Warehouse shuttle vehicle path planning method and system based on multi-source coupling collaborative optimization

Abstract

The application belongs to the technical field of intelligent warehousing logistics, and particularly relates to a warehouse shuttle vehicle path planning method and system based on multi-source coupling collaborative optimization. The method firstly quantifies the dynamic urgency weight by fusing task chain logic and space-time graph neural network congestion prediction; then a mixed integer programming model is constructed with the minimum weighted total completion time as the target and space-time mutual exclusion as the constraint, and a conflict-free space-time corridor is solved at one time to realize conflict pre-dissolution and global optimization collaboration; the system stability is monitored online by designing a Lyapunov function based on task weight, and a rapid local adjustment is triggered when the risk exceeds the threshold; finally, the meta-learner in the digital twin environment is used to automatically optimize the urgency fusion function and the target model parameters to form a closed-loop evolution. The application is based on multi-mechanism deep coupling, which significantly improves the throughput efficiency, robustness and intelligent level of the multi-vehicle system in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehousing and logistics technology, and in particular to a method and system for planning the path of a warehouse shuttle based on multi-source coupling and collaborative optimization. Background Technology

[0002] In highly automated modern warehousing systems, the efficient collaborative scheduling of multiple shuttle vehicles (or multiple AGVs) is key to improving operational efficiency. Existing technologies mainly suffer from problems at three levels:

[0003] (1) Static decision-making basis: Most schemes (such as patent CN113031603B) rely on limited static attributes (such as order timeliness and power consumption) for task priority calculation, and fail to incorporate the temporal dependencies between tasks (task chain) and the real-time evolution of global traffic situation prediction into the decision-making basis, resulting in insufficient scheduling foresight.

[0004] (2) Conflict resolution is disconnected from overall optimization: Traditional conflict search algorithms (CBS, such as those used in patent CN118706124A) take the elimination of immediate conflicts as their primary goal. Their cost function design fails to form a closed loop with system-level optimization goals such as "maximizing the long-term throughput efficiency of the system" or "minimizing the average latency of high-value tasks".

[0005] (3) Loose coupling between learning and optimization modules: Although artificial intelligence and digital twin technology have been introduced, the learning model (such as reinforcement learning agent) often provides auxiliary suggestions as an independent module and fails to be deeply coupled with the mathematical optimization model of the core scheduler to form an integrated self-consistent system of "perception-decision-verification-evolution", resulting in low learning efficiency and limited adaptability to dynamic environments. Summary of the Invention

[0006] This invention proposes a warehouse shuttle route planning method and system based on multi-source coupling and collaborative optimization, aiming to solve the technical problems in the prior art, such as incomplete scheduling decision basis, inconsistency between local conflict resolution and the overall system optimization goal, and the bottleneck of system dynamic adaptability and overall efficiency caused by loose coupling of learning optimization mechanism.

[0007] In a first aspect, embodiments of the present invention provide a warehouse shuttle route planning method based on multi-source coupling and collaborative optimization, including:

[0008] S1 receives a set of tasks, identifies the temporal dependencies between tasks to construct a task chain, and calculates the dynamic urgency weight of each task by fusing the inherent urgency of the task, the urgency propagation factor of the task chain, and the predicted congestion coefficient through a nonlinear fusion function.

[0009] S2, with the optimization objective of minimizing the weighted completion time of all tasks, the weights are the dynamic urgency weights, and a mixed integer nonlinear programming model is constructed to solve the problem using shuttle motion constraints, static obstacle constraints, and shuttle spacetime mutual exclusion constraints as hard constraints, and a conflict-free spacetime corridor is assigned to each vehicle.

[0010] S3, during the execution of the spatiotemporal corridor, a Lyapunov function is constructed based on the deviation between the actual state and the planned state of the shuttle to assess the stability risk; when the risk exceeds a preset threshold, a fast local adjustment algorithm is triggered to adjust the spatiotemporal corridor parameters of the affected shuttle to reduce the stability risk.

[0011] S4. Run a meta-learner in a digital twin environment. The meta-learner automatically adjusts the parameters of the nonlinear fusion function and the weight coefficients of the optimization target by using simulation data and historical data to optimize long-term scheduling performance indicators.

[0012] The technical advantages of the warehouse shuttle path planning method based on multi-source coupling and collaborative optimization disclosed in this invention are: it directly integrates conflict resolution (achieved through constraints) with system-level efficiency optimization (achieved through an objective function), proactively avoiding conflicts from the planning stage and searching towards optimal performance. Through an online intervention mechanism based on stability theory, it can maintain operation with minimal disturbance and prevent deadlock. Finally, through online and automatic optimization and adjustment of core model parameters by a meta-learner, it achieves a leap from "automation based on fixed rules" to "intelligentization with learning and evolution capabilities," enabling the system to adapt to changes in specific warehouse operation modes and long-term efficiency improvement needs.

