| At the present stage,China is the country with the largest number of tunnels and underground projects in the world.Since the 13th Five-Year Plan,with the strategic goal of China's transportation power,the construction of underground engineering needs to develop in the direction of high quality,precision and intelligent construction.Because of the simplicity,flexibility,low cost and strong applicability of drilling and blasting method,rock tunnels such as railways,highways and metros are mainly excavated by drilling and blasting,and smooth blasting technology is the most widely used technology.Due to the complicated geological conditions,a large number of bedding and joint fissures often occur in the rock mass.When using smooth blasting technology to excavate tunnels,the number,spacing,distribution of blastholes and charge quantity under the corresponding geological conditions are still lack of systematic theoretical support.After tunnel excavation,serious overbreak and underbreak often occur,and even causes tunnel collapse and other accidents.Artificial intelligence methods such as image recognition and deep learning have been widely used in tunnel lining disease detection and vehicle recognition.However,the application of related technologies in the field of tunnel intelligent blasting is relatively lacking.Because the parameters of tunnel blastholes directly affect the effect after blasting,intelligent optimization of blasting parameters such as the number and spacing of blastholes can provide powerful technical support for fine blasting of tunnels.In this paper,to accurately and quickly acquire tunnel blasthole parameters and optimize smooth blasting parameters of horizontal layered rock mass tunnels,automatic recognition and coordinate positioning of tunnel blastholes were realized by means of image acquisition and classification of tunnel blastholes,improvement of deep learning method,theoretical analysis and field tests and monitoring.Based on this,parameters of number and spacing of blastholes were further obtained.Finally,the intelligent optimization of smooth blasting parameters was realized,and the engineering application was carried out.The main research contents and results of this paper include:(1)Based on the Panlong mountain tunnel project and Hailuoyu tunnel project,the datasets of tunnel blasthole target detection are constructed,including single blasthole image dataset,multiple blasthole image dataset and multiple blasthole difficult image dataset,which contain 28288 blasthole images.The datasets take into account the change of surrounding rock type,light intensity,shooting distance and shooting angle,and different factors such as shadows and obstructions.Three kinds of blasthole labels reflecting the types of blastholes and the fragmentation of surrounding rock were proposed:circular blasthole(bhc),elliptical blasthole(bhe)and irregular blasthole(bhi).The size range of the blastholes and the proportion of the small-scale blastholes were counted.(2)An improved Faster Region-based Convolutional Neural Network(R-CNN)method based on the lightweight network SqueezeNet for single blasthole object detection is proposed.The deep learning method integrates multi-scale input,top-down multi-layer feature fusion and distance-constrained Non-maximum Suppression(NMS)blastholes filtering algorithm,which can effectively recognize blastholes with different sizes and eliminate fake blastholes and misdetected blastholes.At the same time,the improved Faster R-CNN model hyperparameters were selected based on the control variable method.The number of training iterations,the number of anchors,the number of NMS max boxes,the strategy of the learning rate,the value of the classifier min overlap threshold,and the value of the overlap threshold were determined.Finally,the intelligent recognition of single blasthole with fast,high precision and high recall was achieved.(3)Two improved Faster R-CNN methods based on the deeper network ResNet-51 are proposed for target detection of multiple blastholes and multiple difficult blastholes respectively.The first deep learning method integrates two-stage training,top-down multi-layer feature fusion,online hard example mining and distance-constrained NMS multiple blastholes filtering algorithm,and realizes the recognition of high recall and precision of multiple blastholes with weak,large number and interval distribution.The second deep learning method integrates three-stage training,top-down multi-level regions of interest pooling,focal loss and distance-constrained NMS multiple blastholes filtering algorithm,and realizes the recognition of high recall and precision of multiple difficult blastholes with dark light,shadows and poor resolution.(4)The mechanism of damage and instability of tunnel surrounding rock in horizontal layered rock mass is explained from the blasting dynamic action and the static action after excavation.The blasting forming mechanisms of different parts of tunnel section in horizontal layered rock mass are explained in terms of blasting damage range of peripheral holes,angle change of between peripheral holes and joints.Through field tests,the smooth blasting parameters of the horizontal layered rock mass tunnel were optimized from adjusting distance between peripheral holes,thickness of smooth blasting layer,charge-concentration,charge structure,maximum charge of cut hole and adding empty hole,and the better shaping effect of tunnel excavation was achieved.(5)An algorithm for automatically obtaining the number and spacing parameters of tunnel blastholes based on the object detection results of multiple blastholes is proposed Combined with the relevant geological parameters,charge parameters and blasting effect obtained from field tests and monitoring,a coupling model of particle swarm optimization(PSO)and deep back propagation(DBP)neural network was introduced.Based on the model,the nonlinear mapping relationship among geological conditions of surrounding rock,blasthole and charge parameters of smooth blasting and arch crown settlement,sizes of overbreak and underbreak and rock size after blasting was established.Then,within the given searching range of smooth blasting parameters,the smooth blasting parameters satisfying the stability conditions of surrounding rock were obtained by using the PSO algorithm.The research results were successfully applied to the Panlong mountain tunnel with layered rock mass distribution,which significantly reduced the overbreak of the tunnel.Consequently,the excavation contour was well formed.Therefore,the current method proposed is of a broad application prospects. |
PDF Full Text Request