A Multi-View Dynamic Spatiotemporal Semantic Embedding Method for Predicting Truck Transportation Time in Open-Pit Mines

Abstract

A multi-view dynamic spatiotemporal semantic embedding method for predicting transportation time of open-pit mine trucks is presented, belonging to the field of open-pit mine transportation. The method involves preprocessing the collected raw GPS trajectory data of open-pit mine trucks to form a gridded trajectory sequence; extracting the velocity and structural features of the gridded trajectory sequence through a velocity feature extraction module and a grid structure feature extraction module; embedding the semantic information features through a semantic information embedding module; and then extracting the contextual spatiotemporal embedding of the gridded trajectory sequence using a contextual spatiotemporal embedding extraction module to train the model. The prediction results are obtained by inputting the start and end points to be predicted into the trained prediction model. Advantages: The method fully extracts the dynamic spatiotemporal information of the trajectory data and utilizes velocity and semantic information, improving the accuracy of time prediction, which is beneficial for the scheduling of open-pit mine trucks and improves transportation efficiency.

Description

Technical Field

[0001] This invention belongs to the field of open-pit mine transportation and relates to a method for predicting the transportation time of open-pit mine trucks using multi-view dynamic spatiotemporal semantic embedding. Background Technology

[0002] Current methods for predicting transportation time mainly fall into two categories: one is traditional statistical methods, such as linear regression and autoregressive moving averages, but these methods cannot accurately predict transportation times with nonlinear characteristics; the other is machine learning methods such as fuzzy inference systems, support vector machines, and neural networks. Fuzzy inference systems lack self-learning capabilities, and support vector machines cannot handle large-scale data well. Currently, neural networks perform well in nonlinearity, local minima, and handling large-scale data, and are widely used. However, time prediction for a given trajectory plays an indispensable role in intelligent transportation systems.

[0003] In open-pit mines, the main type of vehicle is the truck. These trucks are responsible for transporting excavated soil. When there are many trucks, a method is needed to schedule them effectively to prevent congestion and inefficiency in transporting excavated soil. Therefore, predicting truck transport time in open-pit mines can optimize truck scheduling and improve transport efficiency.

[0004] Predicting transport time for open-pit mine trucks presents the following challenges:

[0005] (1) Due to the weak GPS signal in some areas of the open-pit mine, the GPS trajectory data generated in these areas is too sparse or even non-existent, resulting in a large number of missing speed values ​​when extracting short-term speed features.

[0006] (2) The grid structure of open-pit mines is relatively simple. If grid features are extracted based solely on grid connectivity, it is difficult to obtain accurate grid feature representations. Furthermore, existing methods only focus on the spatial features of the grid while ignoring temporal correlation. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting the transportation time of open-pit mine trucks using multi-view dynamic spatiotemporal semantic embedding, which solves the problems of traditional graph networks not fully utilizing the temporal and spatial information of the grid and the large amount of missing speed values ​​when extracting short-term speed features.

[0008] The objective of this invention is achieved by employing a grid-based trajectory learning method; through a grid velocity feature extraction module, a grid structure feature extraction module, a semantic information embedding module, and a contextual spatiotemporal embedding extraction module, the prediction of open-pit mine truck transportation time is realized.

[0009] First, the raw GPS trajectory data of open-pit mining trucks is preprocessed to reduce the impact of outliers and extract trajectory sequences. The preprocessed trajectory data is then divided into regional grids, and the trajectory sequences are gridded to form gridded trajectory sequences. Based on the average speed of the original trajectory sequence within each grid, a gridded trajectory sequence with speed attributes is obtained.

[0010] Secondly, the velocity features of the gridded trajectory sequence are extracted using the grid velocity feature extraction module.

[0011] Secondly, the long-term historical trajectory data is divided according to time intervals to construct multi-view grid adjacency matrices for different time periods, forming a multi-view spatiotemporal network; the DynGEM module is used to extract the grid structure feature matrix with dynamic transportation information, and combined with the trajectory sequence to obtain the structural features of the corresponding gridded trajectory sequence.

[0012] Then, the semantic information in the gridded trajectory sequence of the open-pit mine truck is embedded into the semantic information embedding module to obtain the embedded representation of the semantic information features; the velocity features, structural features and semantic information features of the gridded trajectory sequence are extracted into the contextual spatiotemporal embedding extraction module, which is composed of GL-Transformer network units and BiLSTM network units.

[0013] Furthermore, the contextual spatiotemporal embedding of the final gridded trajectory sequence is combined with the transportation time in the sample using an MLP prediction model, and a loss function is designed to train the model; the MLP is represented as a multilayer perceptron.

[0014] Finally, based on the start and end points of the transportation time to be predicted and combined with grid information, a path planning method is designed to obtain several candidate gridded trajectory sequences. The contextual spatiotemporal embedding of each gridded trajectory sequence is generated and input into the trained MLP prediction model to obtain the transportation time prediction result.

[0015] The specific steps are as follows:

[0016] Step S1: Sample the location data of the open-pit mining truck to form the original GPS trajectory data of the open-pit mining truck. Clean the original GPS trajectory data by writing a program. Then, serialize the GPS trajectory data by writing a trajectory sequence extraction program. Divide the processed trajectory data into regions to obtain a region grid. Map the trajectory sequence to the corresponding region grid and calculate the average speed of the trajectory sequence in each grid to obtain a gridded trajectory sequence with speed attributes.

[0017] Step S2: Obtain short-term velocity matrix and long-term velocity matrix from the region mesh obtained in step S1 according to different time intervals; recover the missing velocity values ​​in the velocity matrix through the non-negative tensor decomposition module; and then extract the velocity features of the meshed trajectory sequence through the mesh velocity feature extraction module.

[0018] Step S3: Segment the preprocessed trajectory data according to time, and fill the trajectory data of different time periods into the grid. Construct a dynamic multi-view spatiotemporal network of the region grid based on the grid's connectivity, connection length, and similarity. Input the dynamic multi-view spatiotemporal network into the DynGEM module to extract the structural feature matrix under the gridded trajectory multi-view. Finally, perform a weighted summation of the multi-view structural feature matrix to obtain the structural features of the gridded trajectory sequence.

[0019] Step S4: The semantic information in the gridded trajectory sequence obtained in step S1 is processed through the semantic information embedding module to obtain the embedded representation of the semantic information features; the velocity features, structural features and semantic information features of the gridded trajectory sequence are processed through the context spatiotemporal embedding extraction module to extract the context spatiotemporal embedding of the gridded trajectory sequence, wherein the context spatiotemporal embedding extraction module is composed of GL-Transformer network units and BiLSTM network units.

[0020] Step S5: The contextual spatiotemporal embedding of the gridded trajectory sequence obtained in Step S4 is combined with the transportation time in the sample using an MLP prediction model, and a loss function is designed to train the model.

[0021] Step S6: Based on the starting point and ending point of the transportation time to be predicted, and combined with grid information, design a path planning method to obtain several candidate gridded trajectory sequences, and generate the contextual spatiotemporal embedding of each gridded trajectory sequence. Use the MLP prediction model trained in step S5 to obtain the transportation time prediction result.

[0022] The implementation process of step S1 is as follows:

[0023] The trajectory cleaning procedure filters out duplicate and abnormal data from the original GPS trajectory data; the trajectory sequence extraction procedure divides the GPS trajectory data of the same driver into segments according to a sliding time window with a set time of T and a step size of n GPS trajectory points to complete the extraction of the trajectory sequence.

[0024] The specific operation of the region grid division is as follows: First, find the maximum and minimum values ​​of the latitude and longitude of all trajectory points to form the latitude and longitude coordinates [lat1, lng1] and [lat2, lng2] of the upper left and lower right of the region; divide the region into N×N grids, each two-dimensional grid has an ID, namely lat_id and lng_id, where lat_id represents the latitude ID of the grid and lng_id represents the longitude ID of the grid; calculate the length of each grid: lat_len = (lat2 - lat1) / (N - 1 Given the latitude and longitude information lat, lng of each GPS track point, we can map lat_id = (lat-lat1) / lat_len and lng_id = (lng-lng1) / lng_len to obtain the two-dimensional grid ID corresponding to the latitude and longitude of the track point. Furthermore, we sort the two-dimensional grid IDs by row and number the grids sequentially from 1. The number of each grid is the one-dimensional grid ID of that grid.

