Path planning method, system and transportation controller

Abstract

This invention discloses a path planning method, system, and transport controller. The path planning method includes the following steps: S1, obtaining the starting point and ending point of the path to be planned; S2, determining multiple travel routes from the starting point to multiple boundary points within the nearby region and the first estimated travel time on each travel route using a nearby region path planning model; S3, determining the second estimated travel time from each boundary point to the ending point using a trained distant region time estimation model; S4, determining which travel route is selected as the optimal path based on the first estimated travel time for each travel route and the second estimated travel time from the boundary point of the travel route to the ending point. This invention avoids the inefficient blind search in distant regions when executing Dijkstra's algorithm, reduces path planning time and computational cost while ensuring path planning accuracy, improves the efficiency of the transport controller, and can promptly guide overhead cranes away from congested areas, improving handling efficiency.

Description

Technical Field

[0001] This invention relates to the field of automated handling systems, and in particular to path planning methods, systems, and transport controllers for overhead crane handling systems in semiconductor processing. Background Technology

[0002] Overhead crane transport systems (OHT systems) are key equipment used for material handling in automated factories such as semiconductor manufacturing and high-end manufacturing.

[0003] In overhead crane handling systems, a large number of overhead cranes run on tracks. Therefore, rationally planning the movement path of each overhead crane is crucial to improving handling efficiency.

[0004] The path planning method disclosed in the patent document with authorization announcement number CN116705679B only considers the distance between nodes to determine the optimal path. This method ignores the impact of other overhead cranes on the passage efficiency of the optimal path, resulting in poor passage efficiency of the optimal path and affecting the handling efficiency.

[0005] Patent document CN119692588A discloses another path planning method. This method considers the influence of other overhead cranes on each path and has high accuracy. However, this method requires calculating the time of multiple paths during path planning, which requires a lot of real-time calculations. This results in a higher computational cost, longer computation time and response speed. At the same time, in the actual operation of the overhead crane handling system, there are a lot of simultaneous path planning needs, which greatly increases the computing power requirements of the transport controller. Summary of the Invention

[0006] The purpose of this invention is to solve the above-mentioned problems existing in the prior art and to provide a path planning method, system and transportation controller.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] Path planning methods include the following steps:

[0009] S1, obtain the start and end points of the route to be planned;

[0010] S2, determine multiple travel routes from the starting point to multiple boundary points in the neighboring area and the first estimated travel time on each travel route through the neighboring area path planning model;

[0011] S3, determine the second estimated time from each of the boundary points to the endpoint by using the trained distant neighbor region time estimation model;

[0012] S4. Based on the first estimated travel time corresponding to each travel route and the second estimated travel time from the dividing point of the travel route to the destination, determine which travel route is selected as the optimal path.

[0013] Preferably, when exploring travel routes in the neighboring area using the neighboring area path planning model, if it is determined that the explored travel route includes the destination, then the travel route with the shortest estimated time from the starting point to the destination is selected as the optimal path.

[0014] If it is determined that the explored travel route does not include the destination, then multiple travel routes from the starting point to multiple boundary points in the neighboring area are determined, along with a first estimated travel time on each travel route.

[0015] Preferably, the nearest neighbor area path planning model adopts an improved Dijkstra algorithm, and the weight of the edge in the improved Dijkstra algorithm is determined based on the number of overhead cranes on the edge, whether the edge is located on a branch controlled by the area control unit, the number of overhead cranes in the contactless power supply area to which the edge belongs, and the number of available unloading machines between the two nodes of the edge.

[0016] Preferred,

[0017] A dividing point within the nearest neighbor area refers to the node closest to the crane after the crane has moved forward a predetermined distance from its starting point.

[0018] Or, a dividing point within the nearby area refers to the final node where the crane will be located after moving forward a predetermined number of nodes;

[0019] Or, a dividing point within the nearby area refers to the grid cell where the crane is located when it moves forward a predetermined number of grid cells.

[0020] Preferably, the time estimation model for the distant neighbor region is a deep neural network, which includes an input layer, an embedding layer, at least one hidden layer, and an output layer.

[0021] Preferably, the training data required for training the distant region time estimation model is obtained through simulation by a simulation system.

[0022] Preferably, the total estimated time for each route is determined according to the following formula, and the route with the shortest total estimated time is selected as the optimal path;

[0023] ;

[0024] Where t0 is the total estimated time, and t1 is the first estimated time for a single route; t2 is the discount factor; t2 is the second estimated time from the dividing point of the route to the destination.

[0025] Preferably, each time the overhead crane approaches or moves to the boundary point of the selected passageway, the boundary point is taken as the starting point, and S2-S4 are executed.

[0026] Path planning system, including:

[0027] The start and end point acquisition unit is used to obtain the start and end points of the route to be planned.