[0013] Furthermore, S1 also includes: receiving a set of tasks, identifying the temporal dependencies between tasks to construct a task chain, and running a congestion propagation prediction model based on historical and real-time path data. The construction and training process of the congestion propagation prediction model includes: modeling the warehouse map as a spatiotemporal graph, with node features including historical shuttle density and speed, and edge features representing connection relationships and distances; using a gated spatiotemporal graph convolutional network as the model architecture; and outputting a predicted congestion coefficient.

[0014] During training, the node states of historical continuous time slices are used as input, and the actual shuttle density of future time slices is used as labels. The network parameters are trained by minimizing the prediction error loss function.

[0015] Furthermore, in S1, the task chain urgency propagation factor η i The calculation method is as follows:

[0016] ;

[0017] in, The latest allowed completion time for the chain containing the task. λ represents the current deadline for the task, and λ is the preset decay coefficient.

[0018] Furthermore, in S1, the nonlinear fusion function is specifically: ;

[0019] Where, ω i The dynamic urgency weights for task i, Urgency i and Value i These are the inherent urgency and value coefficient of the task, respectively. Let α, β, γ, δ, ε be the predicted congestion coefficients for path i, where α, β, γ, δ, ε are positive parameters to be optimized.

[0020] Furthermore, in S2, the spatiotemporal mutual exclusion constraint of the shuttle workshop is modeled using the Big M method: for any two different shuttles v1 and v2, at any time t and spatial node p, such that x v1,p,t +x v2,p,t ≤1 holds true, where x v1,p,t and x v2,p,t These are binary decision variables representing the positions occupied by shuttle cars v1 and v2, respectively.

[0021] Furthermore, in S2, the branch and bound method is used to solve the mixed integer nonlinear programming model, and a heuristic cutting plane based on conflict constraints is added during the solution process to accelerate the solution.

[0022] Furthermore, in S3, the Lyapunov function V(e) is defined as:

[0023] ;

[0024] Among them, e T Let P be the transpose of the matrix, e be the deviation vector between the actual and planned positions of all shuttles, and P be a diagonal positive definite matrix. The diagonal elements p in P are... i The dynamic urgency weight ω of the corresponding shuttle vehicle performing the task i Positive correlation; the stability risk threshold is determined based on the following formula:

[0025] ;

[0026] Where α and ξ are positive constants. Let V(e) be the time derivative.

[0027] Furthermore, the fast local adjustment algorithm in S3 specifically involves: keeping the spatiotemporal corridor of shuttles with dynamic urgency weights higher than the first threshold unchanged; for other affected shuttles, within a finite time delay adjustment window, solving a linear programming subproblem aimed at minimizing the derivative of the Lyapunov function, and reallocating their time to pass through the conflict area or selecting alternative paths.

[0028] Furthermore, in step S4, the meta-learner adopts a two-layer optimization framework: the inner layer evaluates the scheduling performance J(θ) in the digital twin under given model parameters θ; the outer layer updates θ through gradient optimization to minimize the long-term expected performance loss E[J(θ)]; the model parameters θ include at least the weight parameters (α,β,γ,δ,ε) of the nonlinear fusion function in step S1 and the regularization coefficient of the optimization objective function in step S2.

[0029] Secondly, embodiments of the present invention provide a warehouse shuttle route planning system based on multi-source coupling and collaborative optimization, used to implement the method described above, the system comprising:

[0030] The intelligent scheduling host integrates a multi-source information fusion computing unit, a spatiotemporal collaborative optimization solver, and an online stability monitoring and adjustment module.

[0031] The digital twin and meta-learning server runs a high-fidelity simulation environment and a meta-learning optimizer; multiple vehicle control and communication units are deployed on each shuttle vehicle.

[0032] A high-speed industrial network connects the intelligent scheduling host, the digital twin and meta-learning server, and each of the vehicle control and communication units.

