Attention-based spatiotemporal graph convolution for open-pit mine truck travel time prediction

Abstract

A method for predicting the travel time of open-pit mine trucks based on attention-based spatiotemporal graph convolution belongs to the field of mine scheduling. First, the original trajectory data and road network data of the open-pit mine truck are obtained, and the trajectory data and road network data of the open-pit mine truck with attribute information are obtained by combining the external factor attribute information; secondly, the attribute information characteristics of the factors in the trajectory are extracted; thirdly, the traffic flow information characteristics in the time and space dimensions are captured by the spatiotemporal graph convolution network based on the attention mechanism; then, the characteristics of the adjacent road sections are captured by the convolution network; then, the above three parts of feature information are connected, the prediction model and loss function are designed, and the model is trained; finally, according to the trajectory to be predicted, the trained model is used to predict the travel time. Advantages: Make full use of the trajectory, road network and external information characteristics to achieve accurate prediction of the travel time of open-pit mine trucks. Optimize the open-pit mine truck scheduling system, reduce operating costs, and improve mining efficiency.

Description

Technical Field

[0001] The present invention relates to the field of smart mining systems, and in particular to an attention-based spatiotemporal graph convolution open-pit mine truck travel time prediction method. Background Art

[0002] Open-pit mining is a large-scale production system centered on mining and linked by transportation. Trucks are the most important mining and transportation tools in the open-pit mining process, and the mining and transportation costs account for more than half of the mine production costs. Therefore, in recent years, many open-pit mines have begun to apply open-pit mine truck scheduling systems, and continuously optimize truck scheduling systems to reduce operating costs and improve mining efficiency. Accurately predicting truck travel time is an important part of the open-pit mine scheduling system. Accurate and real-time truck travel time prediction is an important research content for realizing truck scheduling optimization.

[0003] In actual mining operations, truck travel time may be affected by factors such as road environment, truck type, meteorological conditions, and different time parameters. Traditional research only guarantees the accuracy of a single road section, and obtains the travel time of a given trajectory by summing the time spent on each road section, which will lead to cumulative errors. Existing graph network-based methods mostly use structured traffic flow graphs constructed according to road networks to predict travel time. Accurately and effectively obtaining traffic flow characteristics is the key to improving travel time prediction. The complex traffic flow information of open-pit mines is highly closely related to time and space, and existing research has not comprehensively extracted such spatiotemporal features. On the other hand, existing research considers the traffic flow information of the current period, but it is very likely that there are no vehicles passing during this period, resulting in sparse data. Summary of the invention

[0004] The purpose of the present invention is to provide an attention-based spatiotemporal graph convolution open-pit mine truck travel time prediction method to solve the problem that traditional graph networks do not fully utilize the spatiotemporal information characteristics of traffic flow and that there are no vehicles passing through the road for a period of time, resulting in sparse data.

[0005] The object of the present invention is achieved in this way. The present invention adopts an end-to-end learning method;

[0006] Firstly, the original track data of open-pit mine trucks and the original road network data are processed, and combined with the attribute information of external factors, the track data of open-pit mine trucks and the road network data of open-pit mines with attribute information are obtained;

[0007] Secondly, the factor attribute information in the open-pit mine truck trajectory is extracted to obtain the embedded representation of the attribute features;

[0008] Thirdly, the traffic flow information characteristics of the time dimension and the space dimension are captured simultaneously through the spatiotemporal graph convolutional network based on the attention mechanism; the traffic flow information is the recent, daily and weekly historical traffic flow information;

[0009] Then, a convolutional network is used to capture the feature information of neighboring road segments;

[0010] Secondly, the three parts of feature information are connected and the model is trained according to the loss function;

[0011] Finally, the trajectory to be predicted is input, and the final travel time prediction result is obtained based on the trained model.

[0012] The specific steps are as follows:

[0013] Step S1: sampling the position data of open-pit mine trucks to form the original trajectory data of open-pit mine trucks, extracting, cleaning and denoising the original trajectory data, and combining the attribute information such as weather and shifts to establish the trajectory data of open-pit mine trucks with attribute characteristics; correcting the existing road network data through the existing road network generation algorithm to obtain the data of the open-pit mine road network with attribute characteristics;

[0014] Step S2: extracting the attribute information of factors affecting truck travel time, and converting the information into trajectory attribute features by using a word vector embedding method; based on the open-pit mine truck trajectory data and open-pit mine road network data obtained in step S1, using the existing map matching method, obtaining the original route information of each trajectory in the road network; combining the original route information with the trajectory attribute features to obtain the attribute route information;

[0015] Step S3: Based on the open pit truck trajectory data and the open pit road network data obtained in step S1, a traffic flow characteristic map is constructed, and the traffic flow characteristic map is input into a traffic flow feature extraction module to capture recent, daily and weekly traffic flow information features;

[0016] Step S4: Based on the original route information of each trajectory in the road network obtained in step S2, input it into the adjacent road segment feature extraction module to capture the feature information of the adjacent road segment;

[0017] Step S5: Establish a multi-task learning module; connect the attribute route information, traffic flow information features and adjacent road segment features obtained in steps S2-S4, and input them into the prediction model, design a multi-task loss function, combine the known trajectory and road network data set, and train the model for truck travel time prediction;

[0018] Step S6: Finally, according to the real-time truck trajectory obtained in the actual open-pit mine production, the model obtained in step S5 is used to predict the truck travel time.