[0028] The first estimation unit is used to determine multiple travel routes from the starting point to multiple boundary points in the neighboring area and the first estimated travel time on each travel route through a neighboring area path planning model.

[0029] The second estimation unit is used to determine the second estimated time from each of the boundary points to the endpoint by using a trained distant neighbor region time estimation model;

[0030] The optimal path determination unit is used to determine which route to select as the optimal path based on the first estimated travel time corresponding to each determined route and the second estimated travel time from the boundary point of the route to the destination.

[0031] A transportation controller includes a memory and a processor, the memory storing a program executable by the processor, which, when executed, implements any of the path planning methods described above.

[0032] The advantages of the technical solution of this invention are mainly reflected in:

[0033] The path planning method of this invention considers the travel path of the overhead crane from the perspective of global traffic conditions. First, it uses the classic shortest path algorithm to determine the travel routes and travel times of the nearest neighboring areas to the overhead crane. Then, it uses a deep neural network to estimate the future congestion status and travel time of the far neighboring areas to directly participate in the auxiliary decision-making process to determine the optimal path of the nearest neighboring areas. The effective prediction of the far neighboring areas can avoid the inefficient and blind search in the far neighboring areas when executing Dijkstra's algorithm. It can greatly reduce the path planning time of the transportation controller and reduce the computational cost while ensuring the accuracy of path planning, thereby improving the efficiency of the transportation controller. It also reduces the number of times the transportation controller intervenes to a certain extent, and can guide the overhead crane away from congested areas in a timely manner, thereby improving the handling efficiency.

[0034] This invention employs an improved Dijkstra algorithm, which determines the weight of an edge based on factors such as the number of overhead cranes on the edge, whether the edge is located on a branch controlled by a regional control unit, the number of overhead cranes in the contactless power supply area to which the edge belongs, and the number of unloading machines at the edge. This effectively ensures accurate time estimation for neighboring areas, thus providing reliable data support for accurate planning of the optimal path.

[0035] This invention uses a simulation system to obtain the data required for model training, which can simulate as many road conditions as possible that may occur during actual operation, so that the time estimation model in the distant area can fully learn the various conditions on the track for subsequent accurate estimation. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the path planning method of the present invention;

[0037] Figure 2 This is a schematic diagram of the present invention, which determines the nearest neighbor region by moving forward a certain distance;

[0038] Figure 3 This is a schematic diagram of the present invention, which determines the nearest neighbor region by moving forward a predetermined number of nodes;

[0039] Figure 4 This is a schematic diagram of the present invention, which determines the nearest neighbor region by moving forward a predetermined number of grid cells. The area within the dashed box in the diagram represents the nearest neighbor region.

[0040] Figure 5 This is a schematic diagram of the state of the passage route determined by the nearest area path planning model in this invention at different times. The circles with arrows in the diagram represent overhead cranes.

[0041] Figure 6 This is a schematic diagram of an example of the present invention, showing multiple overhead cranes on one side. The circles with arrows in the diagram represent overhead cranes.

[0042] Figure 7 This is a schematic diagram of the structure of the distant neighbor area time estimation model of the present invention;

[0043] Figure 8 This is a schematic diagram illustrating the first path planning performed by the method of the present invention;

[0044] Figure 9 This is a schematic diagram illustrating the second path planning performed by the method of the present invention;

[0045] Figure 10a This is a schematic diagram of the optimal path determined by the method of the present invention in the first path planning of a grid track;

[0046] Figure 10bThis is a schematic diagram of the optimal path determined by the method of the present invention in a second path planning operation on a grid-like track;

[0047] Figure 10c This is a schematic diagram of the optimal path determined by the third path planning in a grid-like track using the method of the present invention;

[0048] Figure 10d This is a schematic diagram of the optimal path determined by the fourth path planning in a grid-like track using the method of the present invention;

[0049] Figure 10e This is a schematic diagram of the complete path determined by the method of the present invention through multiple path planning operations on a grid-like track. Detailed Implementation

[0050] The objectives, advantages, and features of this invention will be illustrated and explained through the following non-limiting description of preferred embodiments. These embodiments are merely typical examples of applying the technical solutions of this invention, and all technical solutions formed by equivalent substitutions or equivalent transformations fall within the scope of protection claimed by this invention.

[0051] In the description of the solution, it should be noted that the terms "center," "upper," "lower," "left," "right," "front," "rear," "vertical," "horizontal," "inner," and "outer," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience and simplification of description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Example 1

[0052] The path planning method disclosed in this invention will be described below with reference to the accompanying drawings. The path planning method can be used in various scenarios. For example, it can be used for path planning when an overhead crane needs to move from its current location to the location of a material to be transported; or it can be used for path planning when an overhead crane grabs a material and moves from its current location to the location where the material to be transported is to be transported; or it can be used for path planning when an overhead crane moves from its current location to a designated location to avoid other overhead cranes, etc.