[0025] The trajectory sequence is: trajectory data It is a set of trajectory data, where M represents the number of trajectory sequences; Tr i ∈Tr (1≤i≤M) represents a trajectory sequence, denoted as in Represents the j-th GPS trajectory point in the i-th trajectory sequence, i R This represents the number of GPS trajectory points in the i-th trajectory sequence; each GPS trajectory point

[0026] in and for Latitude and longitude of the trajectory point for The grid number of the trajectory point for

[0027] The velocity of the trajectory point and for The timestamps of the trajectory points; in addition, each trajectory contains other additional attributes:

[0028] represent Driver information represent Travel time, represent travel distance, represent Weekly information represent Number of stays and represent The length of stay.

[0029] The gridded trajectory sequence is obtained by merging multiple trajectory points within a grid according to the above GPS trajectory sequence, retaining the information of the merged trajectory points. Among them, grid and corresponding GPS track points The content information is consistent.

[0030] The method for merging multiple trajectory points within a grid is as follows: If a trajectory sequence contains multiple trajectory points within the same grid, these multiple trajectory points are merged into one trajectory point. The merging method is as follows: Calculate the distance from these multiple trajectory points to the center of the grid, select the trajectory point with the shortest distance as the trajectory point within the merged grid, and calculate the velocity of the trajectory sequence in the corresponding grid using the timestamp of the trajectory sequence and the latitude and longitude of the GPS trajectory points to obtain a gridded trajectory sequence with velocity attributes.

[0031] The implementation process of step S2 is as follows:

[0032] Step S2-1: Construct the short-term velocity matrix A using the region grid obtained in step S1. r Long-term velocity matrix A h Fusion velocity matrix A m Short-term velocity matrix To store tt s The short-term travel speed from hour to t, where M represents the number of time intervals, then t represents the speed of each time interval. s / M hours, N 2 A represents the number of grid cells in the divided region. r (i,j) = a represents the time period j within the i-th region grid. s The average driving speed is calculated, and if multiple trajectory points exist within the regional grid, their average value is taken. To obtain effective transportation characteristics from historical information, a long-term speed matrix is ​​constructed. That is, matrix To store tt l Given the long-term driving speed from day t to t, then the speed for each time period is t. l / M days, A h (i,j) = b represents the time period j within the i-th region grid. l Average driving speed; due to A r and A h With similar structures, the fusion velocity matrix is ​​constructed. To integrate their information, the specific integration strategy is as follows:

[0033]

[0034] Based on the above three matrices A r A m A h As different channels, they form a three-dimensional nonnegative tensor. Using existing nonnegative tensor decomposition algorithms, the originally sparse tensors are decomposed. Become a dense tensor

[0035] Step S2-2, for each region grid g i Design a triplet to record velocity characteristics: Represents the region grid g i The driving speed during time period j, dis i This represents the distance the trajectory travels across the grid area. A rough estimate of the time taken to traverse this grid is given. This is to capture the region grid g. i The trend of speed change, defining the speed feature sequence. Represents the region grid g i Velocity characteristics; in the velocity characteristic sequence X i One-dimensional convolutional filters using CNNs; velocity feature sequences

[0036] X i ∈R 3×M Consider it as a three-channel input; the j-dimensional output after one-dimensional convolution is:

[0037]

[0038] "*" represents a convolution operation, b is the bias term, and "tanh" is the activation function; the output sequence

[0039] c is the number of one-dimensional convolutional filters, W conv ∈R k×M Let be the parameter matrix, where k is the kernel size.

[0040] Step S2-3: Use a recurrent neural network (RNN) to learn the region grid g. i Time correlation of internal driving speed:

[0041]

[0042] W c and W h is a learnable parameter matrix, and b is a bias term. Represents the hidden state at step t; RNN model output

[0043] for Recorded as Reflects the regional grid g i Short-term speed characteristics.

[0044] Step S2-4, define the upstream mesh g up That is, for a given gridded trajectory sequence

[0045] Among them, grid It's a grid. The upstream grid of the current grid is affected by the velocity characteristics of its upstream grid; find the current grid g. i All upstream grids, and extract the time interval t before the current time interval t. up The velocity distribution of all upstream grids is calculated using the existing KL divergence procedure to determine the grid g. i The similarity between the velocity distribution and the velocity distribution of all upstream grids is used to select the k upstream grids with the highest similarity as grid g. i The relevant region grids are used; short-term velocity features of these k relevant region grids are extracted using CNN and RNN. The CNN is a deep neural network with a convolutional structure, and the RNN is a recurrent neural network with a chain structure.

[0046] Besides short-term speed characteristics, long-term speed characteristics can represent stable driving speed patterns over a long period of time, patterns that short-term speed characteristics cannot represent; extending the time window from [tt]... s hour,t] changed to [tt] l Similar to methods for extracting short-term velocity features, the grid g can be extracted using the method of extracting day,t]. i Long-term velocity characteristics

[0047] Step S2-5: Connect the short-term velocity features and the long-term velocity features using a fully connected layer to obtain the region mesh g. i velocity characteristics

[0048]

[0049] The velocity features of each region grid are connected according to the gridded trajectory sequence to obtain the velocity features V of the overall gridded trajectory sequence. speed .

[0050] The implementation process of step S3 is as follows:

[0051] Step S3-1: Based on the divided region grid, design the grid into an undirected graph G = (V, E) using the method of grids as nodes and the connectivity between gridded trajectories as edges, where V and E represent vertices and edges in the graph, respectively; the node set V represents the grid, and the edge set represents the connectivity between two region grids; here, the undirected graph G is defined. con =Adjacency matrix S of (V,E) con (N 2 ×N 2 ),in Represents the region grid g i and g j Connected?

[0052] Step S3-2, Figure G con Only the connectivity between two region grids was considered, but the length of the connecting edge between the two region grids is also very important; therefore, graph G is added here. len adjacency matrix S len (N 2 ×N 2 ),in Represents the grid g i and g j The reciprocal of the number of grid intervals:

[0053]

[0054] Where len i,j =long i,j +wide i,j , long i,j Represents the region grid g i and g j The number of grid cells in the vertical spacing, wide i,j Represents the region grid g i and g j The number of grid cells in the horizontal spacing.

[0055] Step S3-3 uses the number of trajectory points within a region grid to measure the spatial similarity of a region grid over a period of time. Higher similarity reflects more similar traffic conditions among the region grids over that period. Therefore, graph G is added here. imp adjacency matrix S imp (N 2 ×N 2 ),in Represents the region grid g i and g j Spatial similarity between them:

[0056]

[0057] Among them, M i Represents the region grid g i The number of internal trajectory points, M j Represents the region grid g j The number of points on the inner trajectory.

[0058] Step S3-4: Design a time sliding window w based on historical trajectory data, then divide the historical trajectory data into W segments (T1, T2, ..., T) according to the time of day. K And extract the adjacency matrix from each historical trajectory data segment. and

[0059] The adjacency matrix S mentioned above con S len and S imp It is dynamic, so the DynGEM module is used to handle this; a Deep Belief Network (DBN) is used to pre-train the parameters on the existing SDNE model to obtain the initial parameter set. Adjacency Matrix and The parameters are input into the SDNE model, and after training, the corrected parameter set is obtained. and Using the parameter set from the previous time step and Initialize the current time step parameter group and After training iterations (W-1), multiple region grid map embedding feature matrices can be obtained.