[0033] The technical effect of the warehouse shuttle path planning system based on multi-source coupling and collaborative optimization disclosed in this invention is as follows: by deeply coupling the intelligent scheduling host, the digital twin evolutionary engine and the vehicle unit, a hardware-in-the-loop intelligent agent with a "decision-simulation-learning" closed loop is constructed, realizing the leap from static preset to dynamic adaptation of the scheduling strategy; enabling the entire multi-vehicle system to continuously perceive performance feedback and self-optimize core model parameters in actual operation, thereby obtaining continuously improved operational efficiency and robustness for specific warehouse scenarios. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the warehouse shuttle route planning method based on multi-source coupling and collaborative optimization provided in an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram illustrating the path planning and conflict pre-resolution principle provided in an embodiment of the present invention;

[0036] Figure 3This is a schematic diagram of the closed loop of meta-learning optimization and model parameter self-update provided in an embodiment of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0038] This invention provides a warehouse shuttle route planning method based on multi-source coupling and collaborative optimization, referencing... Figures 1 to 3 As shown, the specific steps include:

[0039] Step S1: Multi-source information fusion and dynamic urgency quantification: Receive a set of tasks, identify the temporal dependencies between tasks to construct a task chain, and run a congestion propagation prediction model based on historical and real-time path data; fuse the inherent urgency of tasks, the urgency propagation factor of the task chain, and the predicted congestion coefficient, and calculate the dynamic urgency weight of each task through a nonlinear fusion function.

[0040] (1) Receive a batch of task sets C={1,2,...,n}. Use association rule algorithm to analyze historical task order data, identify frequently co-occurring task patterns, and combine them with the batch number of real-time orders to dynamically construct a task chain L with strong temporal dependencies. For task i in task chain L, the task chain urgency propagation factor η i The calculation method is as follows:

[0041] ;

[0042] in, The latest allowed completion time for the chain containing the task. λ represents the current deadline for the task, and λ is the preset decay coefficient. This difference measures the remaining buffer time between the completion time of the current task i and the time needed to ensure the timely completion of the entire chain. The smaller the difference, the less buffer time there is, and the higher the urgency.

[0043] In this process, nonlinear amplification is achieved through an exponential function. When the remaining buffer time (difference) decreases to a certain critical point, the factor η... iThe value increases dramatically, rather than linearly. This sends a strong signal to the entire scheduling system—when a task chain approaches its deadline, the urgency of all remaining tasks on the chain is significantly increased, even if their static attributes (such as the value of the goods) are not high. The system will prioritize resource allocation to ensure the smooth operation of the chain, preventing delays in one link from causing delays in the entire process.

[0044] λ is the decay coefficient, used to control the intensity and gradient of urgency as it propagates forward along the task chain. The larger the value of λ, the faster the urgency decays as it moves away from the final node, emphasizing the pressure of the "deadline"; the smaller the value of λ, the smoother the propagation of urgency in the chain, emphasizing the synergy of the entire process.

[0045] The task chain urgency propagation factor reflects how the urgency of the tail task propagates to the preceding tasks. In other words, when a delay risk occurs at a certain link in the task chain, the scheduling priority of subsequent related tasks is increased in advance, proactively preventing cascading failures and system deadlocks. Furthermore, by introducing the task chain urgency propagation factor, the scheduling optimization objective is elevated from "minimizing the delay of a single task" to "minimizing the delay of critical processes," which is more in line with the realities of modern flexible manufacturing and logistics.

[0046] In a specific implementation scenario, consider a production replenishment chain: Task A (material pickup, deadline 10:00) → Task B (material delivery, deadline 10:05) → Task C (return, deadline 10:10). DL max The ratio is 10:10.

[0047] At 9:50: the remaining buffer time for each task is relatively long, and the calculated η i The values ​​are all close to 1 (e.g., 0.9). At this point, the scheduling system mainly allocates shuttles based on the urgency and value of the task itself, and the impact of the task chain is relatively weak.

[0048] At 10:03 (Task B should have been completed but has not yet started): The remaining buffer time for Task B has decreased sharply (DL). max -DL B (reduced from 5 minutes to 2 minutes), its η B The value could jump from 0.9 to 1.5 (non-linear amplification). Meanwhile, the η of the task C, which has not yet begun,... C The value will also increase accordingly.

[0049] Scheduling decision: At this point, even if a high-value "isolated" task D appears, when the system calculates the dynamic urgency weight, it will prioritize scheduling the shuttle to repair task B because of the amplified chain propagation factor η of tasks B and C, and reserve resources for task C, thereby ensuring the continuous operation of the entire production line and avoiding production line shutdown due to material supply interruption.

[0050] (2) Run the congestion propagation prediction model. The model uses a gated spatiotemporal graph convolutional network. Its input is the full graph node features (number of shuttles, average speed) of the historical T consecutive time slices and the currently known future K-step path plan. The network outputs the predicted traffic impedance coefficient of each node in the next H time steps, which is the predicted congestion coefficient ρ (0≤ρ≤1). The closer ρ is to 1, the more congested it is.