[0053] Correspondingly, as shown in the appendix Figure 1 As shown, the path planning method includes the following steps:

[0054] S1, obtain the start and end points of the route to be planned;

[0055] S2, determine multiple travel routes from the starting point to multiple boundary points in the neighboring area and the first estimated travel time on each travel route through the neighboring area path planning model;

[0056] S3, determine the second estimated time from each of the boundary points to the endpoint by using the trained distant neighbor region time estimation model;

[0057] S4. Based on the first estimated travel time corresponding to each travel route and the second estimated travel time from the dividing point of the travel route to the destination, determine which travel route is selected as the optimal path.

[0058] In step S1, the starting point can be the current location of the overhead crane in each of the scenarios described above, or it can be the endpoint of the optimal path within the nearest neighbor region determined by the path planning described below. The endpoint is the location that the overhead crane needs to move to in each of the scenarios described above.

[0059] In step S2, the nearest neighbor region can be determined in different ways. For example, in an overhead crane transport system, the track may be an irregularly shaped track laid out according to the equipment layout, as shown in patent documents with publication numbers CN119671441A and CN119692588A, and the overhead crane travels unidirectionally on the track. Of course, in another embodiment, the track may also be a grid-like track, as shown in patent documents with authorization publication numbers CN112533813B and CN108698757B. Multiple travel routes may exist within the nearest neighbor region. The far neighbor region is the track area between the boundary point and the endpoint.

[0060] When the track is irregularly shaped, the nearest neighbor region can be understood as the track area between the starting point and a position in the forward direction that is a certain distance from the starting point. This certain distance can be, for example, 10m, 15m, 20m, 50m, etc., and can be determined according to specific needs; it is not limited here. Correspondingly, a dividing point within the nearest neighbor region refers to the node that the crane is closest to after moving forward a predetermined length from the starting point. Alternatively, it can be the position the crane has moved to. The predetermined length is the aforementioned certain distance.

[0061] For example, see attached Figure 2As shown, starting from node 1, the crane can move in two directions. After moving a predetermined distance in each direction, there are nodes 2 and 3. From node 2, the crane can continue moving in two directions towards nodes 4 and 5, and from node 3, it can continue moving forward in two directions towards nodes 5 and 6. When the crane moves from node 2 to node 4 to a certain position, and the distance the crane has traveled between node 1 and that certain position is equal to the predetermined length, it is determined that the distance from that certain position to node 4 is less than its distance to node 2, meaning that the certain position is closer to node 4. Therefore, node 4 is determined to be a boundary point within the nearest neighbor region. Correspondingly, the nearest neighbor region also includes two other boundary points, namely nodes 5 and 6.

[0062] Of course, when the track is irregularly shaped, the nearest neighbor region can also be understood as the track region between the starting point and the nth node downstream of the starting point in the forward direction. The nth node is, for example, the 3rd node, the 4th node, etc., and is determined according to needs, without limitation here. Correspondingly, a dividing point within the nearest neighbor region refers to the node where the crane finally reaches after moving forward a predetermined number of nodes.

[0063] For example, as shown in the appendix Figure 3 As shown, starting from node 1, there are two directions of travel. The predetermined number is, for example, 2. After the crane moves in the first direction, passing through nodes 2 and 4, the final node 4 is a boundary point within the nearest neighbor region. Correspondingly, the nearest neighbor region also includes two other boundary points, namely nodes 5 and 6.

[0064] When the track is a grid-like track, the nearest neighbor region is the track area between the starting point and the grid cell where the crane is located when it moves forward a predetermined number of grid cells from the starting point. Correspondingly, a boundary point within the nearest neighbor region refers to the grid cell where the crane is located when it moves forward a predetermined number of grid cells.

[0065] As attached Figure 4 As shown, when the crane is in grid cell 11, it can move in three directions. Assuming the predetermined number is 4, the grid cells where the crane is located when it moves 4 grid cells in the three directions are grid cells 33, grid cells 51 and grid cells 42, respectively. Correspondingly, the boundary points in the nearby area are grid cells 33, grid cells 51 and grid cells 42.

[0066] Furthermore, the predetermined length of forward movement, the predetermined number of nodes, and the predetermined number of grid cells are defined as hyperparameter K. The value of hyperparameter K determines the range of the nearest and far neighbor regions, which plays a crucial role in the entire path planning.