[0060] Steps S3-5: After iterative training, the region grid feature matrix Y is obtained. con Y len and Y imp The final mesh structure feature matrix is ​​obtained by weighted summation:

[0061] Y = W con *Y con +W len *Y len +W imp *Y imp

[0062] Obtain the feature matrix of the grid structure

[0063]

[0064] p i .lat represents GPS track point p i latitude, p i.lon represents GPS track point p i The longitude is obtained by connecting the grid structure features and the latitude and longitude of the original GPS trajectory points through a fully connected layer to obtain the region grid g. i Structural features By connecting the structural features of each region grid according to the gridded trajectory sequence, the velocity feature V of the overall gridded trajectory sequence is obtained. road .

[0065] The implementation process of step S4 is as follows:

[0066] The GL-Transformer network unit is used to extract the spatiotemporal information of the transport of the gridded trajectory sequence. The GL-Transformer network unit consists of two main parts: a global encoder and a local encoder, which ensures that the GL-Transformer network unit has both global and local dependencies. The global encoder consists of four parts: position embedding, multi-head self-attention, feedforward neural network, and normalized-residual connection. The local encoder consists of four parts: position embedding, masked multi-head self-attention, feedforward neural network, and normalized-residual connection. Among these four parts, only the masked multi-head self-attention is different from the multi-head self-attention in the global encoder. The other three parts are the same as the global encoder, so only the masked multi-head self-attention is introduced here.

[0067] Step S4-1: Calculate the attention weight matrix A∈R n×n The calculation formula is as follows:

[0068]

[0069] The attention weight matrix A is used to calculate the mask adjacency matrix M∈R using the existing K-hop algorithm. n×n The mask self-attention is calculated using the mask adjacency matrix, and the calculation formula is as follows:

[0070]

[0071]

[0072]

[0073] Here, Mask(·) is a function used to calculate the mask attention weight matrix. Attention(Q,K,V) It is the self-attention mask obtained through calculation.

[0074] Step S4-2, V speed and V road The velocity feature vector H of the gridded trajectory sequence is obtained through GL-Transformer network units.speed The structural feature vector H of the gridded trajectory sequence road ,in

[0075] Step S4-3, after extracting the velocity and structural features of the gridded trajectory sequence, also needs to consider other factors affecting transportation time prediction, such as travel time (timeID), weekday information (weekID), driver information (driverID), number of stops (DwellNumber), stop time (DwellTime), and spatial distance. The features of these six factors are then embedded and concatenated to form the final semantic information feature V. add ∈R 21 .

[0076] Step S4-4, the above V add Separately and Embedded connections are made and input into the BiLSTM network element:

[0077]

[0078]

[0079]

[0080] The contextual spatiotemporal embedding representation H of the gridded trajectory sequence is obtained. i .

[0081] The implementation process of step S5 is as follows:

[0082] The contextual spatiotemporal embedding of the gridded trajectory sequence is obtained in step S4. i R The number of grid cells in the region representing the i-th gridded trajectory sequence; transforming the transportation time prediction problem into a sequence H i Compared with actual transportation time The regression problem between them.

[0083] Step S5-1, a vector is constructed. Represents the distance from the grid start point to the grid. Actual transportation time; in Used to correspond to sequence H i In Right now It corresponds The nonlinear mapping f(·) of a multilayer perceptron (MLP) is used to predict the transport time of each gridded trajectory sequence from the grid start point to the subsequent grids of that gridded trajectory sequence, i.e.:

[0084]

[0085] In the above formula f(·) represents the predicted transport time from the starting point of the i-th gridded trajectory sequence to each of the other grids; f(·) is a specific nonlinear function or neural network.

[0086] Step S5-2 establishes the transportation time prediction vector. and the actual transportation time vector T i The loss function between them is as follows:

[0087]

[0088] The molecule represents the transportation time prediction vector. With the actual transportation time vector T i The L2 norm of the relative error percentage between them, with the denominator designed as Because of vectors and T i The first element of each element is 0; the model is trained using the loss function described above to obtain the MLP prediction model.

[0089] The implementation process of step S6 is as follows:

[0090] Step S6-1: Map the GPS coordinates of the starting point and ending point of the actual trucks in the open-pit mine to the divided area grid, thereby forming the starting point grid and the ending point grid.

[0091] Step S6-2, the path planning method is designed as follows:

[0092] Using the undirected graph G constructed in step S3 con =Adjacency matrix S of (V,E) con (N 2 ×N 2 Extending this to the entire historical trajectory data yields the grid connectivity matrix S. c ′ on The entire historical trajectory data is then mapped onto a regional grid, yielding the number of GPS trajectory points mapped within each regional grid, denoted as the historical regional grid frequency vector. Where count i This represents the number of GPS trajectory points mapped within region grid i; starting from the starting grid according to the grid connectivity matrix S c ′ onA breadth-first search is performed until the destination grid is reached. If the destination grid is unreachable during the search, a breadth-first search is performed from the destination grid. The unreachable grid nodes from the starting grid and the destination grid are combined and shortest path searches are performed on each group to obtain the predicted grid path. Each grid in the obtained predicted grid path is then used to perform a Breadth-First Search on the V grid. fre Find the corresponding frequencies and add them together to get the total count. Sort the predicted grid paths from largest to smallest according to the total count, find the top k grid paths, output the top k predicted grid paths and generate the contextual spatiotemporal embedding of each gridded trajectory sequence.

[0093] Step S6-3: Input the contextual spatiotemporal embeddings of the first k gridded trajectory sequences predicted, and then use the trained MLP prediction model to predict the transportation time of open-pit mine trucks.

[0094] Beneficial effects: By adopting the above scheme and utilizing a grid-based trajectory learning method, the cumulative error caused by segmented prediction of transportation time is avoided. First, the long-term and short-term velocity features of the gridded trajectory sequence are captured by a velocity feature extraction module based on non-negative tensor decomposition. Next, the structural features of the gridded trajectory sequence are captured by a grid structure feature extraction module based on multi-view dynamic spatiotemporal semantic embedding. Then, semantic information, such as driver information, travel time, travel distance, weekday information, number of stops, and stop time, is processed by a semantic information embedding module to obtain semantic information features. Finally, the velocity features, structural features, and semantic information features of the gridded trajectory sequence are concatenated, and the final transportation time is obtained through a contextual spatiotemporal embedding extraction module for the gridded trajectory sequence. Compared with existing technologies, the above technical solution of this invention has the following significant advantages:

[0095] This invention introduces dynamic spatiotemporal graph embedding into open-pit mine transportation time prediction. Building upon traditional graph embedding, it adds a dynamic graph embedding mechanism that evolves over time. Through time segmentation and the construction of a multi-view dynamic network with grids, it captures the dynamic spatiotemporal information of the grid. Secondly, by using non-negative tensor decomposition, it addresses the problem of missing speed values ​​in short-term speed features. Then, a speed feature extraction module obtains long-term and short-term speed features of the gridded trajectory sequence. Finally, it fully utilizes semantic information from open-pit mines, such as driver information, travel time, travel distance, weekday information, number of stops, and stop duration, to extract features, resulting in a more accurate prediction of open-pit mine truck transportation time. This leads to more accurate open-pit mine truck transportation time prediction, optimized open-pit mine truck scheduling, and improved open-pit mine truck transportation efficiency.

[0096] This addresses the problem that traditional graph networks do not fully utilize the temporal and spatial information of the grid and that velocity values ​​are largely missing when extracting short-term velocity features.

[0097] This invention provides a method for predicting the transportation time of open-pit mine trucks by comprehensively considering factors such as GPS trajectory information, semantic information, and grid information of the open-pit mine area. Compared with existing methods, the method provided by this invention can achieve more stable and accurate prediction of the transportation time of open-pit mine trucks. Attached image description:

[0098] Figure 1 This is a flowchart of the method of the present invention.

[0099] Figure 2 This is a diagram of the method architecture of the present invention. Detailed implementation method:

[0100] The following are preferred embodiments of the present invention, described in conjunction with the accompanying drawings, to further illustrate the technical solution of the present invention. However, the present invention is not limited to these embodiments. Therefore, all variations made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

[0101] This invention provides a method for predicting the transportation time of open-pit mine trucks using multi-view dynamic spatiotemporal semantic embedding. It employs a grid-based trajectory learning method and achieves the prediction of the transportation time of open-pit mine trucks through a grid velocity feature extraction module, a grid structure feature extraction module, a semantic information embedding module, and a contextual spatiotemporal embedding extraction module.

