Industrial delivery order-combining and vehicle-assembling method

Abstract

The invention discloses an industrial delivery order-combining and vehicle-combining method, which comprises the following steps that A, an undirected graph G (V, E) is constructed according to an order-combining and vehicle-combining scheme, any V is all arrival order sets (v1. V2,... vn), the round-trip frequency between arrival orders is taken as the weight vij between two points, vij is taken as the weight between a point vi and a point vj, vij is equal to vji due to the undirected graph, and for any point vi in the graph, vij is equal to vji; the value of the degree di is the sum of the weights of all edges connected with the degree di, that is, an order combining history exists between two arrival points, that is, di = # imgabs0 #. By adopting an artificial intelligence model and configuring various constraints of a model type, the model automatically calculates an optimal order combining and vehicle combining scheme for industrial delivery, and under the condition that the process of an existing manual order combining and vehicle combining result is not changed, the order combining and vehicle combining efficiency is greatly improved. Industrial companies can refer to or adopt the generated carpooling scheme to improve the efficiency of transportation resources and reduce the cost.

Description

Technical Field

[0001] The invention relates to the field of transportation technology, and in particular to a method for assembling vehicles for industrial delivery. Background Art

[0002] The existing carpooling is inefficient, which reduces the cargo loading rate and the transportation efficiency of the vehicle owner and increases the transportation cost. Therefore, we propose a method of carpooling for industrial delivery. Summary of the invention

[0003] The object of the present invention is to provide a method for assembling vehicles for industrial delivery so as to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for assembling a vehicle for industrial delivery, comprising the following steps:

[0005] A. Construct an undirected graph G(V,E) based on the carpooling and grouping scheme, where any V is the set of all arrival orders (v1.v2,...vn), the frequency of round trips between arrival orders is the weight between two points, and vij is the weight between point vi and point vj. Since it is an undirected graph, vij=vji. For any point vi in ​​the graph, its degree di is the sum of the weights of all edges connected to it, that is, there is a grouping history between the two arrival points, that is:

[0006] B. Get the degree matrix of each arrival order:

[0007] C. Use the full connection method based on Gaussian kernel function to construct the adjacency matrix W.

[0008] D. Construct the Laplace matrix according to the degree matrix and the adjacency matrix, L = DW, and use Ncut to perform spectral clustering.

[0009] E. By dividing the clusters where daily orders are located, that is, packaging the historical frequencies of transportation between different arrival orders, a priori merged orders are formed to be fed to the deep dual Q-value network for calculation;

[0010] F. Grouping orders in the matching space according to the direction angle, the latitude and longitude of the starting point, the latitude and longitude of the end point, and the departure time, including: normalizing the direction angle and departure time; calculating the absolute difference in direction angles and departure time between orders; referring to the historical transportation routes, when it is determined that the absolute difference in direction angles between the current orders is less than the first direction angle threshold and the absolute difference in departure time is less than the set first time threshold, the number of different vehicle models and the maximum load weight of different vehicle models are used as constraints to construct the state of the deep dual Q-value network agent;

[0011] G. By discretizing and classifying the full load rates of all different types of vehicles as the agent behavior space, the limitation of the deep dual Q-value network that the agent behavior selection can only be one-dimensional is broken, and the agent behavior space is raised to a high dimension;

[0012] H. The deep dual Q-value network enters the iteration process, input: number of iterations T, state feature dimension n, behavior space, step size α, decay factor γ, exploration rate ∈, current Q network benefit Q, target Q, network benefit Q′, number of samples of batch gradient descent m, target Q network parameter update frequency C, output: current Q network parameters;

[0013] I. Initialize the values ​​Q corresponding to all states and actions, randomly initialize all parameters w of the current Q network, initialize the parameters w′=w of the target Q network Q′, clear the set D of experience replays, use the transportation order volume as the step size for the agent's single-round calculation, and give priority to feeding the prior order data based on spectral clustering into the deep dual Q-value network to prepare for iterative training.

[0014] Preferably, in step D, the Ncut graph cut improves the indicator vector h. The indicator vector of the RatioCut graph cut uses $\frac{1}{\sqrt{|A_j|}}$ to indicate the sample attribution, while the Ncut graph cut uses the subgraph weight $\frac{1}{\sqrt{vol(A_j)}}$ to indicate the indicator vector h, and finally obtains the relationship cluster between different arrival orders, thereby completing clustering.