[0067] If the hyperparameter K is small, it means the nearest-neighbor area route planning model searches only a small range, resulting in lower computational cost. However, the optimal route in the nearest-neighbor area heavily relies on the second estimated time from the distant-neighbor area time estimation model. Since the distant-neighbor area time estimation model cannot grasp the real-time dynamic changes of the entire track map (track distribution map), the smaller the hyperparameter K value, the longer the distance between the boundary point and the destination in the distant-neighbor area. Estimating the time over long distances increases the uncertainty of route planning. Simultaneously, a small hyperparameter K will also lead to frequent execution of nearest-neighbor area route planning, increasing the frequency of route planning and placing higher demands on the response capability of the Transportation Controller (TCS).

[0068] Conversely, a large value for hyperparameter K makes the entire path planning more reliant on the path planning of neighboring regions. The larger the value of hyperparameter K, the larger the search area the neighboring region path planning model needs to cover, resulting in a higher computational cost. Furthermore, a larger value for hyperparameter K may lead to greater changes in the weights of edges far from the starting position over time, meaning the reliability of the weights of edges far from the starting position is lower.

[0069] As attached Figure 5 As shown, at time t1, the nearest neighbor path planning model plans a route from node 1001, passing through nodes 1002, 1017, 1018, and 1019. There is one vehicle on road segment 1 (1017->1018), but no vehicle on road segment 2 (1018->1019). Therefore, the edge weights of road segments 1 (1017->1018) and 2 (1018->1019) are relatively small.

[0070] At time t2, when the overhead crane travels from node 1001 to node 1002, one crane on segment 1 (1017->1018) is still unloading and hasn't left. Simultaneously, two other overhead cranes move to segment 1 (1017->1018), and two more move to segment 2 (1018->1019). At this point, the weights of the edges on segments 1 (1017->1018) and 2 (1018->1019) increase. Therefore, the optimal path determined at time t1 may not be a optimal path at time t2. Thus, the larger the value of the hyperparameter K, the greater the unreliability of the route planning model in the nearest neighbor region.

[0071] Therefore, the hyperparameter K needs to be selected with a suitable value. To obtain a suitable hyperparameter K, it can be determined in a simulation environment by trial runs of several nearest-neighbor area path planning models and far-neighbor area time estimation models with different hyperparameter Ks. The simulation environment, for example, uses a simulation system to generate transport commands, inputting the start and end points of the transport commands into the nearest-neighbor area path planning models and far-neighbor area time estimation models with different hyperparameter Ks for path planning. The simulation then simulates the overhead crane moving to the end point according to the paths planned by the two models, and records the transport time taken by the overhead crane from the start point to the end point during the simulation.

[0072] After each nearest-neighbor area path planning model corresponding to hyperparameter K has undergone a trial run for a predetermined period (the predetermined period can be determined as needed and is not limited here), the performance of each nearest-neighbor area path planning model and its corresponding far-neighbor area time estimation model can be measured by calculating the average transport time and total throughput (i.e., the total number of transport commands executed by a nearest-neighbor area path planning model of hyperparameter K within the aforementioned predetermined period) of all transport commands used for path planning using the nearest-neighbor area path planning model of each hyperparameter K. Then, the hyperparameter K with the best performance is selected as the target value. Of course, the optimal hyperparameter K varies for different track maps and needs to be determined experimentally in a simulation environment.

[0073] The neighborhood path planning model can use the known Dijkstra algorithm to explore the boundaries of the neighborhood, the routes from the starting point to each boundary, and the first estimated time of movement on each route.

[0074] More preferably, the nearest neighbor area path planning model adopts an improved Dijkstra algorithm. A major change in the improved Dijkstra algorithm is that it improves the calculation method of the weight of each edge based on the existing Dijkstra algorithm. That is, the weight of each edge is determined based on the number of overhead cranes on the edge, whether the edge is on a branch controlled by a ZCU (Zone Control Unit), the number of overhead cranes in the contactless power supply area to which the edge belongs, and the number of available unloading machines between the two nodes of the edge.

[0075] In detail, an edge is the track between two nodes in the track map. Each node can be a merging point, a fork point, or a designated location point in the track.

[0076] When an edge corresponds to a line with a crane, the weight of that edge needs to be determined by considering the potential delay that the crane on that edge might cause to other cranes that need to pass through that edge.

[0077] For example, see attached Figure 6 As shown, the road segment from node N5 to node N6 is defined as edge E56. There are two overhead cranes on edge E56. Therefore, when calculating the weight of edge E56, the node distance D between node N5 and node N6 on edge E56 can be used. 56 Or the travel time T for the overhead crane to move from node N5 to node N6 at a set speed. 56 In addition, there is the extra delay T caused by each crane on edge E56 to other cranes that need to pass through that edge. V That is, when calculating the weight of edge E56, the node distance D is used. 56 Or travel time T 56 Add 2T V .