[0102] First, the raw GPS trajectory data of open-pit mining trucks is preprocessed to reduce the impact of outliers and extract trajectory sequences. The preprocessed trajectory data is then divided into regional grids, and the trajectory sequences are gridded to form gridded trajectory sequences. Based on the average speed of the original trajectory sequence within each grid, a gridded trajectory sequence with speed attributes is obtained.

[0103] Secondly, the velocity features of the gridded trajectory sequence are extracted using the grid velocity feature extraction module.

[0104] Secondly, the long-term historical trajectory data is divided according to time intervals to construct multi-view grid adjacency matrices for different time periods, forming a multi-view spatiotemporal network; the DynGEM module is used to extract the grid structure feature matrix with dynamic transportation information, and combined with the trajectory sequence to obtain the structural features of the corresponding gridded trajectory sequence.

[0105] Then, the semantic information in the gridded trajectory sequence of the open-pit mine truck is embedded into the semantic information embedding module to obtain the embedded representation of the semantic information features; the velocity features, structural features and semantic information features of the gridded trajectory sequence are extracted into the contextual spatiotemporal embedding extraction module, which is composed of GL-Transformer network units and BiLSTM network units.

[0106] Furthermore, the contextual spatiotemporal embedding of the final gridded trajectory sequence is combined with the transportation time in the sample using an MLP prediction model, and a loss function is designed to train the model; the MLP is represented as a multilayer perceptron.

[0107] Finally, based on the start and end points of the transportation time to be predicted, combined with grid information, a path planning method is designed to obtain several candidate gridded trajectory sequences. The contextual spatiotemporal embedding of each gridded trajectory sequence is generated and input into the trained MLP prediction model to obtain the transportation time prediction result.

[0108] like Figure 1 As shown, the specific steps are as follows:

[0109] Step S1: Sample the location data of the open-pit mining truck to form the original GPS trajectory data of the open-pit mining truck. Clean the original GPS trajectory data by writing a program. Then, serialize the GPS trajectory data by writing a trajectory sequence extraction program. Divide the processed trajectory data into regions to obtain a region grid. Map the trajectory sequence to the corresponding region grid and calculate the average speed of the trajectory sequence in each grid to obtain a gridded trajectory sequence with speed attributes.

[0110] Step S2: Obtain short-term velocity matrix and long-term velocity matrix from the region mesh obtained in step S1 according to different time intervals; recover the missing velocity values ​​in the velocity matrix through the non-negative tensor decomposition module; and then extract the velocity features of the meshed trajectory sequence through the mesh velocity feature extraction module.

[0111] Step S3: Segment the preprocessed trajectory data according to time, and fill the trajectory data of different time periods into the grid. Construct a dynamic multi-view spatiotemporal network of the region grid based on the grid's connectivity, connection length, and similarity. Input the dynamic multi-view spatiotemporal network into the DynGEM module to extract the structural feature matrix under the gridded trajectory multi-view. Finally, perform a weighted summation of the multi-view structural feature matrix to obtain the structural features of the gridded trajectory sequence.

[0112] Step S4: The semantic information in the gridded trajectory sequence obtained in step S1 is processed through the semantic information embedding module to obtain the embedded representation of the semantic information features; the velocity features, structural features and semantic information features of the gridded trajectory sequence are processed through the context spatiotemporal embedding extraction module to extract the context spatiotemporal embedding of the gridded trajectory sequence, wherein the context spatiotemporal embedding extraction module is composed of GL-Transformer network units and BiLSTM network units.

[0113] Step S5: The contextual spatiotemporal embedding of the gridded trajectory sequence obtained in Step S4 is combined with the transportation time in the sample using an MLP prediction model, and a loss function is designed to train the model.

[0114] Step S6: Based on the starting point and ending point of the transportation time to be predicted, and combined with grid information, design a path planning method to obtain several candidate gridded trajectory sequences, and generate the contextual spatiotemporal embedding of each gridded trajectory sequence. Use the MLP prediction model trained in step S5 to obtain the transportation time prediction result.

[0115] The implementation process of step S1 is as follows:

[0116] The trajectory cleaning procedure filters out duplicate and abnormal data from the original GPS trajectory data; the trajectory sequence extraction procedure divides the GPS trajectory data of the same driver into segments according to a sliding time window with a set time of T and a step size of n GPS trajectory points to complete the extraction of the trajectory sequence.

[0117] The specific operation of the area grid division is as follows: First, find the maximum and minimum values ​​of the latitude and longitude of all trajectory points to form the latitude and longitude coordinates [lat1, lng1] and [lat2, lng2] of the upper left and lower right of the area; divide the area into N×N grids, and each two-dimensional grid id is (lat_id, lng_id), where lat_id represents the latitude id of the two-dimensional grid id and lng_id represents the longitude id of the two-dimensional grid id. By calculating the length of each grid: lat_len = (lat2-lat1) / (N-1) and the width lng_len = (lng2-lng1) / (N-1), the latitude and longitude information (lat, lng) of each GPS track point can be mapped into the grid, i.e., lat_id = (lat-lat1) / lat_len and lng_id = (lng-lng1) / lng_len, to obtain the two-dimensional grid ID corresponding to the latitude and longitude of the track point; further, the two-dimensional grid IDs are sorted by row, and the grids are numbered sequentially from 1 according to the sorted grids, and the number of each grid is the one-dimensional grid ID of that grid.

[0118] The trajectory sequence is: trajectory data It is a set of trajectory data, where M represents the number of trajectory sequences. Tr i ∈Tr (1≤i≤M) represents a trajectory sequence, denoted as in Represents the j-th GPS trajectory point in the i-th trajectory sequence, i R This represents the number of GPS track points in the i-th track sequence. Each GPS track point...

[0119] in and for Latitude and longitude of the trajectory point for The grid number of the trajectory point for The velocity of the trajectory point and for The timestamps of the trajectory points. In addition, each trajectory contains other additional attributes: represent Driver information represent Travel time, represent travel distance, represent Weekly information represent Number of stays and represent The length of stay.

[0120] The gridded trajectory sequence is obtained by merging multiple trajectory points within a grid according to the above GPS trajectory sequence, retaining the information of the merged trajectory points. Among them, grid and corresponding GPS track points The content information is consistent.

[0121] The method for merging multiple trajectory points within a grid is as follows: If a trajectory sequence contains multiple trajectory points within the same grid, these multiple trajectory points are merged into one trajectory point. The merging method is as follows: Calculate the distance from these multiple trajectory points to the center of the grid, select the trajectory point with the shortest distance as the trajectory point within the merged grid, and calculate the velocity of the trajectory sequence in the corresponding grid using the timestamp of the trajectory sequence and the latitude and longitude of the GPS trajectory points to obtain a gridded trajectory sequence with velocity attributes.

[0122] The implementation process of step S2 is as follows:

[0123] Step S2-1: Construct the short-term velocity matrix A using the region grid obtained in step S1. r Long-term velocity matrix A h Fusion velocity matrix A m Short-term velocity matrix To store tt s The short-term travel speed from hour to t, where M represents the number of time intervals, then t represents the speed of each time interval. s / M hours, N 2 A represents the number of grid cells in the divided region. r (i,j) = a represents the time period j within the i-th region grid. s The average driving speed is calculated, and if multiple trajectory points exist within the regional grid, their average value is taken. To obtain effective transportation characteristics from historical information, a long-term speed matrix is ​​constructed. That is, matrix To store tt l Given the long-term driving speed from day t to t, then the speed for each time period is t. l / M days, A h (i,j) = b represents the time period j within the i-th region grid. l Average driving speed; due to A r and A h With similar structures, the fusion velocity matrix is ​​constructed. To integrate their information, the specific integration strategy is as follows:

[0124]

[0125] Based on the above three matrices A r A m A h As different channels, they form a three-dimensional nonnegative tensor. Using existing nonnegative tensor decomposition algorithms, the originally sparse tensors are decomposed. Become a dense tensor

[0126] Step S2-2, for each region grid g i Design a triplet to record velocity characteristics:

[0127]

[0128] Represents the region grid g i The driving speed during time period j, dis i This represents the distance the trajectory travels across the grid area. A rough estimate of the time taken to traverse the grid.