[0015] Preferably, the carpooling order combination scheme in step A also includes carpooling order combination order pre-processing: (1) obtaining orders to be shipped and merging orders with the same arrival point; (2) maintaining carrier and vehicle information and transportation contract billing formulas; (3) creating a knowledge graph for all orders to be merged, with the arrival warehouse as the node and the edge as the two-point freight. Whether the two nodes are connected depends on the historical combination of the arrival warehouse and the arrival warehouse. If they have never been combined in history, they are not connected.

[0016] Preferably, a coordinate system is established with the order hotspot as the coordinate origin, and the coordinate origin has a first longitude and a first latitude; the longitude and latitude of the starting point of the order are obtained, and the order sets the end point and departure time; the longitude of the starting point is recorded as the second longitude, and the latitude is recorded as the second latitude, and the longitude of the end point is recorded as the third longitude, and the latitude is recorded as the third latitude; the direction angle of the space vector between the starting point and the end point of the order is calculated; and the order is grouped in the matching space according to the direction angle, the longitude and latitude of the starting point, the raindrop type of the longitude and latitude of the end point, and the departure time.

[0017] Preferably, the calculation of the direction angle includes: mapping the longitude and latitude of the end point and the starting point set in the order to the coordinate system, connecting the starting point and the end point to form a vector; projecting the vector onto the coordinate axis, setting the angle between the vector and the projection on the coordinate axis as the direction angle, and grouping the direction angle, the longitude and latitude of the starting point, the longitude and latitude of the end point and the departure time in the matching space, including a grouping method based on clustering analysis: normalizing the direction angle and the departure time; establishing a fuzzy similarity matrix and determining its similarity coefficient; setting a first threshold, and forming a group order when the similarity coefficient between orders is less than the first threshold.

[0018] Preferably, step F also includes adding a mileage / piece number attenuation factor as a soft constraint condition for each carpooling transport route combination, and weighting the attenuation factor by analyzing the actual situation of historical transport routes.

[0019] Preferably, the iterative process in step H includes:

[0020] a) Initialize S as the first state of the current state sequence and obtain its feature vector φ(S);

[0021] b) Use φ(S) as input in the Q network to obtain the Q-value output corresponding to all actions of the Q network, and use the ∈-greedy method to select the corresponding action A from the current Q-value output;

[0022] c) Execute the current action A in state S, obtain the feature vector φ(S′) and reward R$ corresponding to the new state S′, and determine whether the state is terminated;

[0023] d) Store the five-tuple {φ(S), A, R, φ(S′), is_end} into the experience replay set D;

[0024] e) S = S′;

[0025] f) Sample m, from the experience replay set D, and calculate the current target Q value,;

[0026] g) Update all parameters w of the Q network using the mean squared error loss function through gradient back propagation of the neural network;

[0027] h) If T%C=1, then update the target Q network parameter w′=w. If S′ is in the terminal state, the current round of iteration is completed, otherwise go to step b).

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention adopts an artificial intelligence model and configures various constraints in the model to enable the model to automatically calculate the best carpooling plan for industrial delivery. Without changing the existing process of manual carpooling results, industrial companies can refer to or adopt the generated carpooling plan to improve the efficiency of transportation resources and reduce costs, and meet various constraints used in daily industrial delivery: (1) meet the arrival time of each arrival point in the order, and reserve at least 0.5 days; (2) Orders of the same arrival order are preferentially combined under the constraint of the arrival time window; (3) adopt the raindrop-type delivery principle, and the maximum radius from the shipping point to the farthest arrival point is 300 kilometers as an ellipse. (4) The combination meets the maximum load capacity of the vehicle model without overloading. It makes full use of intelligent technologies and means such as intelligent hardware, the Internet of Things, and large models. Through intelligent hardware and the Internet of Things, that is, information uploading, intelligent processing, intelligent storage, cleaning and other technologies, a method of carpooling is studied to reasonably and efficiently carpool, improve the cargo loading rate and the transportation efficiency of the vehicle owner, reduce transportation costs, and effectively integrate and optimize the allocation of transportation resources, thereby achieving the overall goal of improving efficiency, reducing costs, integrating resources, ensuring quality, and improving services. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a three-layer neural network diagram used by the ordinary DQN of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] See also Figure 1 , a method for grouping vehicles for industrial delivery, comprising the following steps:

[0033] A. Construct an undirected graph G(V,E) based on the carpooling and grouping scheme, where any V is the set of all arrival orders (v1.v2,...vn), the frequency of round trips between arrival orders is the weight between two points, and vij is the weight between point vi and point vj. Since it is an undirected graph, vij=vji. For any point vi in ​​the graph, its degree di is the sum of the weights of all edges connected to it, that is, there is a grouping history between the two arrival points, that is:

[0034] B. Get the degree matrix of each arrival order:

[0035] C. Use the full connection method based on Gaussian kernel function to construct the adjacency matrix W (in this case, the similarity matrix is ​​the same as the adjacency matrix).

[0036] D. Construct the Laplace matrix according to the degree matrix and the adjacency matrix, L = DW, and use Ncut to perform spectral clustering.

[0037] E. By dividing the clusters where daily orders are located, that is, packaging the transportation history frequencies between different arrival orders, a priori merged orders are formed to be fed to the deep dual Q-value network (DDQN) for calculation;

[0038] F. Grouping orders in the matching space according to the direction angle, the latitude and longitude of the starting point, the latitude and longitude of the end point and the departure time, including: normalizing the direction angle and departure time; calculating the absolute difference of the direction angle and the absolute difference of the departure time between orders; referring to the historical transportation routes, when it is judged that the absolute difference of the direction angle between the current orders is less than the first direction angle threshold and the absolute difference of the departure time is less than the set first time threshold, the number of different models and the maximum load of different models are used as constraints to construct the states of the deep dual Q-value network agent;

[0039] G. By discretizing and classifying the full load rates of all different types of vehicles as the agent action space (ActionSpace), the limitation of the deep dual Q-value network that the agent action selection can only be one-dimensional is broken, and the agent action space (ActionSpace) is upgraded to a high dimension. On the basis of greatly increasing the agent's behavior selection, it also avoids the problem of slow convergence speed of the decision gradient-based reinforcement learning model;

[0040] H. The deep dual Q-value network enters the iteration process, input: number of iterations T, state feature dimension n, behavior space (ActionSpace), step size α, decay factor γ, exploration rate ∈, current Q network benefit Q, target Q, network benefit Q′, number of samples of batch gradient descent m, target Q network parameter update frequency C, output: current Q network parameters;

[0041] I. Initialize the values ​​Q corresponding to all states and actions, randomly initialize all parameters w of the current Q network, initialize the parameters w′=w of the target Q network Q′, clear the set D of experience replays, use the transportation order volume as the step size for the agent's single-round calculation, and give priority to feeding the prior order data based on spectral clustering into the deep dual Q-value network to prepare for iterative training.

[0042] In step D, Ncut graph cutting improves the indicator vector h. The indicator vector of RatioCut graph cutting uses $\frac{1}{\sqrt{|A_j|}}$ to indicate the sample attribution, while Ncut graph cutting uses the subgraph weight $\frac{1}{\sqrt{vol(A_j)}}$ to indicate the indicator vector h. Finally, the relationship clusters between different arrival orders are obtained, thus completing the clustering.

[0043] The carpooling order combination scheme in step A also includes the pre-processing of carpooling order combination: (1) obtaining the orders to be shipped and merging the orders with the same arrival point; (2) maintaining the carrier and vehicle information (including loading capacity) and the transportation contract billing formula; (3) creating a knowledge graph for all the orders to be merged, with the arrival warehouse as the node and the edge as the two-point freight. Whether the two nodes are connected depends on the historical combination of the arrival warehouse and the arrival warehouse. If they have never been combined in history, they are not connected.

[0044] A coordinate system is established with the order hotspot as the coordinate origin, and the coordinate origin has a first longitude and a first latitude; the longitude and latitude of the starting point of the order are obtained, and the order sets the end point and departure time; the longitude of the starting point is recorded as the second longitude, and the latitude is recorded as the second latitude, and the longitude of the end point is recorded as the third longitude, and the latitude is recorded as the third latitude; the direction angle of the space vector between the start point and the end point of the order is calculated; the order is grouped in the matching space according to the direction angle, the longitude and latitude of the starting point, the raindrop type of the longitude and latitude of the end point, and the departure time.

