Research On Travel Time Estimation Of Open-pit Trucks Based On Spatio-temporal Graph Networks

Unreasonable scheduling of haul trucks in the process of open-pit mine production can lead to truck congestion and queueing,raising production costs.Precise travel time estimation allows the trucks to plan routes and avoid busy roadways,thereby reducing production costs.Therefore,real-time travel time estimation of open-pit trucks is the basis for achieving intelligent truck scheduling,and it’s critical for improving open-pit mining efficiency.This thesis conducts research on the travel time estimation of open-pit trucks.The main research contents are as follows:(1)Attention based spatio-temporal graph convolutional networks for travel time estimation of open-pit trucksThe traffic flow of open-pit trucks is highly correlated with time and space in travel time estimation.Existing methods have the problem of incomplete spatio-temporal feature extraction and ignore periodic traffic flow features when applying the travel time estimation of open-pit trucks.To solve these problems,this thesis proposes an Attention based spatio-temporal Graph convolutional networks for Travel Time Estimation of open-pit trucks(AGTTE).AGTTE includes three feature extraction modules: attribute feature,traffic flow feature,and adjacent road segments,as well as a travel time estimation module.Firstly,a spatio-temporal feature extraction component including a graph convolutional network and a convolutional neural network is used to capture traffic flow features from spatio-temporal dimensions.Furthermore,an attention mechanism is incorporated into graph convolutions to capture the dynamic correlation among spatial nodes.Secondly,three independent spatio-temporal feature extraction components are used to capture periodic features of surface mine trucks from recent,daily and weekly periods respectively.Then,a convolutional network is used to obtain the static features of adjacent road segments The word embedding is used to capture the external attribute information.Finally,the above features are connected and fused through a connection layer and multiple fully connected layers to estimate travel time.Experiments show that on a real dataset of an open-pit mine in Inner Mongolia,the mean absolute percentage error of AGTTE is reduced by 3.17% compared with the baseline method.(2)Transfer learning based graph convolutional networks for travel time estimation of open-pit trucksOpen-pit mine locations change frequently,and most advanced deep learningbased models cannot be easily applied to new areas with little traffic data.The road network and traffic flow distribution in different areas are very different,and the transfer performance of the overall feature transfer method is poor.In response to the above problems,this thesis proposes a Transfer Learning based Graph convolutional networks for Travel Time Estimation of open-pit trucks(TLGTTE).Firstly,in order to better capture the similar features of different open-pit mines,the road network and traffic flow feature graph of the source mine and the target mine are decomposed into smaller road network structures using the method of graph division.Then,the modelbased graph transfer learning is used to solve the problem that the new mine has little data and cannot perform traditional deep learning.Finally,the travel time of open-pit trucks is obtained through the multi-task learning module.Experiments show that on a real dataset of an open-pit mine in Inner Mongolia,the mean absolute percentage error of TLGTTE is reduced by 2.71% compared with the baseline method.(3)Prototype system and application for travel time estimation of open-pit trucksBased on the verification of the above method,a prototype system for travel time estimation of open-pit trucks is developed and integrated into the intelligent dispatching platform for open-pit mine.Firstly,the system composition structure of the platform is introduced.Secondly,it introduces the software functions,versions and hardware configurations used in system development.Thirdly,the logic of data preprocessing,model training and travel time estimation function of the truck travel time estimation module is introduced.Finally,the operation interface is shown in detail,including the operation steps of data preprocessing,parameter settings of model training,front-end display of training progress,and the actual application display of the model in open-pit trucks.The prototype system provides a visual operation interface for data input and output of each module,which can facilitate model training and estimation operations.The thesis has 44 figures,9 tables and 102 references.