[0051] Ultimately, the dynamic urgency weight ω of task i i Calculated by the following nonlinear fusion function: ;

[0052] Where, ω i The dynamic urgency weights for task i, Urgency i and Value i These are the inherent urgency and value coefficient of the task, respectively. Let α, β, γ, δ, and ε be the predicted congestion coefficients for path i, and let α, β, γ, δ, and ε be the positive parameters to be optimized. Where Urgency i Based on the order type or promised completion time corresponding to the task, orders can be marked as "urgent," "normal," "economy," etc., and mapped to different numerical values ​​(e.g., Urgency). i ∈ {2.0, 1.0, 0.5}). Urgency i The value coefficient is determined based on the value of the goods transported in the task, the corresponding customer level, or the importance of the production line, and is normalized (e.g., Urgency). i ∈[0,1]), used to reflect the economic value or operational importance of the task. η i The urgency propagation factor of the chain; The function estimates the average predicted traffic impedance coefficient for each node on the path of task i; α, β, γ, δ, and ε are adjustable positive parameters. This function clearly expresses that the higher the fundamental value of a task and the more urgent its task chain, the greater its weight; conversely, the more congested the path, the less its weight. The technical effect of this step is to generate a precise quantitative scheduling guidance index that integrates static attributes, dynamic correlations, and forward-looking predictions.

[0053] In a specific implementation scenario:

[0054] 1. Task Background:

[0055] T123 is a picking task for a "same-day delivery" customer's order. The order needs to be retrieved from shelf A and delivered to packing station P.

[0056] This order belongs to a combined order chain: Task T122 (picking item A) → Task T123 (picking item B) → Task T124 (packing and shipping). The order can only be delivered after all tasks in this chain are completed.

[0057] Current system time: 10:00.

[0058] Order commitment delivery time (i.e., the latest completion time of the blockchain DL) max ): 12:00.

[0059] 2. Model parameter presets (to simplify calculations, assume the initial weight coefficients are as follows):

[0060] α=1.0 (urgency weight), β=0.8 (value coefficient weight), γ=1.5 (task chain factor weight), δ=1.0 (congestion adjustment base), ε=2.0 (congestion penalty weight).

[0061] First, based on the attributes of task T123, we can obtain:

[0062] Urgency i (Inherent urgency of the task): Assign a value of 1.5 based on the "same day delivery" type (assuming the following rules: normal = 1.0, same day delivery = 1.5, expedited = 2.0).

[0063] Value i (Task Value Coefficient): The total value of the goods in this order is relatively high. After calculation and normalization, a value coefficient of 0.9 (range 0-1) is assigned.

[0064] η i (Task chain urgency propagation factor):

[0065] The current deadline for task T123 is DL. i By DL max The time was set backwards to 11:30.

[0066] Calculate the difference: DL max -DL i =12:00-11:30= 30 minutes.

[0067] Substitute into the formula Set the attenuation coefficient λ = 0.05. Calculate: η i =exp(-0.05*30)=exp(-1.5)≈0.223.

[0068] This indicates that, since there is still some buffer before the chain ends, the urgency of the chain is of moderate to low pressure on the current task.

[0069] (Path Average Predicted Traffic Impedance): Based on the current full-path planning, the Congestion Propagation Prediction Model (GSTGCN) predicts that the planned path of Task T123 will experience average congestion within the next 30 minutes. The output... =0.4 (range 0-1, 0.4 indicates mild congestion).

[0070] Secondly, substitute all the above values ​​into the nonlinear fusion function:

[0071] (1) Calculate the first part of the numerator (basic urgency): α∙Urgency_i+β∙Value_i=1.0*1.5+0.8*0.9=1.5+0.72=2.22.

[0072] (2) Calculate the second part of the molecule (chain urgency amplification factor):

[0073] 1+γ∙η i =1+1.5*0.223=1+0.3345=1.3345.

[0074] (3) Calculate the numerator (comprehensive positive excitation): 2.22*1.3345≈2.962.

[0075] (4) Calculate the denominator (congestion penalty term):

[0076] =1.0 + 2.0 * 0.4 = 1.0 + 0.8 = 1.8.

[0077] (5) Calculate the final dynamic urgency weight ω i =2.962 / 1.8≈1.645.

[0078] Suppose an isolated task with equal urgency (1.5) and value (0.9) (no task chain, i.e., η_i=0), and its path is completely unobstructed. =0.1).

[0079] The calculation is: (1.5+0.72)*(1+0) / (1+2.0*0.1)=2.22 / 1.2≈1.85.

[0080] Comparison reveals that despite the task chain advantage, T123's higher path prediction congestion (0.4 vs 0.1) significantly lowers its final weight (1.645 < 1.85). Therefore, it can be concluded that when another task in the system with a higher weight (e.g., ω) experiences this congestion, the final weight will be significantly lower. j When the value of 2.0 appears, it will gain a stronger optimization "pull". When the paths of two tasks conflict in time and space, the optimization model will tend to "sacrifice" a small amount of time of the task with lower weight (T123) in exchange for a greater time saving for the task with higher weight, thereby achieving the global objective optimization.