[0019] In step S1:

[0020] The original trajectory data T of the open pit truck is defined as: T = {T1, T2, ..., T i , ..., T N} is a series of raw track data of open pit trucks, N represents the number of tracks; T in T i Expressed as in represents the jth trajectory point in the i-th trajectory, Includes the time, longitude and latitude of the trajectory point; k is the number of trajectory points in the i-th trajectory.

[0021] The open pit truck trajectory data T′ is defined as: T′={T′1, T′2, ..., T′ i , ..., T′ N} is a series of open pit truck trajectory data, where N represents the number of trajectories; Represents the trajectory T i And its four factor attribute information. Indicates the travel distance of the changed trajectory, Indicates weather conditions, Indicates vehicle schedule information. Indicates driver.

[0022] The open pit mine network G is defined as: G = (V, E) is an undirected graph, V and E represent the nodes and edges in the graph respectively; V = {v1, v2, ..., v l , ..., v m}, where v l is a road segment in the road network; e ij =(v i , v j )∈E represents the road segment v i and v j Connect or not.

[0023] In step S2:

[0024] The travel time of open pit trucks is affected by multiple factor attribute information, which are all in the open pit truck trajectory data of step S1. The factor attribute information includes: weather, distance, driver and shift; the factor attribute information is a classification value, and each classification attribute is converted into a low-dimensional real vector using the traditional word vector embedding method as a trajectory attribute feature. Through the open pit truck trajectory data and the open pit road network data, the original route information of each trajectory in the road network is obtained using the existing map matching method. The original route information is a series of road sections in the open pit road network; a route R is defined as: Where R dThe number of sections in the route. Assign the attribute information of the trajectory to each node in the original route to obtain the attribute route of the trajectory. The attribute route information of section l at time t can be expressed as The formula is as follows:

[0025]

[0026] Where W e represents the weight of the fully connected layer, A dis Indicates the travel distance (Distance), A wea Indicates weather conditions (Weather), A duty Indicates the vehicle schedule (Duty), A dri Indicates the driver (DriverID), symbol Represents a connection between attributes.

[0027] In step S3:

[0028] The vehicle flow characteristic diagram is defined as: represents the characteristics of all nodes at time t; represents the characteristic value of node h at time t; the structure of the traffic flow characteristic graph is the same as that of the road network, and the node information is represented as the traffic flow characteristic value. Ψ=(ψ1,ψ2,...,ψ t , ..., ψ ξ ) represents a set of traffic flow characteristic graphs, and ξ represents the number of traffic flow characteristic graphs.

[0029] The vehicle flow feature extraction module is a spatiotemporal graph convolutional network based on the attention mechanism. The process is as follows: First, a vehicle flow feature graph is constructed based on the open-pit mine truck trajectory data. Then, the spatiotemporal feature extraction component is used to comprehensively extract the vehicle flow features from the time dimension and the spatial dimension. The component consists of a spatial feature extraction unit and a temporal feature extraction unit. The spatial feature extraction unit is used to extract the dynamic vehicle flow features of the spatial dimension of each road section in the vehicle flow information, and the temporal feature extraction unit is used to extract the features of the temporal dimension of each road section. Next, three independent spatiotemporal feature extraction components are used to capture the recent, daily and weekly historical vehicle flow information respectively, and finally a fully connected layer is used to merge and connect the data generated by the above three components.

[0030] The specific process of generating the traffic flow characteristic map is as follows: the average speed and traffic volume of passing vehicles on each road section are calculated based on the open pit truck trajectory data, and used as traffic flow characteristics. Combined with different time nodes, the recent, daily and weekly historical traffic flow characteristic map data are obtained from the traffic flow characteristic map. Specifically, the recent traffic flow characteristic map data is The traffic flow characteristic diagram data of the daily cycle is The traffic flow characteristic diagram data of the weekly cycle is Where t0 is the current time, f is the time interval for generating the traffic flow characteristic graph, and q is the number of traffic flow characteristic graphs generated every day. p =kf (k=1, 2, ..., n) is the time segment. b T p The number of traffic flow characteristic graphs in the time series segment. b F is the traffic flow feature vector.

[0031] The spatial feature extraction unit is a convolutional network with attention. The convolutional network uses a spectral domain graph convolution calculation method, which can effectively capture the correlation of neighboring nodes in the traffic flow feature graph. First, the Laplace matrix of the traffic flow feature graph is constructed. Then, the Laplace operator in the frequency domain is derived using the Laplace matrix, and then the convolution in the Euclidean space in the frequency domain is analogized to finally obtain the formula for graph convolution.

[0032] The eigenvalue decomposition of the Laplace matrix is ​​L = U(diag(A))U T , where diag(A) is the diagonal matrix of A, U is the Fourier basis, and L is the eigenvector matrix of the normalized Laplace.

[0033] Formula (2) is the graph convolution operation formula:

[0034] Θ*Gψ=Θ(L)ψ=Θ(U(diag(A))U T )ψ=UΘ(diag(A))U T ψ (2)

[0035] Among them, *G represents the graph convolution operation, Θ is the convolution kernel, and ψ represents the traffic flow feature map.