[0078] Meanwhile, the maximum number of power supplies a single CPS (Contactless Power Supply) device can provide determines the total number of overhead cranes that can enter a contactless power supply area (the track area powered by a single CPS device). When the number of overhead cranes in a contactless power supply area reaches the upper limit warning value for that area, there is a risk of power shortage. In this case, other overhead cranes wishing to enter the contactless power supply area will be controlled to wait at the entrance or give way. Therefore, during path planning, when determining the weight of each edge in the contactless power supply area, it is necessary to use the node distance of the edge or the corresponding travel time of the edge plus the corresponding delay weight T. cps .

[0079] For example, if the number of overhead cranes in the contactless power supply area to which edge E56 belongs in the example above reaches the upper limit warning value for that contactless power supply area, then when determining the weight of edge E56, the node distance D is used. 56 Or travel time T 56 Add delay weight T cps .

[0080] Furthermore, the number of available unloading machines at each edge will significantly impact the passage of other overhead cranes. Therefore, when determining the weight of an edge, it is also necessary to consider the number N of available unloading machines between the two nodes corresponding to that edge. port And the additional delay T caused by the overhead crane picking up and unloading goods at each available loading / unloading station to other overhead cranes passing through. port To determine this. For example, there is a pick-up and unloading machine at edge E56, i.e., N. port =1, then when determining the weight of edge E56, the node distance D is used. 56 Or travel time T 56 Add N port *T port .

[0081] Additionally, the ZCU is used to control the overhead cranes to pass through the intersecting track segments in sequence. When a node of an edge belongs to an upstream or downstream end of an intersecting track segment controlled by a ZCU (i.e., the edge is on a branch controlled by the ZCU), the additional delay T generated when the ZCU controls the overhead cranes to pass through the intersecting track segment one by one needs to be considered. zcu That is, when determining the weight of an edge, if the edge is on a branch of an intersecting track segment managed by a ZCU, the weight of the edge is the node distance or travel time corresponding to the edge plus T. zcu .

[0082] Furthermore, the T V T cps T port T zcu These can be set to fixed values, or they can be derived from other system analysis methods; no restrictions are imposed here.

[0083] In S3, the distant neighbor region time estimation model can employ a known deep neural network (DNN), as shown in the attached figure. Figure 7 As shown, the deep neural network is, for example, a multilayer perceptron, which includes an input layer, an embedding layer, at least one hidden layer, and an output layer. The specific functions and working mechanisms of the input layer, embedding layer, hidden layer, and output layer are known technologies and not innovative in this invention, so they will not be elaborated here. Furthermore, the input layer contains two neurons, and the embedding layer and hidden layer are fully connected, meaning each neuron in the embedding layer is connected to each neuron in the hidden layer. Preferably, there are at least two hidden layers. When the track map is large or there are many nodes, more hidden layers and neurons can be considered to learn more details. During the training phase, if underfitting occurs, the number of hidden layers can be increased; if overfitting occurs, the number of hidden layers can be reduced or Dropout can be introduced to enhance the model's generalization ability. The hidden layers are fully connected to each other, and the hidden layer and output layer are also fully connected. The output layer has only one neuron, representing the prediction time between the boundary point and the destination. The model uses the Rectified Linear Unit (ReLU) activation function to calculate the output.

[0084] Furthermore, when the orbital information is particularly large, convolutional layers can be added between the embedding layer and the hidden layer to achieve dimensionality reduction, thereby reducing the problem of excessive computation during training.

[0085] To facilitate model training, this invention proposes a simulation system to improve the predictive capability of the time estimation model for distant regions. The simulation system consists of three parts: a transport controller (TCS) holding the algorithm, a higher-level simulation module for generating transport instructions, and a lower-level simulation system for simulating crane operation.

[0086] The host simulation module generates transport instructions by simulating the MCS (Material Control System) to generate transport instructions and send them to the TCS. These instructions include a start point, an end point, and a transport priority, with the start and end points randomly configured. Alternatively, the MCS can be used directly to send transport instructions to the TCS, replacing the host simulation module.

[0087] The TCS plans the path according to the existing path planning method based on the handling instructions and instructs the overhead crane to carry out the handling.

[0088] The lower-level simulation system for simulating crane operation uses software to simulate the crane's movement on the track. It includes a track section and a crane section, wherein the track section is based on actual... Figure 1 1. A simulated track map, including nodes, FOUP (front-opening wafer cassette) access ports, ZCU (Zone Control Unit), and CPS (Cyberpower System) power supply equipment. The crane section simulates the actual operation logic of the crane, such as the simulated crane traveling along the simulated track at a preset speed, the simulated crane slowing down or even stopping when encountering traffic congestion ahead, the simulated crane receiving traffic control from the ZCU at intersecting track sections, and receiving transport instructions from the TCS (Transport System).