[0081] Step S2: Path planning and conflict pre-resolution based on spatiotemporal collaborative optimization model: The optimization objective is to minimize the weighted completion time of all tasks, where the weight is the dynamic urgency weight. A mixed integer nonlinear programming model is constructed to solve the problem, with shuttle motion constraints, static obstacle constraints, and spatiotemporal mutual exclusion constraints between shuttles as hard constraints, and each vehicle is assigned a conflict-free spatiotemporal corridor.

[0082] This step formalizes multi-vehicle routing into a mixed-integer nonlinear programming problem:

[0083] Decision variable: Define a binary variable x for each shuttle vehicle v. v,p,t This indicates whether shuttle v occupies path node p at time t; simultaneously, a continuous variable t is defined. vstart and t vend This indicates the start and end times of the task.

[0084] Objective function: Minimize the weighted total completion time of the system.

[0085] ;

[0086] Where ω i Let v(i) be the dynamic urgency weight of task i, and v(i) be the shuttle that executes the task. This objective directly drives the system to prioritize high-weight tasks.

[0087] The constraints include:

[0088] 1. Flow conservation constraint: ensures path continuity.

[0089] 2. Shuttle dynamics constraints: maximum speed and acceleration limits.

[0090] 3. Spatiotemporal Mutual Exclusion Constraint (Core): For any two shuttles v1 and v2, and any time t and spatial node p, the Big M method is introduced to construct constraints: x v1,p,t +x v2,p,t ≤1+M∙b, and ensures through auxiliary variables and constraints that there will be no "brush-by" conflicts in space and time. This is equivalent to planning a safe spacetime corridor for each vehicle.

[0091] 4. Time Window Constraints: There are temporal constraints between tasks in a task chain. This model is solved using a branch and bound method combined with a heuristic cutting plane technique. The initial solution can be provided by a fast heuristic algorithm (such as A* with resource reservation). The solver outputs the execution time window and spatial path assigned to each task, i.e., the spacetime corridor.

[0092] By using a unified mathematical model, conflict resolution (achieved through constraints) and system-level efficiency optimization (achieved through objective functions) are integrated, proactively avoiding conflicts from the planning stage and directly searching for optimal system performance.

[0093] Specifically, A* with resource reservations includes:

[0094] For each shuttle's assigned task sequence, an A* algorithm incorporating time windows is used, specifically a spatiotemporal A* algorithm that incorporates dynamic urgency weights as heuristic factors. Starting from the current pose, paths to each subtask's target point (pickup point, unloading point) are planned sequentially. During planning, the system maintains a global spatiotemporal reservation table. For high-priority tasks currently being planned (e.g., those with a score higher than a threshold θ), the algorithm directly writes its planned path unit (position + time) into the reservation table. For subsequent low-priority tasks, a "conflict penalty term" is added to the cost function of the A* algorithm. If a spatiotemporal unit that the task intends to occupy has already been reserved by a high-priority task, this term has an extremely high cost, thus guiding the algorithm to automatically avoid it. This process generates an initial path for each task that minimizes infringing on high-priority reservations.

[0095] In traditional A* algorithms, the heuristic function h(n) is typically the geometric distance to the destination. This invention employs an A* algorithm with a fused time window, reconstructing h(n) as h(n) = w1∙A + w2∙B + w3∙C. Here, w1, w2, and w3 are adjustable weights, A is the geometric distance, B is the average predicted congestion index of the path nodes, and C is the dynamic urgency of the task. This makes the algorithm more inclined to choose faster and smoother paths when finding routes for high-urgency tasks, even if the geometric distance is slightly longer.

[0096] By prioritizing resources as a prerequisite for planning, "planning is avoidance" is achieved, reducing the pressure of resolving conflicts later and improving the success rate of the initial planning.

[0097] Step S3: Online Local Adjustment and Deadlock Prevention Based on Lyapunov: During the execution of the spatiotemporal corridor, a Lyapunov function is constructed based on the deviation between the actual state and the planned state of the shuttle to assess stability risk; when the risk exceeds a preset threshold, a fast local adjustment algorithm is triggered to adjust the spatiotemporal corridor parameters of the affected shuttle to restore system stability, while ensuring the high dynamic urgency weight task corridor.

[0098] Assume that the initial plan defines an ideal spatiotemporal trajectory r for each vehicle v. v(t) The system state is defined as the vector of deviations between the actual and planned positions of all shuttles. .