[0129] In order to capture the region grid gi The trend of speed change, defining the speed feature sequence. Represents the region grid g i Velocity characteristics; in the velocity characteristic sequence X i A one-dimensional convolutional filter is applied using a CNN. The velocity feature sequence X i ∈R 3×M Consider it as a three-channel input; the j-dimensional output after one-dimensional convolution is:

[0130]

[0131] "*" represents the convolution operation, b is the bias term, and "tanh" is the activation function. Output sequence c is the number of one-dimensional convolutional filters, W conv ∈R k×M Let be the parameter matrix, where k is the kernel size.

[0132] Step S2-3: Use a recurrent neural network (RNN) to learn the region grid g. i Time correlation of internal driving speed:

[0133]

[0134] W c and W h is a learnable parameter matrix, and b is a bias term. This represents the hidden state at step t. The RNN model output is... Recorded as Reflects the regional grid g i Short-term speed characteristics.

[0135] Step S2-4, define the upstream mesh g up That is, for a given gridded trajectory sequence Among them, grid It's a grid. The current mesh velocity characteristics are influenced by its upstream mesh; find the current mesh g. i All upstream grids, and extract the time interval t before the current time interval t. up The velocity distribution of all upstream grids is calculated using the existing KL divergence procedure to determine the grid g. i The similarity between the velocity distribution and the velocity distribution of all upstream grids is used to select the k upstream grids with the highest similarity as grid g. i The relevant region grids are used; short-term velocity features of these k relevant region grids are extracted using CNN and RNN. The CNN is a deep neural network with a convolutional structure, and the RNN is a recurrent neural network with a chain structure.

[0136] Besides short-term speed characteristics, long-term speed characteristics can represent stable driving speed patterns over a long period of time, patterns that short-term speed characteristics cannot represent; extending the time window from [tt]... s hour,t] changed to [tt] l Similar to methods for extracting short-term velocity features, the grid g can be extracted using the method of extracting day,t]. i Long-term velocity characteristics

[0137] Step S2-5: Connect the short-term velocity features and the long-term velocity features using a fully connected layer to obtain the region mesh g. i velocity characteristics

[0138]

[0139] The velocity features of each region grid are connected according to the gridded trajectory sequence to obtain the velocity features V of the overall gridded trajectory sequence. speed .

[0140] The implementation process of step S3 is as follows:

[0141] Step S3-1: Based on the divided region grid, design the grid into an undirected graph G = (V, E) using the method of grids as nodes and the connectivity between gridded trajectories as edges, where V and E represent vertices and edges in the graph, respectively; the node set V represents the grid, and the edge set represents the connectivity between two region grids; here, the undirected graph G is defined. con =Adjacency matrix S of (V,E) con (N 2 ×N 2 ),in Represents the region grid g i and g j Connected?

[0142]

[0143] Step S3-2, Figure G con Only the connectivity between two region grids was considered, but the length of the connecting edge between the two region grids is also very important; therefore, graph G is added here. len adjacency matrix S len (N 2 ×N 2 ),in Represents the grid g i and gj The reciprocal of the number of grid intervals:

[0144] Where len i,j =long i,j +wide i,j , long i,j Represents the region grid g i and g j The number of grid cells in the vertical spacing, wide i,j Represents the region grid g i and g j The number of grid cells in the horizontal spacing.

[0145] Step S3-3 uses the number of trajectory points within a region grid to measure the spatial similarity of a region grid over a period of time. Higher similarity reflects more similar traffic conditions among the region grids over that period. Therefore, graph G is added here. imp adjacency matrix S imp (N 2 ×N 2 ),in Represents the region grid g i and g j Spatial similarity between them:

[0146]

[0147] Among them, M i Represents the region grid g i The number of internal trajectory points, M j Represents the region grid g j The number of points on the inner trajectory.

[0148] Step S3-4: Design a time sliding window w based on historical trajectory data, then divide the historical trajectory data into W segments (T1, T2, ..., T) according to the time of day. K And extract the adjacency matrix from each historical trajectory data segment.

[0149] and

[0150] The adjacency matrix S mentioned above con S len and S imp It's dynamic, so the DynGEM module is used to handle this. A Deep Belief Network (DBN) is used to pre-train the parameters on the existing SDNE model to obtain the initial parameter set.

[0151] Adjacency Matrix and The parameters are input into the SDNE model, and after training, the corrected parameter set is obtained. and Using the parameter set from the previous time step and Initialize the current time step parameter group and After training iterations (W-1), multiple region grid map embedding feature matrices can be obtained.

[0152] Steps S3-5: After iterative training, the region grid feature matrix Y is obtained. con Y len and Y imp The final mesh structure feature matrix is ​​obtained by weighted summation:

[0153] Y = W con *Y con +W len *Y len +W imp *Y imp

[0154] Obtain the feature matrix of the grid structure

[0155]

[0156] p i .lat represents GPS track point p i latitude, p i .lon represents GPS track point p i The longitude is obtained by connecting the grid structure features and the latitude and longitude of the original GPS trajectory points through a fully connected layer to obtain the region grid g. i Structural features By connecting the structural features of each region grid according to the gridded trajectory sequence, the velocity feature V of the overall gridded trajectory sequence is obtained. road .

[0157] The implementation process of step S4 is as follows:

[0158] The GL-Transformer network unit is used to extract the spatiotemporal information of transport in the gridded trajectory sequence. The GL-Transformer network unit consists of two main parts: a global encoder and a local encoder, ensuring both global and local dependencies. The global encoder comprises four parts: position embedding, multi-head self-attention, a feedforward neural network, and normalized-residual connections. The local encoder also comprises four parts: position embedding, masked multi-head self-attention, a feedforward neural network, and normalized-residual connections.

[0159] Of these four parts, only the mask multi-head self-attention is different from the multi-head self-attention in the global encoder. The other three parts are the same as the global encoder, so only the mask multi-head self-attention will be introduced here.

[0160] Step S4-1: Calculate the attention weight matrix A∈R n×n The calculation formula is as follows:

[0161]

[0162] The attention weight matrix A is used to calculate the mask adjacency matrix M∈R using the existing K-hop algorithm. n×n The mask self-attention is calculated using the mask adjacency matrix, and the calculation formula is as follows:

[0163]

[0164]

[0165]

[0166] Here, Mask(·) is a function used to calculate the mask attention weight matrix. Attention(Q,K,V) It is the self-attention mask obtained through calculation.

[0167] Step S4-2, V speed and V road The velocity feature vector H of the gridded trajectory sequence is obtained through GL-Transformer network units. speed The structural feature vector H of the gridded trajectory sequence road ,in

[0168] Step S4-3, after extracting the velocity and structural features of the gridded trajectory sequence, also needs to consider other factors affecting transportation time prediction, such as travel time (timeID), weekday information (weekID), driver information (driverID), number of stops (DwellNumber), stop time (DwellTime), and spatial distance. The features of these six factors are then embedded and concatenated to form the final semantic information feature V. add ∈R 21 .

[0169] Step S4-4, the above V add Separately and Embedded connections are made and input into the BiLSTM network element:

[0170]

[0171]

[0172]

[0173] The contextual spatiotemporal embedding representation H of the gridded trajectory sequence is obtained. i .

[0174] The implementation process of step S5 is as follows:

[0175] The contextual spatiotemporal embedding of the gridded trajectory sequence is obtained in step S4. i R This represents the number of grid cells in the region of the i-th gridded trajectory sequence. The transportation time prediction problem is transformed into a sequence H... i Compared with actual transportation time The regression problem between them.

