Research And System Implementation Of TBM Rock Fragments Form Recognition Based On Deep Learning

During the construction of tunnel boring machine(TBM),mucking is an important link after cutterhead excavation.Rock fragmentation can directly reflect the tunneling state.The uneven distribution of rock fragments often reflects the failure of roadheader cutters or unreasonable setting of tunneling parameters.In addition,fragments with too large particle size will also cause physical damage to the conveyor belt.The traditional manual monitoring method of mucking requires a lot of manpower and material resources and the judgment results are affected by subjectivity,which leads to low construction efficiency and potential construction risks.In this paper,deep learning and computer vision are used as the main technical means to study the automatic recognition technology of TBM mucking state,which provides technical support for unmanned mucking monitoring in TBM construction.The main research contents and work are listed as follows:1.According to the construction requirements,the overall scheme of TBM mucking state recognition system based on deep learning is designed.The functional requirements of each part of the system are analyzed,the system hardware selection is carried out according to the requirements,and the flow of system algorithm is designed;2.Install our image acquisition system at a selected location of the construction site to acquire the original image.Methods of clipping,filtering and contrast saturation enhancement are used to preprocess the original image.According to the needs of TBM mucking monitoring,the geological conditions of the site and relevant national standards,a three-classification scheme of TBM mucking images is concluded,and the manual labelling is completed according to the scheme.In order to solve the problem of data imbalance between classes,the method of image transformation with label reservation is used to augment the data;3.Based on the comparative study of various models,according to the test accuracy and algorithm running speed,AlexNet is selected as the basic network for structure and parameters design.The network models are trained with the collected rock fragments dataset.The optimal parameter set for rock fragment image recognition is obtained through multiple sets of parameter settings comparative experiments.The accuracy,stability and generalization of the optimal model are verified on the test dataset;4.Transplant the trained convolutional neural network model to the embedded platform of Jetson TX2,use Tensor RT inference optimizer to accelerate the model,design the hardware and develop the front-end interface integrating all functions of the system,so as to make the whole system portable and easy to use,and meet the requirements of construction site deployment.The research of this paper shows that the TBM mucking state recognition model based on deep learning has 93.81% accuracy.The algorithm has high real-time performance on Jetson TX2,and the recognition speed can reach 12 fps,which meets the deployment requirements of TBM mucking monitoring.