[0099] Design a candidate Lyapunov function based on the weighted sum of squares:

[0100] ;

[0101] Among them, e T Let P be the transpose of the matrix, e be the deviation vector between the actual and planned positions of all shuttles, and P be a diagonal positive definite matrix. The diagonal elements p in P are... i The dynamic urgency weight ω of the corresponding shuttle vehicle performing the task i Positive correlation; the stability risk threshold is determined based on the following formula:

[0102] ;

[0103] Where α and ξ are positive constants. Let V(e) be the time derivative. This indicates that the trajectory stability of high-urgency tasks contributes more to the overall stability of the system.

[0104] During online operation, estimates of V(e) and its time derivative are calculated in real time. Set a stability risk threshold ξ. When When α is a positive constant, the system is considered to have a high risk of instability and may evolve into deadlock or severe congestion.

[0105] Triggering a fast local adjustment algorithm: This algorithm fixes the high-urgency weights (i.e., ω) i The corridor of the larger shuttle remains unchanged, serving as an "anchor point." For the affected low-urgency shuttles, within a limited search space (such as the maximum allowed delay time), a fast linear programming problem is solved to reallocate their time slices through the conflict zone or switch to a pre-calculated alternative shorter path, making the new... The estimated value satisfies the stability condition. A lightweight online intervention mechanism can maintain stable shuttle operation with minimal disturbance, preventing cumulative performance degradation.

[0106] A further refinement involves addressing conflicts arising from bugs or unforeseen circumstances. Specifically, if shuttles R1 and R2 are detected occupying the same location at a given time, a conflict is identified. Traditional CBS generates two child nodes, each constraining R1 and R2 to prevent their presence at that specific time point. Therefore, step S3 further includes introducing a penalty function: Penalty = f(ΔScore, ConflictType), where ΔScore is the difference in priority scores between the tasks performed by the two shuttles, and ConflictType is categorized as head-on conflict, rear-end collision, etc. If the priority of task R1 is significantly higher than that of R2, the penalty function will result in a minimal increase in the cost of constraining the child nodes of R2, while the cost of constraining the child nodes of R1 will increase dramatically. This ensures that during constraint tree expansion, nodes constraining lower-priority shuttles are prioritized. Meanwhile, a pre-trained reinforcement learning AI agent (whose state is constrained tree node features, whose action is to choose which conflict to resolve or which branch to select, and whose reward is the cost of finally obtaining a conflict-free path) will quickly evaluate newly generated nodes, predict their probability of leading to a high-quality solution, and prune nodes with extremely low evaluations, significantly accelerating the search process.

[0107] By employing a priority-oriented penalty function, business logic is deeply embedded into the mathematical model of conflict resolution; and by using AI agent pruning, the problem of combinatorial explosion in the CBS algorithm is solved. The technical effect is to achieve fast, high-quality conflict resolution, with superior solution quality (overall task completion time).

[0108] Step S4: Digital twin-driven meta-learning optimization and model parameter self-update: Run the meta-learner in the digital twin environment. The meta-learner automatically adjusts the parameters of the nonlinear fusion function and the weight coefficients of the optimization target by using simulation and historical data to optimize long-term scheduling performance indicators.

[0109] A meta-learner runs within the digital twin model. At the heart of the meta-learner is a two-layer optimization framework.

[0110] Inner task: Given a set of model parameters θ (such as the fusion weights α, β, γ, δ, ε in S1, and other adjustment coefficients of the objective function in S2), simulate running a scheduling cycle in the digital twin (including multiple task scenarios that are randomly generated or extracted from history) to obtain a set of performance indicators J(θ), such as the actual weighted total completion time and the number of deadlocks.

[0111] Outer layer optimization: The goal of the meta-learner is to find the optimal parameters θ* such that the expected performance index E[J(θ)] is optimal over long-term, multi-scenario conditions. The outer layer optimization employs a gradient-based meta-learning algorithm, which calculates the gradient estimate of J(θ) with respect to θ by sampling different scenarios, and updates θ using gradient descent.

[0112] In the actual system, after each scheduling cycle (or every N cycles), the real data of that cycle (task set, decision, actual completion time) is input into the digital twin to initiate a one-dimensional learning iteration and update the parameter θ. The updated θ is then synchronized to the S1 and S2 models of the online system.

[0113] The key to achieving long-term autonomous evolution of the system lies in constructing a learning loop that enables the core model parameters of the system to automatically and continuously adapt to environmental changes and efficiency goals.

[0114] In a specific implementation scenario, this automatic adjustment process is triggered once at a fixed cycle (e.g., after processing 100 tasks, or at midnight every day). The specific process is as follows:

[0115] Assume the current parameter is θ current =(α=1.0,β=0.8,γ=1.5,δ=1.0,ε=2.0).