[0176] Step S5-1, a vector is constructed. Represents the distance from the grid start point to the grid. Actual transportation time; in Used to correspond to sequence H i In Right now It corresponds The nonlinear mapping f(·) of a multilayer perceptron (MLP) is used to predict the transport time of each gridded trajectory sequence from the grid start point to the subsequent grids of that gridded trajectory sequence, i.e.:

[0177] In the above formula This represents the predicted transport time from the starting point of the i-th gridded trajectory sequence to each of the other grids. f(·) is a specific nonlinear function or neural network.

[0178] Step S5-2 establishes the transportation time prediction vector. and the actual transportation time vector T i The loss function between them is as follows:

[0179]

[0180] The molecule represents the transportation time prediction vector. With the actual transportation time vector T i The L2 norm of the relative error percentage between them, with the denominator designed as Because of vectors and T i The first element of each element is 0. The model is trained using the loss function described above to obtain the MLP prediction model.

[0181] The implementation process of step S6 is as follows:

[0182] Step S6-1: Map the GPS coordinates of the starting point and ending point of the actual trucks in the open-pit mine to the divided area grid, thereby forming the starting point grid and the ending point grid.

[0183] Step S6-2, the path planning method is designed as follows:

[0184] Using the undirected graph G constructed in step S3 con =Adjacency matrix S of (V,E) con (N 2 ×N 2 Extending this to the entire historical trajectory data yields the grid connectivity matrix S′. con The entire historical trajectory data is then mapped onto a regional grid, yielding the number of GPS trajectory points mapped within each regional grid, denoted as the historical regional grid frequency vector. Where count i This represents the number of GPS trajectory points mapped within region grid i; starting from the starting grid according to the grid connectivity matrix S′ con A breadth-first search is performed until the destination grid is reached. If the destination grid is unreachable during the search, a breadth-first search is performed from the destination grid. The unreachable grid nodes from the starting grid and the destination grid are combined and shortest path searches are performed on each group to obtain the predicted grid path. Each grid in the obtained predicted grid path is then used to perform a Breadth-First Search on the V grid. fre Find the corresponding frequencies and add them together to get the total count. Sort the predicted grid paths from largest to smallest according to the total count, find the top k grid paths, output the top k predicted grid paths and generate the contextual spatiotemporal embedding of each gridded trajectory sequence.

[0185] Step S6-3: Input the contextual spatiotemporal embeddings of the first k gridded trajectory sequences predicted, and then use the trained MLP prediction model to predict the transportation time of open-pit mine trucks.

[0186] Example 1:

[0187] The following is in conjunction with the appendix Figure 1-2 Specific embodiments of the present invention are described to enable those skilled in the art to understand the invention. However, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0188] An implementation of a multi-view dynamic spatiotemporal semantic embedding method for predicting open-pit mine truck transportation time, based on a real GPS dataset of open-pit mine truck transportation, with a GPS sampling frequency of approximately 30 seconds and a data collection period of one month; includes the following steps:

[0189] Step 1: Based on the obtained GPS trajectory data of the open-pit mining trucks, preprocess the GPS trajectory point sequence and divide the area into grids.

[0190] Step 2: Map the trajectory sequences obtained in Step 1 onto the region grid, calculate the average velocity of each trajectory sequence within the corresponding grid, and obtain the gridded trajectory sequence. Step 3: Based on the gridded trajectory sequences obtained in Step 2, obtain the velocity features of the gridded trajectory sequences using the grid velocity feature extraction module.

[0191] Step 4: Based on the gridded trajectory sequence obtained in Step 2, obtain the gridded structure feature matrix through the gridded structure feature extraction module, and obtain the structural features of the gridded trajectory sequence from the gridded structure feature matrix according to the grid number of the gridded trajectory sequence.

[0192] Step 5: Extract the contextual spatiotemporal embedding of the gridded trajectory sequence obtained in Steps 3 and 4 using the contextual spatiotemporal embedding extraction module. Combine this with the transportation time in the sample using the MLP prediction model to obtain the predicted transportation time. Construct a loss function to train the model.

[0193] Step 6: Based on the path planning, obtain several candidate gridded trajectory sequences and generate the contextual spatiotemporal embedding of each gridded trajectory sequence. Input the embedding into the trained MLP prediction model to obtain the predicted transportation time.

[0194] Step 7: Experimental Environment and Hyperparameter Settings:

[0195] The deep learning framework used was PyTorch 1.13.0, and the programming language was Python 3.7. All experiments were conducted on a computer equipped with an NVIDIA GeForce RTX 3080Ti, with CUDA 12.0 and cu DNN 7.0 as the deep learning acceleration environment. During training, the time window size for trajectory sequence division was 1 hour, the step size was 10 GPS trajectory points, the grid size was 128×128, and the number of time periods M in the grid velocity feature extraction module was fixed at 4. s Set to 1 hour, t l Set to 7 days, t upThe timeout is set to 15 minutes; the CNN kernel extracts 2 short-term velocity features and 3 long-term velocity features, with an output channel size c of 4; the number of relevant grids in the RNN is fixed at 2; the number of encoder and decoder layers in the DynGEM is set to 2, and the hyperparameter α = 0.1. 4 In the GL-Transformer network unit, K is set to 3 for the K-hop algorithm; in the BiLSTM network unit, the hidden units are set to 64, and the dropout value is set to 0.25; in the path planning, k is set to 8.

Claims

1. A method for predicting truck transportation time in open-pit mines using multi-view dynamic spatiotemporal semantic embedding, characterized by: The prediction of open-pit mine truck transportation time was achieved through the grid velocity feature extraction module, grid structure feature extraction module, semantic information embedding module, and contextual spatiotemporal embedding extraction module. First, the raw GPS trajectory data of open-pit mining trucks is preprocessed to reduce the impact of outliers and extract trajectory sequences. The preprocessed trajectory data is then divided into regional grids, and the trajectory sequences are gridded to form gridded trajectory sequences. Based on the average speed of the original trajectory sequence within each grid, a gridded trajectory sequence with speed attributes is obtained. Secondly, the velocity features of the gridded trajectory sequence are extracted using the grid velocity feature extraction module. Secondly, the long-term historical trajectory data is divided according to time intervals to construct multi-view grid adjacency matrices for different time periods, forming a multi-view spatiotemporal network; The DynGEM module is used to extract the grid structure feature matrix with dynamic transportation information, and the structural features of the corresponding gridded trajectory sequence are obtained by combining the trajectory sequence. Then, the semantic information in the gridded trajectory sequence of the open-pit mine truck is embedded into the semantic information embedding module to obtain the embedded representation of the semantic information features; the velocity features, structural features and semantic information features of the gridded trajectory sequence are extracted into the contextual spatiotemporal embedding extraction module, which is composed of GL-Transformer network units and BiLSTM network units. Furthermore, the contextual spatiotemporal embedding of the final gridded trajectory sequence is combined with the transportation time in the sample using an MLP prediction model, and a loss function is designed to train the model; the MLP is referred to as a multilayer perceptron. Finally, based on the start and end points of the transportation time to be predicted, combined with grid information, a path planning method is designed to obtain several candidate gridded trajectory sequences. The contextual spatiotemporal embedding of each gridded trajectory sequence is generated and input into the trained MLP prediction model to obtain the transportation time prediction result.