[0089] The simulation system is used to obtain the data required for training the time estimation model for the distant region. Specifically, during the simulation, the starting point, ending point, road condition information before route planning, and the calculated travel time of the overhead crane from the starting point to the ending point for each handling instruction are recorded as a data sample. The road condition information includes edge information features and load information features. The edge information features refer to the number of overhead cranes on each edge, reflecting the current congestion situation on each edge. The load information features reflect the recent or future congestion situation related to the workload, specifically divided into three types: unassigned load (the number of handling tasks that have not yet been assigned to the target overhead crane), pickup load (the number of handling tasks that have been assigned and are on their way to pickup), and unloading load (the number of unloading tasks that have been assigned and are currently being carried out by overhead cranes to the designated destination).

[0090] After a period of simulation to obtain sufficient data samples for training, the training data is fed into the distant neighbor area time estimation model for training. The origin, destination, and road condition information of each training sample are fed into the input layer, and the travel time is used as the output layer data.

[0091] After a period of training, the deep neural network can fully learn the relationship between various road condition information and travel time. When the time estimation model of the distant neighbor area tends to stabilize, it is considered that the model has converged, and the training process can be stopped. The specific training process of the time estimation model of the distant neighbor area is a known technique and is not an innovation of this invention, so it will not be described in detail here. When updating the module parameters during training, Adam is preferably selected as the optimizer to achieve adaptive adjustment of the learning rate and prevent overfitting.

[0092] Once the distant neighbor time estimation model stabilizes, the travel time generated by the simulation system can be further used to test the model. For example, the simulated travel time from the starting point to the destination can be compared with the estimated time between the starting point and the destination calculated by the distant neighbor time estimation model. During testing, the structural parameters in the distant neighbor time estimation model can still be adjusted, such as the number of hidden layers and the number of neurons in each hidden layer. The adjusted model can then be further trained. Ultimately, a DNN model with strong learning ability and stability is obtained as the distant neighbor time estimation model.

[0093] The performance of the distant area time estimation model is affected by whether various scenarios occur during training. For scenarios that do not occur during training, the model may perform poorly in actual decision-making. However, this invention can simulate various situations encountered during the operation of the overhead crane through a simulation system, thereby enabling the distant area time estimation model to learn the degree of congestion under various road conditions and thus improve the prediction accuracy of the distant area time estimation model.

[0094] Of course, the data required for training the time estimation model in the distant region can also be collected from the stored historical transport log data, which will not be elaborated here.

[0095] Furthermore, in order to better adapt to changes in actual road conditions, it is necessary to continuously improve the time estimation model for neighboring areas so that it can learn new rail traffic conditions.

[0096] Therefore, this invention employs an incremental training method, which continuously updates the model parameters using small batches of real-time data during actual operation. For example, at regular intervals or when the transport density is low, recent historical transport records are randomly selected to update the model parameters.

[0097] When determining the second estimated time using a trained neighboring region time estimation model, the current road condition information (edge ​​information features, load information features) for each determined boundary point, endpoint, and route planning can be input into the trained neighboring region time estimation model. The neighboring region time estimation model predicts the second estimated time from each boundary point to the endpoint. Since the neighboring region time estimation model only needs to estimate time and does not need to perform route planning, it can greatly reduce unnecessary route exploration and computational costs.

[0098] In step S4, the total estimated time for each travel route is determined according to the following formula, and the travel route with the shortest total estimated time is selected as the optimal path.

[0099] ;

[0100] Where t0 is the total estimated time, and t1 is the first estimated time for a single route; The discount factor is set because the time estimation model for distant areas is inherently a prediction of future congestion in areas far from the starting point, which has a certain degree of uncertainty, so a discount factor needs to be added; t2 is the second estimated time from the dividing point of the travel route to the destination.

[0101] The discount factor can be determined based on different situations. When the accuracy of the time estimation model for the distant neighboring region is high, it can be considered... ≥1; When the time estimation model for distant areas is in its initial stage or the prediction is not very stable, it can be considered More preferably, the discount factor can be determined as follows: After training the distant neighbor time estimation model, historical data or data obtained from simulation by a simulation system are re-input into the determined nearby neighbor path planning model and distant neighbor time estimation model, and the discount factor is determined by grid search. For example, in the range of 0.5 to 1.5, grid search is performed with a step size of 0.1. The specific process of grid search is a known technique and will not be elaborated here.

[0102] As attached Figure 1 As shown, in actual route planning, there may be situations where the distance between the starting point or dividing point and the destination is very short. Especially after multiple route plannings, the distance between the dividing point (starting point) of the latest selected route and the destination is already very short. Therefore, in S2, when exploring routes in the neighboring area through the neighboring area route planning model, if it is determined that the explored route includes the destination, the route with the shortest estimated time from the starting point to the destination is selected as the optimal route, and the route planning can be terminated.