[0036] Formula (2) is to perform Fourier transform on Θ and ψ respectively, then multiply their transformed results, and then perform inverse Fourier transform to finally get the result of convolution operation. The cost of directly calculating the equation by graph convolution method is extremely high, with a complexity of O(n 2 ). When the scale of the graph is large, the computational efficiency will be greatly reduced. Chebyshev polynomials can be used to approximately but effectively solve this problem.

[0037] Update itself through k-1 neighbor nodes, the formula is as follows:

[0038]

[0039]

[0040] T k (ψ)=2ψT k-1 (ψ)-T k-2 (ψ) (5)

[0041] Among them, θ k are the coefficients of the polynomial vector, I N is the identity matrix, λ max is the largest eigenvalue of the Laplacian matrix, is a Chebyshev polynomial, and its derivation is shown in formula (5), where the initial terms T0(ψ)=1 and T2(ψ)=ψ.

[0042] In actual road traffic scenarios, the road conditions at different locations affect each other, and this mutual influence changes dynamically. In order to make full use of this feature, an attention mechanism is used to adaptively capture the dynamic correlation between spatial nodes. The attention mechanism formula is as follows:

[0043] S=V s tanh((ΨW1)W2(W3Ψ) T +b s ) (6)

[0044]

[0045] in, Represents graph input, C r is the number of channels of the graph input data, T r is the length of the time dimension of the traffic flow feature graph, tanh is the activation function, W1, W2, W3, V s and b s is a learnable parameter. The element S in S i，j The value of represents the correlation strength between node i and node j in semantics. Softmax represents the normalization process, and finally the attention matrix S′ is obtained. During the operation, the adjacency matrix A and the attention matrix S′ are used to dynamically adjust the influence weights between nodes.

[0046] The attention mechanism allows the network to automatically pay more attention to valuable information. It allows the spatiotemporal graph convolution to update node information more effectively. In order to dynamically update the value of each node, the attention matrix S′ is added to each item in the Chebyshev formula to obtain the attention-based update equation for each node:

[0047]

[0048] The time feature extraction unit is a convolutional network that uses a common network to extract traffic flow features and updates the feature values ​​of its own nodes by capturing information at adjacent time slices. For the convolution of the time dimension, the input value is According to formula (8) and the convolution formula, the calculation formulas are shown in (9) to (11):

[0049]

[0050]

[0051]

[0052] From the above formula we can get the recent traffic flow characteristic matrix ST h , the traffic flow characteristic matrix ST of the daily cycle d And the weekly traffic flow characteristic matrix ST w Since the characteristic matrices of the three time periods have different influences on the predicted traffic flow characteristics, it is necessary to add the parameter W before the three matrices. h , W d and W w , and the final traffic flow characteristic matrix is ​​obtained:

[0053]

[0054] In step S4:

[0055] In the route R of the open-pit mine, the characteristic information of the adjacent sections of the route R plays an important role in estimating the travel time of the truck; the adjacent sections are defined as SP l =[v l-w , v l-w+1 , ..., v l-1 , v l , v l+1 , ..., v l+w ], indicating v l The neighboring road segments are represented by w, where w is the context window size; the feature information of the neighboring road segments includes the width of the road, the angle of the corner, and the slope of the road segment. The traditional CNN convolutional layer is used to capture the feature information of the neighboring road segments, and the formula is as follows:

[0056]

[0057] Among them, σ represents the fully connected layer calculation, and Emb is the representation layer.

[0058] In step S5:

[0059] The purpose of the multi-task learning module is to aggregate the previously obtained features and then predict the travel time. First, a connection layer is used to aggregate the attribute information features Traffic flow characteristics and the static characteristics of the road network of adjacent road segments Multiple fully connected layers are then used to predict the travel time for each segment of a given route.

[0060] The travel time for a given path Q is calculated as follows:

[0061]

[0062] MLP stands for multiple fully connected layers. represents the predicted travel time of segment l of path Q, and then the travel time of each segment is added up to get the travel time of the entire path.

[0063]

[0064] Among them, m is the number of sections contained in the given path Q, represents the travel time to be predicted;

[0065] The multi-task loss function obtains the travel time of each section and the travel time of the entire path through the fully connected layer. During the training process, the mean absolute percentage error is selected as the loss function for the estimation of the travel time of the entire path and the section. The mean absolute percentage error is expressed in English as MAPE; a path loss function is defined as:

[0066]

[0067] in, is the predicted travel time of path i, Y i is the actual travel time of path i;

[0068] The segment loss function of a path is defined as:

[0069]

[0070] Among them, m is the number of road segments contained in the i-th path, is the predicted travel time of the jth segment in the i-th path, t j is the actual travel time of the jth segment in the i-th path;

[0071] The loss functions of the path and segment are combined into:

[0072]

[0073] Finally, the final model is obtained by training based on the multi-task loss function.

[0074] Beneficial effects, due to the adoption of the above scheme, the end-to-end learning idea is adopted. First, some external information, such as weather, shift, date, distance and other attribute information, is extracted and embedded. Then, through the spatiotemporal graph convolution based on the attention mechanism, the recent, daily and weekly traffic flow information characteristics are captured, and the time and space information characteristics can be captured at the same time. The information characteristics of the adjacent road sections are captured through a convolutional network. Finally, the three parts of feature information are connected to obtain the final travel time.