[0116] Data collection: Collect dataset D of all scheduled tasks over the past 24 hours.

[0117] Scenario construction: Randomly select 100 typical busy task time segments from D as training scenarios.

[0118] Inner layer assessment:

[0119] In digital twins, θ current The parameters are used to simulate these 100 scenarios.

[0120] The average weighted completion time T was calculated. avg =520 seconds, average number of deadlocks D avg =0.05 times.

[0121] If the deadlock penalty coefficient is 1000, then L(θ) current =520 + 1000 * 0.05 = 570.

[0122] Gradient estimation and update:

[0123] Meta-learners discovered through calculation that increasing the gradient direction of ε (congestion penalty weight) and slightly decreasing γ (task chain factor weight) can effectively reduce latency and deadlock caused by congestion.

[0124] Update according to the learning rate to obtain θ new =(α=1.0,β=0.8,γ=1.4,δ=1.0,ε=2.3).

[0125] With the new ε=2.3, there is a greater "aversion" to path congestion, resulting in lower urgency weights being assigned to tasks on congested paths in stage S1, thus favoring detours during stage S2 planning. This is expected to reduce the average latency and deadlock risk in the next runtime cycle.

[0126] A further proposed solution involves a policy neural network obtained through policy self-evolutionary learning, with the specific training process as follows:

[0127] A digital twin simulation environment synchronized with the physical warehouse is constructed. In this environment, a near-end policy optimization (PPO) reinforcement learning algorithm is adopted, with the shortest total task completion time and the fewest deadlock occurrences as the combined reward function. The policy neural network is trained offline. During online operation, the initial path generated in step S2 is continuously collected as the expert demonstration trajectory. The policy neural network is then fine-tuned online through a behavior cloning algorithm.

[0128] In digital twins, the agent's state is defined as: the position, speed, task urgency, and destination of all shuttles, as well as a tensor representing the predicted congestion index for each node on the map. Actions are defined as the direction for assigning the next target corridor to a waiting shuttle. The reward is designed as follows:

[0129] ;

[0130] D, E, and F are weighting coefficients, respectively, and T... S S represents the total task time. C Let N be the number of deadlocks. S This represents the number of high-urgency tasks completed. The PPO algorithm is used to iteratively train the policy network, enabling it to learn to make global plans that balance efficiency and fairness in complex scenarios.

[0131] Step S5: When the digital twin or actual system confirms a deadlock (e.g., multiple vehicles waiting in a loop at an intersection). First-level processing: Analyze all shuttles on the deadlock loop, select the shuttle with the lowest task priority, and have it back one node along its original path (one-step backtracking) to release the occupied resources. Then, immediately replan a new path for it starting from that backtracking point (replanning). If the deadlock cannot be broken due to environmental limitations after backtracking, then initiate the second level: perform a temporary "priority inversion," for example, allowing a low-priority shuttle in the loop to temporarily obtain high-priority access to the next required resource, breaking the loop waiting condition. After the task is completed, its priority is restored. If the above methods fail (e.g., the deadlock area is too large), then initiate the third level: define a minimum range containing the deadlocked shuttles and their directly affected areas, and only perform global path replanning for all shuttles within this range; the plans for shuttles outside the range remain unchanged.

[0132] The layered mechanism provides a progressive solution, from local fine-tuning to local readjustment and finally to minimal-scale global adjustments. The technical effect is to efficiently resolve deadlocks with minimal system disruption, ensuring high system availability.

[0133] Based on the same inventive concept, embodiments of the present invention also provide a warehouse shuttle path planning system based on multi-source coupling and collaborative optimization, used to implement the aforementioned method, the system comprising:

[0134] The intelligent scheduling host integrates a multi-source information fusion computing unit, a spatiotemporal collaborative optimization solver, and an online stability monitoring and adjustment module.

[0135] The digital twin and meta-learning server runs a high-fidelity simulation environment and a meta-learning optimizer; multiple vehicle control and communication units are deployed on each shuttle vehicle.

[0136] A high-speed industrial network connects the intelligent scheduling host, the digital twin and meta-learning server, and each of the vehicle control and communication units.

[0137] The system's modules work together to translate methodological innovations into a physical system, providing an efficient, reliable, and intelligent warehouse shuttle scheduling solution.