[0045] The calculation of the direction angle includes: mapping the longitude and latitude of the end point and the starting point set by the order to the coordinate system, connecting the starting point and the end point to form a vector; projecting the vector onto the coordinate axis, and setting the angle between the vector and the projection on the coordinate axis as the direction angle. The direction angle, the longitude and latitude of the starting point, the longitude and latitude of the end point and the departure time are grouped in the matching space, including a grouping method based on clustering analysis: normalizing the direction angle and the departure time; establishing a fuzzy similarity matrix and determining its similarity coefficient; setting a first threshold, and forming a group order when the similarity coefficient between orders is less than the first threshold.

[0046] Step F also includes adding a mileage / piece count attenuation factor as a soft constraint condition for each carpooling transport route combination. By analyzing the actual situation of historical transport routes and weighting this attenuation factor, it is possible to effectively avoid the rigid exclusion of unreasonable routes due to the results of algorithm calculations, thereby avoiding the "top-heavy" situation. The matching space is a circular area with the starting point of the order as the origin and M as the radius.

[0047] The iterative process in step H includes:

[0048] a) Initialize S as the first state of the current state sequence and obtain its feature vector φ(S);

[0049] b) Use φ(S) as input in the Q network to obtain the Q-value output corresponding to all actions of the Q network, and use the ∈-greedy method to select the corresponding action A from the current Q-value output;

[0050] c) Execute the current action A in state S, obtain the feature vector φ(S′) and reward R$ corresponding to the new state S′, and determine whether the state is terminated;

[0051] d) Store the five-tuple {φ(S), A, R, φ(S′), is_end} into the experience replay set D;

[0052] e) S = S′;

[0053] f) Sample m, from the experience replay set D, and calculate the current target Q value,;

[0054] g) Update all parameters w of the Q network using the mean squared error loss function through gradient back propagation of the neural network;

[0055] h) If T%C=1, then update the target Q network parameter w′=w. If S′ is in the terminal state, the current round of iteration is completed, otherwise go to step b).

[0056] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for assembling vehicles for industrial delivery, characterized in that: The following steps are involved: A. Construct an undirected graph G(V,E) based on the carpooling and grouping scheme, where any V is the set of all arrival orders (v1.v2,...vn), the frequency of round trips between arrival orders is the weight between two points, and vij is the weight between point vi and point vj. Since it is an undirected graph, vij=vji. For any point vi in ​​the graph, its degree di is the sum of the weights of all edges connected to it, that is, there is a grouping history between the two arrival points, that is: di = ; B. Get the degree matrix of each arrival order: D = ; C. Use the full connection method based on Gaussian kernel function to construct the adjacency matrix W, Vij = Vji = exp(-; ); D. Construct the Laplace matrix according to the degree matrix and the adjacency matrix, L=DW, and use Ncut to perform spectral clustering, NCut(A1,A2,...Ak) = ; E. By dividing the clusters where daily orders are located, that is, packaging the historical frequencies of transportation between different arrival orders, a priori merged orders are formed to be fed to the deep dual Q-value network for calculation; F. Grouping orders in the matching space according to the direction angle, the latitude and longitude of the starting point, the latitude and longitude of the end point, and the departure time, including: normalizing the direction angle and departure time; calculating the absolute difference in direction angles and departure time between orders; referring to the historical transportation routes, when it is determined that the absolute difference in direction angles between the current orders is less than the first direction angle threshold and the absolute difference in departure time is less than the set first time threshold, the number of different vehicle models and the maximum load weight of different vehicle models are used as constraints to construct the state of the deep dual Q-value network agent; G. By discretizing and classifying the full load rates of all different types of vehicles as the agent behavior space, the limitation of the deep dual Q-value network that the agent behavior selection can only be one-dimensional is broken, and the agent behavior space is raised to a high dimension; H. The deep dual Q-value network enters the iteration process, input: number of iterations T, state feature dimension n, behavior space, step size α, decay factor γ, exploration rate ϵ, current Q network benefit Q, target Q, network benefit Q′, number of samples of batch gradient descent m, target Q network parameter update frequency C, output: current Q network parameters; I. Initialize the values ​​Q corresponding to all states and actions, randomly initialize all parameters w of the current Q network, initialize the parameters w′=w of the target Q network Q′, clear the set D of experience replay, use the transportation order volume as the step length of the agent's single-round calculation, and give priority to feeding the prior order data based on spectral clustering into the deep dual Q-value network to prepare for iterative training.