[0075] Compared with the prior art, the present invention adopts the above technical solution and has the following significant advantages:

[0076] The present invention introduces an attention-based spatiotemporal graph network into the open-pit mine travel time prediction. On the one hand, it makes full use of the attribute information of the open-pit mine, such as weather, shifts, and distance, to extract features. On the other hand, on the basis of the traditional graph network, an attention mechanism is added to realize the simultaneous use of road section spatiotemporal information. Furthermore, the graph network based on the attention mechanism captures the recent, daily, and weekly historical cycle characteristics, which effectively improves the prediction accuracy of the model. It realizes more accurate open-pit mine truck travel time prediction, optimizes the truck dispatching system, reduces operating costs, and improves mining efficiency.

[0077] It solves the problem that traditional graph networks do not fully utilize the spatiotemporal information characteristics of traffic flow and that there are no vehicles passing through the road for a period of time, resulting in sparse data.

[0078] The method provided by the present invention can provide a method for predicting the travel time of open-pit mine trucks by comprehensively considering factors such as truck GPS trajectory information, weather information, open-pit mine road network information and periodicity, which can achieve more accurate and stable travel time prediction compared to current existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flow chart of the steps of the method of the present invention.

[0080] Figure 2 It is a diagram of the architecture of the method of the present invention.

[0081] Figure 3 It is a schematic diagram of the attention-based graph convolution module of the present invention. DETAILED DESCRIPTION

[0082] The following is a preferred embodiment of the present invention and is combined with the accompanying drawings to further describe the technical solution of the present invention, but the present invention is not limited to this embodiment. Therefore, all changes made according to the shape and principle of the present invention should be included in the protection scope of the present invention.

[0083] This paper proposes an attention-based spatiotemporal graph convolution open-pit mine truck travel time prediction method using an end-to-end learning approach;

[0084] Firstly, the original track data of open-pit mine trucks and the original road network data are processed, and combined with the attribute information of external factors, the track data of open-pit mine trucks and the road network data of open-pit mines with attribute information are obtained;

[0085] Secondly, the factor attribute information in the open-pit mine truck trajectory is extracted to obtain the embedded representation of the attribute features;

[0086] Thirdly, the traffic flow information characteristics of the time dimension and the space dimension are captured simultaneously through the spatiotemporal graph convolutional network based on the attention mechanism; the traffic flow information is the recent, daily and weekly historical traffic flow information;

[0087] Then, a convolutional network is used to capture the feature information of neighboring road segments;

[0088] Secondly, the three parts of feature information are connected and the model is trained according to the loss function;

[0089] Finally, the trajectory to be predicted is input, and the final travel time prediction result is obtained based on the trained model.

[0090] like Figure 1 As shown, the specific steps are as follows:

[0091] Step S1: sampling the position data of open-pit mine trucks to form the original trajectory data of open-pit mine trucks, extracting, cleaning and denoising the original trajectory data, and combining the attribute information such as weather and shifts to establish the trajectory data of open-pit mine trucks with attribute characteristics; correcting the existing road network data through the existing road network generation algorithm to obtain the data of the open-pit mine road network with attribute characteristics;

[0092] Step S2: extracting the attribute information of factors affecting truck travel time, and converting the information into trajectory attribute features by using a word vector embedding method; based on the open-pit mine truck trajectory data and open-pit mine road network data obtained in step S1, using the existing map matching method, obtaining the original route information of each trajectory in the road network; combining the original route information with the trajectory attribute features to obtain the attribute route information;

[0093] Step S3: Based on the open pit truck trajectory data and the open pit road network data obtained in step S1, a traffic flow characteristic map is constructed, and the traffic flow characteristic map is input into a traffic flow feature extraction module to capture recent, daily and weekly traffic flow information features;

[0094] Step S4: Based on the original route information of each trajectory in the road network obtained in step S2, input it into the adjacent road segment feature extraction module to capture the feature information of the adjacent road segment;

[0095] Step S5: Establish a multi-task learning module; connect the attribute route information, traffic flow information features and adjacent road segment features obtained in steps S2-S4, and input them into the prediction model, design a multi-task loss function, combine the known trajectory and road network data set, and train the model for truck travel time prediction;

[0096] Step S6: Finally, based on the real-time truck trajectory obtained in the actual open-pit mine production, the model obtained in step S5 is used to predict the truck travel time.

[0097] Furthermore, in step S1:

[0098] The original trajectory data T of the open pit truck is defined as: T = {T1, T2, ..., T i , ..., T N} is a series of raw track data of open pit trucks, N represents the number of tracks; T in T i Expressed as in represents the jth trajectory point in the i-th trajectory, Includes the time, longitude and latitude of the trajectory point; k is the number of trajectory points in the i-th trajectory.

[0099] The open pit truck trajectory data T′ is defined as: T′={T′1, T′2, ..., T′ i , ..., T′ N} is a series of open pit truck trajectory data, where N represents the number of trajectories; Represents the trajectory T i And its four factor attribute information. Indicates the travel distance of the changed trajectory, Indicates weather conditions, Indicates vehicle schedule information. Indicates driver.

[0100] The open pit mine network G is defined as: G = (V, E) is an undirected graph, V and E represent the nodes and edges in the graph respectively; V = {v1, v2, ..., v l , ..., v m}, where v l is a road segment in the road network; e ij =(v i , v j )∈E represents the road segment v i and v j Connect or not.