[0103] If it is determined that the explored travel route does not include the destination, then multiple travel routes from the starting point to multiple boundary points in the neighboring area and the first estimated time of movement on each travel route are determined, and S3 is executed.

[0104] As attached Figure 1 As shown, after one path planning, the overhead crane moves according to the selected optimal path. Each time the overhead crane moves close to or to the boundary point of the selected optimal path, the boundary point is used as the starting point, and the nearest neighbor area is determined again according to the same hyperparameter K as mentioned above. Then, S2-S4 are executed. That is, each time the crane moves close to or to the boundary point of the optimal path, the boundary point is used as the starting point to perform path planning again, thereby continuously planning new optimal paths.

[0105] The following example illustrates how to determine the optimal path during each path planning process:

[0106] As attached Figure 8 As shown, a transport task starts at node 1 and ends at node 99. The nearest neighbor region is determined by moving forward 20 meters using hyperparameter K. In each path planning, the node closest to the location of the crane after moving forward 20 meters from the starting point is used as the dividing point. The dividing points reached after moving forward 20 meters from node 1 are node 4, node 5, and node 6. The travel routes estimated by the nearest neighbor region path planning model, the first estimated time for each travel route, the second estimated time determined by the distant neighbor region time estimation model, and the total estimated time are shown in Table 1. The discount factor a is 1.

[0107] Table 1

[0108]

[0109] As shown in Table 1, starting from node 1, four routes were explored: route 1 (1→2→4), route 2 (1→2→5), route 3 (1→3→5), and route 4 (1→3→6). The estimated first travel times for these four routes, calculated using the nearest-neighbor area path planning model, are 7s, 8s, 9s, and 10s, respectively. The estimated second travel times from nodes 4, 5, and 6 to destination 99, predicted by the distant-neighbor area time estimation model, are 50s, 40s, and 60s, respectively.

[0110] Therefore, during the first path planning, the estimated total time for the four routes in the neighboring area is as follows: route 1 has an estimated total time of 57 seconds, route 2 has an estimated total time of 48 seconds, route 3 has an estimated total time of 49 seconds, and route 4 has an estimated total time of 70 seconds. Therefore, when determining the optimal route for the first path planning, route 2 will be selected as the optimal route.

[0111] During the crane's journey along route 2, when the crane is about to move to node 5 or has already moved to node 5, the transport controller needs to plan the next stage of the crane's journey route again. (See attached...) Figure 9 As shown, starting from node 5, after traveling forward 20 meters, you can reach nodes 9 and 10. At this point, nodes 9 and 10 are the newly determined boundaries of the neighboring regions. The travel routes determined by the neighboring region path planning model, the first estimated time for each travel route, the second estimated time determined by the distant region time estimation model, and the total estimated time are shown in Table 2. The discount factor a is 1.

[0112] Table 2

[0113]

[0114] As shown in Table 2, starting from node 5, two routes were explored: route 5 (5→7→9) and route 6 (5→8→10). The nearest-neighbor area path planning model calculated the first estimated time for route 5 to be 7 seconds, and the distant-neighbor area time estimation model determined the second estimated time from node 9 to destination 99 to be 35 seconds, for a total estimated time of 42 seconds. Correspondingly, route 6 has a first estimated time of 10 seconds, a second estimated time of 33 seconds, and a total estimated time of 43 seconds. Therefore, in this path planning, route 5 is selected as the optimal route for the overhead crane.

[0115] Each time the overhead crane moves close to or to the end point (boundary point) of the optimal path, the above path planning process is repeated until the end point is reached.

[0116] As attached Figures 10a-10e As shown in the example of a 9x9 grid-based track map, each grid cell is numbered according to its corresponding x-coordinate and y-coordinate combination. The crane needs to move from the starting position (grid cell 11 in the lower left corner) to the ending position (grid cell 99 in the upper right corner), advancing four grid cells at a time. In the first path planning, the optimal path determined by the crane from grid cell 11 ends at grid cell 33. In the second path planning, the optimal path determined by continuing path planning from grid cell 33 ends at grid cell 64. Similarly, in the third path planning, the optimal path determined by continuing path planning from grid cell 64 ends at grid cell 77. In the fourth decision, the optimal path determined by continuing path planning from grid cell 77 ends at grid cell 99, and path planning stops. Example 2

[0117] This embodiment discloses a path planning system, including:

[0118] The start and end point acquisition unit is used to obtain the start and end points of the route to be planned.

[0119] The first estimation unit is used to determine multiple travel routes from the starting point to multiple boundary points in the neighboring area and the first estimated travel time on each travel route through a neighboring area path planning model.