2. The method for predicting open-pit mine truck transportation time using multi-view dynamic spatiotemporal semantic embedding according to claim 1, characterized in that, The specific steps are as follows: Step S1: Sample the location data of the open-pit mining truck to form the original GPS trajectory data of the open-pit mining truck. Clean the original GPS trajectory data by writing a program. Then, serialize the GPS trajectory data by writing a trajectory sequence extraction program. Divide the processed trajectory data into regions to obtain a region grid. Map the trajectory sequence to the corresponding region grid and calculate the average speed of the trajectory sequence in each grid to obtain a gridded trajectory sequence with speed attributes. Step S2: Obtain short-term velocity matrix and long-term velocity matrix from the region mesh obtained in step S1 according to different time intervals; recover the missing velocity values ​​in the velocity matrix through the non-negative tensor decomposition module; and then extract the velocity features of the meshed trajectory sequence through the mesh velocity feature extraction module. Step S3: Divide the preprocessed trajectory data into segments according to time, fill the trajectory data of different time periods into the grid, and construct a regional grid dynamic multi-view spatiotemporal network based on the grid's connectivity, connection length, and similarity. The dynamic multi-view spatiotemporal network is input into the DynGEM module to extract the structural feature matrix under the multi-view gridded trajectory; finally, the multi-view structural feature matrix is ​​weighted and summed to obtain the structural features of the gridded trajectory sequence. Step S4: The semantic information in the gridded trajectory sequence obtained in step S1 is processed through the semantic information embedding module to obtain the embedded representation of the semantic information features; the velocity features, structural features and semantic information features of the gridded trajectory sequence are processed through the context spatiotemporal embedding extraction module to extract the context spatiotemporal embedding of the gridded trajectory sequence, wherein the context spatiotemporal embedding extraction module is composed of GL-Transformer network units and BiLSTM network units. Step S5: The contextual spatiotemporal embedding of the gridded trajectory sequence obtained in Step S4 is combined with the transportation time in the sample using an MLP prediction model, and a loss function is designed to train the model. Step S6: Based on the starting point and ending point of the transportation time to be predicted, and combined with grid information, design a path planning method to obtain several candidate gridded trajectory sequences, and generate the contextual spatiotemporal embedding of each gridded trajectory sequence. Use the MLP prediction model trained in step S5 to obtain the transportation time prediction result.

3. The method for predicting open-pit mine truck transportation time using multi-view dynamic spatiotemporal semantic embedding according to claim 2, characterized in that: The implementation process of step S1 is as follows: The trajectory cleaning procedure filters out duplicate and abnormal data from the original GPS trajectory data; the trajectory sequence extraction procedure divides the GPS trajectory data of the same driver into segments according to a sliding time window with a set time of T and a step size of n GPS trajectory points to complete the extraction of the trajectory sequence. The specific operation of the region grid division is as follows: First, find the maximum and minimum values ​​of the latitude and longitude of all trajectory points to form the latitude and longitude coordinates [lat1, lng1] and [lat2, lng2] of the upper left and lower right of the region; divide the region into N×N grids, each two-dimensional grid has an ID, namely lat_id and lng_id, where lat_id represents the latitude ID of the two-dimensional grid and lng_id represents the longitude ID of the two-dimensional grid; calculate the length of each grid: lat_len = (lat2-lat1) / (N-1 Given the latitude and longitude information lat and lng of each GPS track point, lat_id = (lat-lat1) / lat_len and lng_id = (lng-lng1) / lng_len, we can map the latitude and longitude information lat and lng of each GPS track point to the grid, i.e., lat_id = (lat-lat1) / lat_len and lng_id = (lng-lng1) / lng_len, to obtain the two-dimensional grid id corresponding to the latitude and longitude of the track point; further, we sort the two-dimensional grid id by row, and number the grids sequentially from 1 according to the sorted grids, and the number of each grid is the one-dimensional grid id of that grid; The trajectory sequence is: trajectory data Tr: Tr = {Tr1, Tr2, ..., Tr...} i , ...,Tr M } is a set of trajectory data, where M represents the number of trajectory sequences; Tr i ∈Tr (1≤i≤M) represents a trajectory sequence, denoted as in Represents the j-th GPS trajectory point in the i-th trajectory sequence, i R This represents the number of GPS trajectory points in the i-th trajectory sequence; each GPS trajectory point in and for Latitude and longitude of the trajectory point for The grid number of the trajectory point for The velocity of the trajectory point and for timestamps of trajectory points; In addition, each trajectory contains other additional attributes: represent Driver information represent Travel time, represent travel distance, represent Weekly information represent Number of stays and represent The length of stay; The gridded trajectory sequence is obtained by merging multiple trajectory points within a grid according to the above GPS trajectory sequence, retaining the information of the merged trajectory points. Among them, grid and corresponding GPS track points The content information is consistent; The method for merging multiple trajectory points within a grid is as follows: If a trajectory sequence contains multiple trajectory points within the same grid, these multiple trajectory points are merged into one trajectory point. The merging method is as follows: Calculate the distance from these multiple trajectory points to the center of the grid, select the trajectory point with the shortest distance as the trajectory point within the merged grid, and calculate the velocity of the trajectory sequence in the corresponding grid using the timestamp of the trajectory sequence and the latitude and longitude of the GPS trajectory points to obtain a gridded trajectory sequence with velocity attributes.

4. The method for predicting open-pit mine truck transportation time using multi-view dynamic spatiotemporal semantic embedding according to claim 2, characterized in that: The implementation process of step S2 is as follows: Step S2-1: Construct the short-term velocity matrix A using the region grid obtained in step S1. r Long-term velocity matrix A h Fusion velocity matrix A m Short-term velocity matrix To store tt s The short-term travel speed from hour to t, where M represents the number of time intervals, then t represents the speed of each time interval. s / M hours, N 2 A represents the number of grid cells in the divided region. r (i, j) = a represents the time period j within the i-th region grid. s The average driving speed is taken, and if there are multiple trajectory points in the area grid, the average value is taken. In order to obtain effective transportation characteristics from historical information, a long-term velocity matrix is ​​constructed. That is, matrix To store tt l Given the long-term driving speed from day t to t, then the speed for each time period is t. l / M days, A h (i, j) = b represents the time period j within the i-th region grid. l Average driving speed; due to A r and A h With similar structures, the fusion velocity matrix is ​​constructed. To integrate its information, the specific integration strategy is as follows: Based on the above three matrices A r A m A h As different channels, they form a three-dimensional nonnegative tensor. Using existing nonnegative tensor decomposition algorithms, the originally sparse tensors are decomposed. Become a dense tensor Step S2-2, for each region grid g i Design a triplet to record velocity characteristics: Represents the region grid g i The driving speed at time j, where disi represents the distance the trajectory crosses the grid area. A rough estimate of the time taken to traverse the grid; In order to capture the region grid g i The trend of speed change, defining the speed feature sequence. Represents the region grid g i Speed ​​characteristics; in Velocity feature sequence X i One-dimensional convolutional filters using CNN; velocity feature sequence X i ∈R 3×M Consider it as a three-channel input; the j-dimensional output after one-dimensional convolution is: * indicates a convolution operation, b is the bias term, and "tanh" is the activation function; the output sequence... c is the number of one-dimensional convolutional filters, W conv ∈R k×M Let k be the parameter matrix, where k is the kernel size. Step S2-3: Use a recurrent neural network (RNN) to learn the region grid g. i Time correlation of internal driving speed: W c and W h is a learnable parameter matrix, and b is a bias term. Represents the hidden state at step t; the RNN model output is Recorded as Reflects the regional grid g i Short-term velocity characteristics; Step S2-4, define the upstream mesh g up That is, for a given gridded trajectory sequence Among them, grid It's a grid. The upstream grid of the current grid is affected by the velocity characteristics of its upstream grid; find the current grid g. i All upstream grids, and extract the time interval t before the current time interval t. up The velocity distribution of all upstream grids is calculated using the existing KL divergence procedure to determine the grid g. i The similarity between the velocity distribution and the velocity distribution of all upstream grids is used to select the k upstream grids with the highest similarity as grid g. i The relevant region grids are used; short-term velocity features of these k relevant region grids are extracted using CNN and RNN. The CNN is a deep neural network with a convolutional structure, and the RNN is a recurrent neural network with a chain structure. Besides short-term speed characteristics, long-term speed characteristics can represent stable driving speed patterns over a long period of time, patterns that short-term speed characteristics cannot represent; extending the time window from [tt]... s hour, t] changed to [tt] l Similar to methods for extracting short-term velocity features, the grid g can be extracted using the [day, t] method. i Long-term velocity characteristics Step S2-5: Connect the short-term velocity features and the long-term velocity features using a fully connected layer to obtain the region mesh g. i velocity characteristics The velocity features of each region grid are connected according to the gridded trajectory sequence to obtain the velocity features V of the overall gridded trajectory sequence. speed .