[0101] A real data set of truck transportation in an open-pit mine was collected. The data set collected data from 185 trucks during operation. The GPS sampling frequency was about 5 seconds, and the data collection period was 2 months. The truck was considered to start a track when it was in the unloading area or loading area. First, the noise data generated by the GPS device was removed. When the vehicle was in the mine, a large amount of stop data was generated. These data were appropriately removed to reduce the amount of data. These data were divided into training data set, validation data set, and test data set in a ratio of 2:1:1. Among them, the training data set was used for model training, the validation data set was used for parameter calibration, and the test data set was used for performance verification.

[0102] Further, in step S2:

[0103] The travel time of open pit trucks is affected by multiple factor attribute information, which are all in the open pit truck trajectory data of step S1. The factor attribute information includes: weather, distance, driver and shift; the factor attribute information is a classification value, and each classification attribute is converted into a low-dimensional real vector using the traditional word vector embedding method as a trajectory attribute feature. Through the open pit truck trajectory data and the open pit road network data, the original route information of each trajectory in the road network is obtained using the existing map matching method. The original route information is a series of road sections in the open pit road network; a route R is defined as: Where R d The number of sections in the route. Assign the attribute information of the trajectory to each node in the original route to obtain the attribute route of the trajectory. The attribute route information of section l at time t can be expressed as The formula is as follows:

[0104]

[0105] Where W e represents the weight of the fully connected layer, A dis Indicates the travel distance (Distance), A wea Indicates weather conditions (Weather), A duty Indicates the vehicle schedule (Duty), A dri Indicates the driver (DriverID), symbol Represents a connection between attributes.

[0106] Furthermore, in step S3:

[0107] The vehicle flow characteristic diagram is defined as: represents the characteristics of all nodes at time t; represents the characteristic value of node h at time t; the structure of the traffic flow characteristic graph is the same as that of the road network, and the node information is represented as the traffic flow characteristic value. Ψ=(ψ1,ψ2,...,ψ t , ..., ψ ξ ) represents a set of traffic flow characteristic graphs, and ξ represents the number of traffic flow characteristic graphs.

[0108] The vehicle flow feature extraction module is a spatiotemporal graph convolutional network based on the attention mechanism. The process is as follows: First, a vehicle flow feature graph is constructed based on the open-pit mine truck trajectory data. Then, the spatiotemporal feature extraction component is used to comprehensively extract the vehicle flow features from the time dimension and the spatial dimension. The component consists of a spatial feature extraction unit and a temporal feature extraction unit. The spatial feature extraction unit is used to extract the dynamic vehicle flow features of the spatial dimension of each road section in the vehicle flow information, and the temporal feature extraction unit is used to extract the features of the temporal dimension of each road section. Next, three independent spatiotemporal feature extraction components are used to capture the recent, daily and weekly historical vehicle flow information respectively, and finally a fully connected layer is used to merge and connect the data generated by the above three components.

[0109] The specific process of generating the traffic flow characteristic map is as follows: the average speed and traffic volume of passing vehicles on each road section are calculated based on the open pit truck trajectory data, and used as the traffic flow characteristics. In this embodiment, the time interval for calculating the average speed and traffic volume is 5 minutes. Combined with different time nodes, the recent, daily and weekly historical traffic flow characteristic map data are obtained from the traffic flow characteristic map. Specifically, the recent traffic flow characteristic map data is The traffic flow characteristic diagram data of the daily cycle is The traffic flow characteristic diagram data of the weekly cycle is Where t0 is the current time, f is the time interval for generating the traffic flow characteristic graph, and q is the number of traffic flow characteristic graphs generated every day. p =kf (k=1, 2, ..., n) is the time segment. b T p The number of traffic flow characteristic graphs in the time series segment. b F is the traffic flow feature vector.

[0110] The spatial feature extraction unit is a convolutional network with attention. The convolutional network uses a spectral domain graph convolution calculation method, which can effectively capture the correlation of neighboring nodes in the traffic flow feature graph. First, the Laplace matrix of the traffic flow feature graph is constructed. Then, the Laplace operator in the frequency domain is derived using the Laplace matrix, and then the convolution in the Euclidean space in the frequency domain is analogized to finally obtain the formula for graph convolution.

[0111] The eigenvalue decomposition of the Laplace matrix is ​​L = U(diag(A))U T , where diag(A) is the diagonal matrix of A, U is the Fourier basis, and L is the eigenvector matrix of the normalized Laplace.

[0112] Formula (2) is the graph convolution operation formula:

[0113] Θ*Gψ=Θ(L)ψ=Θ(U(diag(A))U T )ψ=UΘ(diag(A))U T ψ (2)

[0114] Among them, *G represents the graph convolution operation, Θ is the convolution kernel, and ψ represents the traffic flow feature map.

[0115] Formula (2) is to perform Fourier transform on Θ and ψ respectively, then multiply their transformed results, and then perform inverse Fourier transform to finally get the result of convolution operation. The cost of directly calculating the equation by graph convolution method is extremely high, with a complexity of O(n 2 ). When the scale of the graph is large, the computational efficiency will be greatly reduced. Chebyshev polynomials can be used to approximately but effectively solve this problem.