[0138] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A warehouse shuttle route planning method based on multi-source coupling and collaborative optimization, characterized in that, include: S1 receives a set of tasks, identifies the temporal dependencies between tasks to construct a task chain, and calculates the dynamic urgency weight of each task by fusing the inherent urgency of the task, the urgency propagation factor of the task chain, and the predicted congestion coefficient through a nonlinear fusion function. In S1, the task chain urgency propagation factor η i The calculation method is as follows: ； in, The latest allowed completion time for the chain containing the task. The current deadline for the task is λ, where λ is the preset decay coefficient. In S1, the nonlinear fusion function is specifically: ; Where, ω i The dynamic urgency weights for task i, Urgency i and Value i These are the inherent urgency and value coefficient of the task, respectively. Let α, β, γ, δ, ε be the predicted congestion coefficients for path i, where α, β, γ, δ, ε are positive parameters to be optimized. S2, with the optimization objective of minimizing the weighted completion time of all tasks, the weights are the dynamic urgency weights, and a mixed integer nonlinear programming model is constructed to solve the problem using shuttle motion constraints, static obstacle constraints, and shuttle spacetime mutual exclusion constraints as hard constraints, and a conflict-free spacetime corridor is assigned to each vehicle. S3, during the execution of the spatiotemporal corridor, a Lyapunov function is constructed based on the deviation between the actual state and the planned state of the shuttle to assess the stability risk; when the risk exceeds a preset threshold, a fast local adjustment algorithm is triggered to adjust the spatiotemporal corridor parameters of the affected shuttle to reduce the stability risk. The Lyapunov function V(e) is defined as follows: ； Among them, e T Let P be the transpose of the matrix, e be the deviation vector between the actual and planned positions of all shuttles, and P be a diagonal positive definite matrix. The diagonal elements p in P are... i The dynamic urgency weight ω of the corresponding shuttle vehicle performing the task i Positive correlation; the stability risk threshold is determined based on the following formula: (e) ； in and For positive integers, (e) is Derivative along time; S4. Run a meta-learner in the digital twin model. The meta-learner automatically adjusts the parameters of the nonlinear fusion function and the weight coefficients of the optimization target by using simulation data and historical data to optimize long-term scheduling performance indicators.

2. The method according to claim 1, characterized in that, S1 further includes: receiving a set of tasks, identifying the temporal dependencies between tasks to construct a task chain, and running a congestion propagation prediction model based on historical and real-time path data. The construction and training process of the congestion propagation prediction model includes: modeling the warehouse map as a spatiotemporal graph, with node features including historical shuttle density and speed, and edge features representing connection relationships and distances; using a gated spatiotemporal graph convolutional network as the model architecture; and outputting a predicted congestion coefficient. During training, the node states of historical continuous time slices are used as input, and the actual shuttle density of future time slices is used as labels. The network parameters are trained by minimizing the prediction error loss function.

3. The method according to claim 1, characterized in that, In S2, the spatiotemporal mutual exclusion constraint of the shuttle workshop is modeled using the Big M method: for any two different shuttles v1 and v2, at any time t and spatial node p, such that x v1,p,t +x v2,p,t ≤1 holds true, where x v1,p,t and x v2,p,t These are binary decision variables representing the positions occupied by shuttle cars v1 and v2, respectively.

4. The method according to claim 3, characterized in that, In S2, the branch and bound method is used to solve the mixed integer nonlinear programming model, and a heuristic cutting plane based on conflict constraints is added during the solution process to accelerate the solution.

5. The method according to claim 1, characterized in that, The fast local adjustment algorithm described in S3 is as follows: the spatiotemporal corridors of shuttles with dynamic urgency weights higher than the first threshold remain unchanged; for other affected shuttles, within a finite time delay adjustment window, a linear programming subproblem aimed at minimizing the derivative of the Lyapunov function is solved, and their time for passing through the conflict area is reallocated or alternative paths are selected.

6. The method according to claim 1, characterized in that, In step S4, the meta-learner adopts a two-layer optimization framework: the inner layer evaluates the scheduling performance J(θ) in the digital twin under given model parameters θ; the outer layer updates θ through gradient optimization to minimize the long-term expected performance loss E[J(θ)]; the model parameters θ include at least the weight parameters (α,β,γ,δ,ε) of the nonlinear fusion function in step S1 and the regularization coefficient of the optimization objective function in step S2.

7. A warehouse shuttle path planning system based on multi-source coupling and collaborative optimization, characterized in that, The system for implementing the method of any one of claims 1-6 comprises: The intelligent scheduling host integrates a multi-source information fusion computing unit, a spatiotemporal collaborative optimization solver, and an online stability monitoring and adjustment module. The digital twin and meta-learning server runs a high-fidelity simulation environment and a meta-learning optimizer. Multiple vehicle control and communication units are deployed on each shuttle vehicle; A high-speed industrial network connects the intelligent scheduling host, the digital twin and meta-learning server, and each of the vehicle control and communication units.

Images