5. The method for predicting open-pit mine truck transportation time using multi-view dynamic spatiotemporal semantic embedding according to claim 2, characterized in that: The implementation process of step S3 is as follows: Step S3-1: Based on the divided region grid, design the grid into an undirected graph G = (V, E) using the method of grids as nodes and the connectivity between gridded trajectories as edges, where V and E represent vertices and edges in the graph, respectively; the node set V represents the grid, and the edge set represents the connectivity between two region grids; here, the undirected graph G is defined. con =Adjacency matrix S of (V, E) con (N 2 ×N 2 ),in Represents the region grid g i and g j Connected? Step S3-2, Figure G con Only the connectivity between two region grids was considered, but the length of the connecting edge between the two region grids is also very important; therefore, graph G is added here. len adjacency matrix S len (N 2 ×N 2 ),in Represents the grid g i and g j The reciprocal of the number of grid intervals: Where len i，j =long i,j +wide i，j , long i,j Represents the region grid g i and g j The number of grid cells in the vertical spacing, wide i，j Represents the region grid g i and g j The number of horizontally spaced grid cells; Step S3-3 uses the number of trajectory points within a region grid to measure the spatial similarity of a region grid over a period of time. Higher similarity reflects more similar traffic conditions among the region grids over that period. Therefore, graph G is added here. imp adjacency matrix S imp (N 2 ×N 2 ),in Represents the region grid g i and g j Spatial similarity between them: Among them, M i Represents the region grid g i The number of internal trajectory points, M j Represents the region grid g j The number of internal trajectory points; Step S3-4: Design a time sliding window w based on historical trajectory data, then divide the historical trajectory data into W segments according to the time of day (T1, T2, ..., T...). K And extract the adjacency matrix from each historical trajectory data segment. and The adjacency matrix S mentioned above con S len and S imp It is dynamic, so the DynGEM module is used to handle this; a Deep Belief Network (DBN) is used to pre-train the parameters on the existing SDNE model to obtain the initial parameter set. Adjacency Matrix and The parameters are input into the SDNE model, and after training, the corrected parameter set is obtained. and Using the parameter set from the previous time step and Initialize the current time step parameter group and After training iterations (W-1), multiple region grid map embedding feature matrices can be obtained; Steps S3-5: After iterative training, the region grid feature matrix Y is obtained. con Y len and Y imp The final mesh structure feature matrix is ​​obtained by weighted summation: Y=W con *Y con +W len *Y len +W imp *Y imp Obtain the feature matrix of the grid structure p i ·lat represents GPS track point p i latitude, p i ·lon represents GPS track point p i The longitude is obtained by connecting the grid structure features and the latitude and longitude of the original GPS trajectory points through a fully connected layer to obtain the region grid g. i Structural features By connecting the structural features of each region grid according to the gridded trajectory sequence, the velocity feature V of the overall gridded trajectory sequence is obtained. road .

6. The method for predicting open-pit mine truck transportation time using multi-view dynamic spatiotemporal semantic embedding according to claim 2, characterized in that: The implementation process of step S4 is as follows: The GL-Transformer network unit is used to extract the spatiotemporal information of the transport of the gridded trajectory sequence. The GL-Transformer network unit consists of two main parts: a global encoder and a local encoder, which ensures that the GL-Transformer network unit has both global and local dependencies. The global encoder consists of four parts: position embedding, multi-head self-attention, feedforward neural network, and normalized-residual connection. The local encoder consists of four parts: position embedding, masked multi-head self-attention, feedforward neural network, and normalized-residual connection. Among these four parts, only the masked multi-head self-attention is different from the multi-head self-attention in the global encoder. The other three parts are the same as the global encoder, so only the masked multi-head self-attention is introduced here. Step S4-1: Calculate the attention weight matrix A∈R n×n The calculation formula is as follows: The attention weight matrix A is used to calculate the mask adjacency matrix M∈R using the existing K-hop algorithm. n×n The mask self-attention is calculated using the mask adjacency matrix, and the calculation formula is as follows: Here, Mask(·) is a function used to calculate the mask attention weight matrix. Attention(Q，K，V) It is the self-attention mask obtained through calculation; Step S4-2, V speed and V road The velocity feature vector H of the gridded trajectory sequence is obtained through GL-Transformer network units. speed The structural feature vector H of the gridded trajectory sequence road ,in Step S4-3, after extracting the velocity and structural features of the gridded trajectory sequence, also needs to consider other factors affecting transportation time prediction, including travel time (timeID), weekday information (weekID), driver information (driverID), number of stops (DwellNumber), stop time (DwellTime), and spatial distance. The features of these six factors are then embedded and concatenated to form the final semantic information feature V. add ∈R 21 ; Step S4-4, the above V add Separately and Embedded connections are made and input into the BiLSTM network element: The contextual spatiotemporal embedding representation H of the gridded trajectory sequence is obtained. i .

7. The method for predicting open-pit mine truck transportation time using multi-view dynamic spatiotemporal semantic embedding according to claim 2, characterized in that: The implementation process of step S5 is as follows: The contextual spatiotemporal embedding of the gridded trajectory sequence is obtained in step S4. i R The number of region grids representing the i-th gridded trajectory sequence; The transportation time prediction problem is transformed into a sequence H i Compared with actual transportation time The regression problem between them; Step S5-1, a vector is constructed. Represents the distance from the grid start point to the grid. Actual transportation time; in Used to correspond to sequence H i In Right now It corresponds The nonlinear mapping f(·) of a multilayer perceptron (MLP) is used to predict the transport time of each gridded trajectory sequence from the grid start point to the subsequent grids of that gridded trajectory sequence, i.e.: In the above formula This represents the predicted transport time from the starting point of the i-th gridded trajectory sequence to each of the other grids; f(·) is a specific nonlinear function or neural network. Step S5-2 establishes the transportation time prediction vector. and the actual transportation time vector T i The loss function between them is as follows: The molecule represents the transportation time prediction vector. With the actual transportation time vector T i The L2 norm of the relative error percentage between them, with the denominator designed as Because of vectors and T i The first element of each element is 0; the model is trained using the loss function described above to obtain the MLP prediction model.

8. The method for predicting open-pit mine truck transportation time using multi-view dynamic spatiotemporal semantic embedding according to claim 2, characterized in that: The implementation process of step S6 is as follows: Step S6-1: Map the GPS coordinates of the starting point and ending point of the actual trucks in the open-pit mine production to the divided area grid, thereby forming the starting point grid and the ending point grid. Step S6-2, the path planning method is designed as follows: Using the undirected graph G constructed in step S3 con =Adjacency matrix S of (V, E) con (N 2 ×N 2 Extending this to the entire historical trajectory data yields the grid connectivity matrix S′. con The entire historical trajectory data is then mapped onto a regional grid, yielding the number of GPS trajectory points mapped within each regional grid, denoted as the historical regional grid frequency vector. Where count i This represents the number of GPS trajectory points mapped within region grid i; starting from the starting grid according to the grid connectivity matrix S′ con Perform a breadth-first search until the endpoint grid is found; If the destination grid becomes unreachable during the search, a breadth-first search is performed starting from the destination grid. Unreachable grid nodes from the starting grid and the destination grid are combined and shortest path searches are performed on each group to obtain the predicted grid path. Each grid in the predicted path is then analyzed in V... fre Find the corresponding frequencies and add them together to get the total Count. Sort the predicted grid paths from largest to smallest according to the total Count, find the top k grid paths, output the top k predicted grid paths and generate the contextual spatiotemporal embedding of each gridded trajectory sequence. Step S6-3: Input the contextual spatiotemporal embeddings of the first k gridded trajectory sequences predicted, and then use the trained MLP prediction model to predict the transportation time of open-pit mine trucks.

Images