[0116] Update itself through k-1 neighbor nodes, the formula is as follows:

[0117]

[0118]

[0119] T k (ψ)=2ψT k-1 (ψ)-T k-2 (ψ) (5)

[0120] Among them, θ k are the coefficients of the polynomial vector, I N is the identity matrix, λ max is the largest eigenvalue of the Laplacian matrix, is a Chebyshev polynomial, and its derivation is shown in formula (5), where the initial terms T0(ψ)=1 and T2(ψ)=ψ.

[0121] In actual road traffic scenarios, the road conditions at different locations affect each other, and this mutual influence changes dynamically. In order to make full use of this feature, an attention mechanism is used to adaptively capture the dynamic correlation between spatial nodes. The attention mechanism formula is as follows:

[0122] S=V s tanh((ΨW1)W2(W3Ψ)T +b s ) (6)

[0123]

[0124] in, Represents graph input, C r is the number of channels of the graph input data, T r is the length of the time dimension of the traffic flow feature graph, tanh is the activation function, W1, W2, W3, V s and b s is a learnable parameter. The element S in S i，j The value of represents the correlation strength between node i and node j in semantics. Softmax represents the normalization process, and finally the attention matrix S′ is obtained. During the operation, the adjacency matrix A and the attention matrix S′ are used to dynamically adjust the influence weights between nodes.

[0125] The attention mechanism allows the network to automatically pay more attention to valuable information. It allows the spatiotemporal graph convolution to update node information more effectively. In order to dynamically update the value of each node, the attention matrix S′ is added to each item in the Chebyshev formula to obtain the attention-based update equation for each node:

[0126]

[0127] The time feature extraction unit is a convolutional network that uses a common network to extract traffic flow features and updates the feature values ​​of its own nodes by capturing information at adjacent time slices. For the convolution of the time dimension, the input value is According to formula (8) and the convolution formula, the calculation formulas are shown in (9) to (11):

[0128]

[0129]

[0130]

[0131] From the above formula we can get the recent traffic flow characteristic matrix ST h , the traffic flow characteristic matrix ST of the daily cycle d And the weekly traffic flow characteristic matrix ST w Since the characteristic matrices of the three time periods have different influences on the predicted traffic flow characteristics, it is necessary to add the parameter W before the three matrices. h , W d and W w , and the final traffic flow characteristic matrix is ​​obtained:

[0132]

[0133] Further, in step S4:

[0134] In the route R of the open-pit mine, the characteristic information of the adjacent sections of the route R plays an important role in estimating the travel time of the truck; the adjacent sections are defined as SP l =[v l-w , v l-w+1 , ..., v l-1 , v l , v l+1 , ..., v l+w ], indicating v l The neighboring road segments are represented by w, where w is the context window size; the feature information of the neighboring road segments includes the width of the road, the angle of the corner, and the slope of the road segment. The traditional CNN convolutional layer is used to capture the context information, and the formula is as follows:

[0135]

[0136] Among them, σ represents the fully connected layer calculation, and Emb is the representation layer.

[0137] Further, in step S5:

[0138] The purpose of the multi-task learning module is to aggregate the previously obtained features and then predict the travel time. First, a connection layer is used to aggregate the attribute information features Traffic flow characteristics and the static characteristics of the road network of adjacent road segments Multiple fully connected layers are then used to predict the travel time for each segment of a given route.

[0139] The travel time for a given path Q is calculated as follows:

[0140]

[0141] MLP stands for multiple fully connected layers. represents the predicted travel time of segment l of path Q, and then the travel time of each segment is added up to get the travel time of the entire path.

[0142]

[0143] Where m is the number of sections contained in the given path Q, represents the travel time to be predicted;

[0144] The multi-task loss function obtains the travel time of each section and the travel time of the entire path through the fully connected layer. During the training process, the mean absolute percentage error is selected as the loss function for the estimation of the travel time of the entire path and the section. The mean absolute percentage error is expressed in English as MAPE; a path loss function is defined as:

[0145]

[0146] in, is the predicted travel time of path i, Y i is the actual travel time of path i;

[0147] The segment loss function of a path is defined as:

[0148]

[0149] Among them, m is the number of road segments contained in the i-th path, is the predicted travel time of the jth segment in the i-th path, t j is the actual travel time of the jth segment in the i-th path;

[0150] The loss functions of the path and segment are combined into:

[0151]

[0152] Finally, the final model is obtained by training based on the multi-task loss function.

[0153] Furthermore, in step S6, actual truck trajectory data of open-pit mine production is input, and then the trained model is used to predict the truck travel time.

Claims

1. A method for predicting truck travel time in open-pit mines based on attention-based spatiotemporal graph convolution, characterized by: The specific steps are as follows: Step S1: sampling the position data of the open-pit mine truck to form the original trajectory data of the open-pit mine truck, extracting, cleaning and denoising the original trajectory data, and combining the weather and shift attribute information to establish the open-pit mine truck trajectory data with attribute characteristics; Correct the existing road network data through the existing road network generation algorithm to obtain the open-pit mine road network data with attribute characteristics; Step S2: extracting the attribute information of factors affecting truck travel time, and converting the information into trajectory attribute features using a word vector embedding method; Based on the open pit truck trajectory data and the open pit road network data obtained in step S1, the original route information of each trajectory in the road network is obtained by using the existing map matching method; The original route information is combined with the trajectory attribute features to obtain the attribute route information; Step S3: Based on the open pit truck trajectory data and the open pit road network data obtained in step S1, a traffic flow characteristic map is constructed, and the traffic flow characteristic map is input into a traffic flow feature extraction module to capture recent, daily and weekly traffic flow information features; Step S4: Based on the original route information of each trajectory in the road network obtained in step S2, input it into the adjacent road segment feature extraction module to capture the feature information of the adjacent road segment; Step S5: Establish a multi-task learning module; The attribute route information, traffic flow information features and adjacent road segment features obtained in steps S2-S4 are connected and input into the prediction model, a multi-task loss function is designed, and a model is trained to predict truck travel time by combining known trajectories and road network datasets; Step S6: Finally, according to the real-time truck trajectory obtained in the actual open-pit mine production, the model obtained in step S5 is used to predict the truck travel time.