[0120] The second estimation unit is used to determine the second estimated time from each of the boundary points to the endpoint by using a trained distant neighbor region time estimation model;

[0121] The optimal path determination unit is used to determine which route to select as the optimal path based on the first estimated travel time corresponding to each determined route and the second estimated travel time from the boundary point of the route to the destination. Example 3

[0122] This embodiment discloses a transportation controller, including a memory and a processor. The memory stores a program that can be executed by the processor. When the program is executed, it implements any of the path planning methods described above.

[0123] This invention has many other embodiments, and all technical solutions formed by equivalent transformation or equivalent transformation fall within the protection scope of this invention.

Claims

1. A path planning method, characterized in that, Includes the following steps: S1, obtain the start and end points of the route to be planned; S2, using a neighborhood path planning model, multiple travel routes from the starting point to multiple boundary points within the neighborhood area are determined, along with the first estimated travel time on each route. The extent of the neighborhood area is determined based on a hyperparameter K, which is a predetermined length for the crane to move forward, a predetermined number of nodes for the crane to move forward, or a predetermined number of grid cells for the crane to move forward. In S2, when exploring travel routes within the neighborhood area using the neighborhood path planning model, if it is determined that the explored travel route includes the destination, the travel route with the shortest estimated travel time from the starting point to the destination is selected as the optimal path. If it is determined that the explored route does not include the destination, then multiple routes from the starting point to multiple boundary points in the neighboring area are determined, and a first estimated time of movement on each route is determined. S3, determine the second estimated time from each of the boundary points to the endpoint by using the trained distant neighbor region time estimation model; S4, determine which route to select as the optimal path based on the first estimated travel time corresponding to each route and the second estimated travel time from the boundary point of the route to the destination; determine the total estimated travel time corresponding to each route according to the following formula, and select the route with the shortest total estimated travel time as the optimal path. ； Where t0 is the total estimated time, and t1 is the first estimated time for a single route; t2 is the discount factor; t2 is the second estimated time from the dividing point of the route to the destination.

2. The path planning method according to claim 1, characterized in that: The nearest neighbor area path planning model adopts an improved Dijkstra algorithm. The weight of the edge in the improved Dijkstra algorithm is determined based on the number of overhead cranes on the edge, whether the edge is located on a branch controlled by the area control unit, the number of overhead cranes in the contactless power supply area to which the edge belongs, and the number of available unloading machines between the two nodes of the edge.

3. The path planning method according to claim 1, characterized in that: A dividing point within the nearest neighbor area refers to the node closest to the crane after the crane has moved forward a predetermined distance from its starting point. Or, a dividing point within the nearby area refers to the final node where the crane will be located after moving forward a predetermined number of nodes; Or, a dividing point within the nearby area refers to the grid cell where the crane is located when it moves forward a predetermined number of grid cells.

4. The path planning method according to claim 1, characterized in that: The time estimation model for the distant neighbor region is a deep neural network, which includes an input layer, an embedding layer, at least one hidden layer, and an output layer.

5. The path planning method according to claim 1, characterized in that: The training data required for training the time estimation model of the distant neighboring region is obtained through simulation by a simulation system.

6. The path planning method according to any one of claims 1-5, characterized in that: Each time the overhead crane moves close to or to the boundary point of the selected passageway, the boundary point is used as the starting point, and S2-S4 are executed.

7. A path planning system, characterized in that, include: The start and end point acquisition unit is used to obtain the start and end points of the route to be planned. The first estimation unit is used to determine multiple travel routes from the starting point to multiple boundary points in the neighboring area and the first estimated travel time on each travel route through a neighboring area path planning model. When exploring travel routes within a neighboring region using a neighboring region path planning model, if it is determined that the explored travel route includes a destination, then the travel route with the shortest estimated time from the starting point to the destination is selected as the optimal path. If it is determined that the explored travel route does not include the destination, then multiple travel routes from the starting point to multiple boundary points in the neighboring area are determined, and the first estimated time for moving on each travel route is determined; the range of the neighboring area is determined according to the hyperparameter K, which is a predetermined length for the crane to move forward, a predetermined number of nodes for the crane to move forward, or a predetermined number of grid cells for the crane to move forward. The second estimation unit is used to determine the second estimated time from each of the boundary points to the endpoint by using a trained distant neighbor region time estimation model; The optimal path determination unit is used to determine which route to select as the optimal path based on the first estimated time corresponding to each route and the second estimated time from the boundary point of the route to the destination; and to determine the total estimated time corresponding to each route according to the following formula, and select the route with the shortest total estimated time as the optimal path. ； Where t0 is the total estimated time, and t1 is the first estimated time for a single route; t2 is the discount factor; t2 is the second estimated time from the dividing point of the route to the destination.

8. A transport controller, comprising a memory and a processor, the memory storing a program executable by the processor, characterized in that: When the program is executed, it implements the path planning method as described in any one of claims 1-6.

Images