate and fast measurement of the PSD of green pellets in haze-free environment is realized,and additional product quality parameters such as the roundness distribution and the proportion of high-quality green pellets can be provided.(2)Under certain conditions of environment temperature,humidity and water content of the material,haze may occur randomly during the pelletizing process,which results in the degradation of image quality,and consequently increased PSD measuring errors or even failures.To solve this problem,a feature fusion based deep network VGG16-LUnet is proposed in this thesis.Firstly,a haze judgment module based on K-means clustering is proposed as the proprocessing step to enhance the saliency of the features of particle areas;Then,taking the LUnet as backbone,a deep network VGG16-LUnet with higher feature extraction ability is established to achieve higher segmentation accuracy;Based on the image segmentation results,the contour information of each particle is extracted by using the contour detection method in(1),and consequently obtains the statistical result of PSD.In this thesis,the dataset Pellet Images2# is established by using the images captured in the industrial pelletizing process with random appearance of haze,and extensive experiments are conducted.The results show that the m Io U index of the image segmentation results by the proposed method for hazy and haze-free images is 0.83 and 0.88 respectively,and the DICE index is over 0.9.The PSD measurement error by the proposed method and the manual measurement is 1.2 and 0.5 for hazy and haze-free images respectively,which outperforms other methods.Additionally,the total time consumption for image segmentation and PSD measurement for an image(256 × 256)is about 110 ms,which meets the requirements of the robust and online measurement of PSD of green pellets in environment with random appearance of haze.(3)In order to solve the difficulties in the measurement of fill level caused by the tilt of the disk,fast movement of the particles in the disk,and the varying shapes of the material pile in disk pelletizing,one RGB-D camera is used in the present work to capture the data of the materials in the disk,and a fill level measuring method based on the fusion of image and point cloud data is proposed.Firstly,K-means clustering is adopted to extract the region of interest(ROI)of the depth data;Then,a method based on graph-cut is proposed to obtain the material region in RGB image;Next,the ROI depth data and the material region in RGB image are fused and the transform matrix is adopted to generate the point cloud of the material pile in the disk;On this basis,a "proportion conversion" algorithm based on mathematical model is derived for fast calculation of the fill level.Extensive experiments are conducted to test the proposed method on a pilot-scale disk pelletizer in the laboratory.The experiment results show that the proposed method achieves high measuring accuracy at various disk rotation speeds and material amounts.The average relative error of the fill level measurement is 5.4%,and the calculation time for the RGB-D data with a projection area of 640×480 is less than 2s.This approach provides an end-to-end solution for the monitoring of the dynamic change of the fill level in disk pelletizers,and is potential in engineering applications.In order to solve the problem that the existing machine vision methods have difficulties to meet the requirements of fast and accurate measurement of particle velocity field when the disk pelletizer rotates at high speed,we adopt an industrial camera for image data capturing of moving particles in the disk,and propose a measuring method for velocity field based on an unsupervised pyramid optical flow network.Firstly,an autoencoder deep newtwork FPNFlow Net is built by introducing the pyramidal structure feature extractor FPN to obtain the raw dense velocity field of the input image pair.The encoder extracts the features that are robust to the “great movements” and the brightness changes,meanwhile,the features have less number of layers,which ensures the processing speed.Additionally,introducing of the weighted fusion calculation can optimize the training procedure,and it effectively solves the high consumption of time and computer memory problem in the existed deep networks with high measurement accuracy.Next,an automatic particle center position detection method based on graph-cut and VGG16-LUnet deep network is designed,and the velocity field is obtained by assigning the detected particle positions to the raw velocity field.In this thesis,the image datasets of moving particles in disk pelletizer and rotary drum are established,and a large number of experiments are conducted for quatitative and quantitative comparative analysis.The results demonstrate that the proposed method achieves competitive measuring accuracy to the state-of-the-art method(RAFT-PIV),while the time consumption is only 13% of that of it,and the computer memory occupation is greatly reduced by more than 80%,which has obvious advantages.The proposed method realizes fast measurement of particle velocity field in disk pelletizers,which plays a positive role in optimizing the pelletizing process and improving the product quality. |
PDF Full Text Request