2. The method for predicting open pit truck travel time based on attention-based spatiotemporal graph convolution according to claim 1 is characterized in that: in step S1: The original trajectory data T of the open pit truck is defined as: T = {T1, T2, ..., T i , ..., T N } is a series of raw track data of open pit trucks, N represents the number of tracks; T in T i Represented as {T i 1 , T i 2 , ..., T i j , ..., T i k },in represents the jth trajectory point in the i-th trajectory, Includes the time, longitude and latitude of the trajectory point; k is the number of trajectory points in the i-th trajectory; The open pit truck trajectory data T′ is defined as: T′={T′1, T′2, ..., T′ i , ..., T′ N } is a series of open pit truck trajectory data, where N represents the number of trajectories; Represents the trajectory T i and its four factor attribute information; Indicates the travel distance of the changed trajectory, Indicates weather conditions. Indicates vehicle schedule information. It means driver; The open pit mine network G is defined as: G = (V, E) is an undirected graph, V and E represent the nodes and edges in the graph respectively; V = {v1, v2, ..., v l , ..., v m }, where v l is a road segment in the road network; e ij =(v i , v j )∈E represents the road segment v i and v j Connect or not.

3. The method for predicting open pit truck travel time based on attention-based spatiotemporal graph convolution according to claim 1, Its characteristics are: in step S2: The travel time of open pit trucks is affected by multiple factor attribute information, which are all in the open pit truck trajectory data of step S1. The factor attribute information includes: weather, distance, driver and shift; the factor attribute information is a classification value, and each classification attribute is converted into a low-dimensional real vector using a traditional word vector embedding method as a trajectory attribute feature; through the open pit truck trajectory data and the open pit road network data, the original route information of each trajectory in the road network is obtained using the existing map matching method; the original route information is a series of road sections in the open pit road network; a route R is defined as: Where R d The number of sections in the route; assign the attribute information of the trajectory to each node in the original route to obtain the attribute route of the trajectory; the attribute route information of section l at time t can be expressed as The formula is as follows: Where W e represents the weight of the fully connected layer, A dis Indicates the travel distance, A wea Indicates weather conditions, A duty Indicates vehicle schedule, A dri Indicates driver, symbol Represents a connection between attributes.