2. The method for assembling vehicles for industrial delivery according to claim 1, characterized in that: In step D, Ncut graph cutting improves the indicator vector h. The indicator vector of RatioCut graph cutting uses $\frac{1}{\sqrt{|A_j|}}$ to indicate the sample attribution, while Ncut graph cutting uses the subgraph weight $\frac{1}{\sqrt{vol(A_j)}}$ to indicate the indicator vector h, and finally obtains the relationship clusters between different arrival orders, thereby completing clustering.

3. The method for assembling vehicles for industrial delivery according to claim 1, characterized in that: The carpooling order combination scheme in step A also includes the pre-processing of carpooling order combination: (1) obtaining orders to be shipped and merging orders with the same arrival point; (2) maintaining carrier and vehicle information and transportation contract billing formula; (3) creating a knowledge graph for all orders to be merged, with the arrival warehouse as the node and the edge as the two-point freight. Whether the two nodes are connected depends on the historical combination of the arrival warehouse and the arrival warehouse. If they have never been combined in history, they are not connected.

4. The method for assembling vehicles for industrial delivery according to claim 3 is characterized in that: A coordinate system is established with the order hotspot as the coordinate origin, and the coordinate origin has a first longitude and a first latitude; the longitude and latitude of the starting point of the order are obtained, and the order sets the end point and departure time; the longitude of the starting point is recorded as the second longitude, and the latitude is recorded as the second latitude, and the longitude of the end point is recorded as the third longitude, and the latitude is recorded as the third latitude; the direction angle of the space vector between the start point and the end point of the order is calculated; the order is grouped in the matching space according to the direction angle, the longitude and latitude of the starting point, the raindrop type of the longitude and latitude of the end point, and the departure time.

5. The method for assembling vehicles for industrial delivery according to claim 4, characterized in that: The calculation of the direction angle includes: mapping the longitude and latitude of the end point and the starting point set in the order to the coordinate system, connecting the starting point and the end point to form a vector; projecting the vector onto the coordinate axis, setting the angle between the vector and the projection on the coordinate axis as the direction angle, and grouping the direction angle, the longitude and latitude of the starting point, the longitude and latitude of the end point and the departure time in the matching space, including a grouping method based on clustering analysis: normalizing the direction angle and the departure time; establishing a fuzzy similarity matrix and determining its similarity coefficient; setting a first threshold, and forming a group order when the similarity coefficient between the orders is less than the first threshold.

6. The method for assembling vehicles for industrial delivery according to claim 1, characterized in that: Step F also includes adding a mileage / piece number attenuation factor as a soft constraint condition for each carpooling transport route combination, and weighting this attenuation factor by analyzing the actual situation of historical transport routes.

7. The method for assembling vehicles for industrial delivery according to claim 1, characterized in that: The iterative process in step H includes: a) Initialize S as the first state of the current state sequence and obtain its feature vector φ(S); b) Use ϕ(S) as input in the Q network to obtain the Q-value output corresponding to all actions of the Q network, and use the ϵ-greedy method to select the corresponding action A from the current Q-value output; c) Execute the current action A in state S, obtain the feature vector ϕ(S′) and reward R$ corresponding to the new state S′, and determine whether the state is terminated; d) Store the five-tuple {ϕ(S),A,R,ϕ(S′),is_end} into the experience replay set D; e) S = S′; f) Sample m from the experience replay set D, , calculate the current target Q value , ; g) Use mean square error loss function Update all parameters w of the Q network through gradient backpropagation of the neural network; h) If T%C=1, update the target Q network parameter w′=w. If S′ is in the terminal state, the current round of iteration is completed, otherwise go to step b).

Images