4. The method for predicting open pit truck travel time based on attention-based spatiotemporal graph convolution according to claim 1 is characterized by: In step S3: The vehicle flow characteristic diagram is defined as: represents the characteristics of all nodes at time t; represents the characteristic value of node h at time t; the structure of the traffic flow characteristic graph is the same as that of the road network, and the node information is represented as the traffic flow characteristic value; Ψ = (ψ1, ψ2, ..., ψ t , ..., ψ ξ ) represents a set of traffic flow characteristic graphs, ξ represents the number of traffic flow characteristic graphs; The vehicle flow feature extraction module is a spatiotemporal graph convolutional network based on the attention mechanism; the process is as follows: first, a vehicle flow feature graph is constructed based on the open-pit mine truck trajectory data; then, the vehicle flow features are comprehensively extracted from the time dimension and the spatial dimension through the spatiotemporal feature extraction component, which is composed of a spatial feature extraction unit and a temporal feature extraction unit, wherein the spatial feature extraction unit is used to extract the dynamic vehicle flow features of the spatial dimension of each road section in the vehicle flow information, and the temporal feature extraction unit is used to extract the features of the temporal dimension of each road section; then, three independent spatiotemporal feature extraction components are used to capture the recent, daily cycle and weekly cycle historical vehicle flow information respectively, and finally the data generated by the above three components are merged and connected using a fully connected layer; The specific process of generating the traffic flow characteristic map is as follows: the average speed and traffic flow of passing vehicles on each road section are calculated based on the open pit truck trajectory data, and used as traffic flow characteristics; combining different time nodes, the recent, daily and weekly historical traffic flow characteristic map data are obtained from the traffic flow characteristic map; specifically, the recent traffic flow characteristic map data is The traffic flow characteristic diagram data of the daily cycle is The traffic flow characteristic diagram data of the weekly cycle is Where t0 is the current time, f is the time interval for generating the traffic flow characteristic graph, and q is the number of traffic flow characteristic graphs generated every day; T p =kf (k=1, 2, ..., n) is the time segment; T b T p The number of traffic flow characteristic graphs in the time series segment; the intercept length along the time axis is T b The three time series segments; F is the traffic flow feature vector; The spatial feature extraction unit is a convolutional network with attention. The convolutional network adopts the spectrum domain graph convolution calculation method, which can effectively capture the correlation of adjacent nodes of the traffic flow feature graph. First, the Laplacian matrix of the traffic flow feature graph is constructed. Then, the Laplacian operator in the frequency domain is derived using the Laplacian matrix, and then the convolution in the Euclidean space in the frequency domain is analogized to finally obtain the formula of graph convolution. The eigenvalue decomposition of the Laplace matrix is ​​L = U(diag(A))U T , where diag(A) is the diagonal matrix of A, U is the Fourier basis, and L is the eigenvector matrix of the normalized Laplace; Formula (2) is the graph convolution operation formula: Θ*Gψ=Θ(L)ψ=Θ(U(diag(A))U T )ψ=UΘ(diag(A))U T ψ(2) Among them, *G represents the graph convolution operation, Θ is the convolution kernel, and ψ represents the traffic flow feature map; Formula (2) is to perform Fourier transform on Θ and ψ respectively, then multiply their transformed results, and then perform inverse Fourier transform to finally get the result of convolution operation; the cost of directly calculating the equation by graph convolution method is extremely high, with a complexity of O(n 2 ); When the scale of the graph is large, the computational efficiency will be greatly reduced; Chebyshev polynomials can be used to approximately but effectively solve this problem; Update itself through k-1 neighbor nodes, the formula is as follows: T k (ψ)=2ψT k-1 (ψ)-T k-2 (ψ)(5) Among them, θ k are the coefficients of the polynomial vector, I N is the identity matrix, λ max is the largest eigenvalue of the Laplacian matrix, is a Chebyshev polynomial, and its derivation is shown in formula (5), where the initial terms T0(ψ) = 1, T2(ψ) = ψ; In actual road traffic scenarios, the road conditions at different locations influence each other, and this mutual influence changes dynamically. In order to make full use of this feature, an attention mechanism is used to adaptively capture the dynamic correlation between spatial nodes. The attention mechanism formula is as follows: S=V s ·tanh((ΨW1)W2(W3Ψ) T +b s )(6) in, Represents graph input, C r is the number of channels of the graph input data, T r is the length of the time dimension of the traffic flow feature graph, tanh is the activation function, W1, W2, W3, V s and b s is a learnable parameter; the element S in S i，j The value of represents the correlation strength between node i and node j semantically; softmax represents normalization processing, and finally the attention matrix S′ is obtained; during the operation, the adjacency matrix A and the attention matrix S′ are used to dynamically adjust the influence weights between nodes; The attention mechanism allows the network to automatically pay more attention to valuable information; it allows the spatiotemporal graph convolution to update node information more effectively; in order to dynamically update the value of each node, the attention matrix S′ is added to each item in the Chebyshev formula to obtain the attention-based update equation for each node: The time feature extraction unit is a convolutional network, which uses a common network to extract traffic flow features and updates the feature values ​​of its own nodes by capturing information at adjacent time slices. For the convolution of the time dimension, the input value is According to formula (8) and the convolution formula, the calculation formulas are shown in (9) to (11): From the above formula we can get the recent traffic flow characteristic matrix ST h , the traffic flow characteristic matrix ST of the daily cycle d And the weekly traffic flow characteristic matrix ST w ; Since the characteristic matrices of the three time periods have different influences on the predicted traffic flow characteristics, it is necessary to add the parameter W before the three matrices h , W d and W w , and the final traffic flow characteristic matrix is ​​obtained:

5. The method for predicting open pit truck travel time based on attention-based spatiotemporal graph convolution according to claim 1, characterized in that: in step S4: In the route R of the open-pit mine, the characteristic information of the adjacent sections of the route R plays an important role in estimating the travel time of the truck; the adjacent sections are defined as SP l =[v l-w , v l-w+1 , ..., v l-1 , v l , v l+1 , ..., v l+w ], indicating v l The neighboring road segments are represented by w, where w is the context window size; the feature information of the neighboring road segments includes the width of the road, the angle of the corner, and the slope of the road segment; the traditional CNN convolutional layer is used to capture the feature information of the neighboring road segments, and the formula is as follows: in, σ represents the calculation of the fully connected layer, and Emb is the representation layer.

6. The attention-based spatiotemporal graph convolution method for open-pit mine truck travel time prediction according to claim 1, Its characteristics are: in step S5: The purpose of the multi-task learning module is to aggregate the previously obtained features and then predict the travel time; first, a connection layer is used to aggregate the attribute information features Traffic flow characteristics and the static characteristics of the road network of adjacent road segments Then use multiple fully connected layers to predict the travel time for each segment of a given route; The travel time for a given path Q is calculated as follows: MLP stands for multiple fully connected layers. represents the predicted travel time of segment l of path Q, and then the travel time of each segment is added to get the travel time of the entire path; Among them, m is the number of sections contained in the given path Q, represents the travel time to be predicted; The multi-task loss function obtains the travel time of each section and the travel time of the entire path through the fully connected layer. During the training process, the mean absolute percentage error is selected as the loss function for the estimation of the travel time of the entire path and the section. The mean absolute percentage error is expressed in English as MAPE; a path loss function is defined as: in, is the predicted travel time of path i, Y i is the actual travel time of path i; The segment loss function of a path is defined as: Among them, m is the number of road segments contained in the i-th path, is the predicted travel time of the jth segment in the i-th path, t j is the actual travel time of the jth segment in the i-th path; The loss functions of the path and segment are combined into: Finally, the final model is trained according to the multi-task loss function.

Images