Multi-objective optimization scheduling method and system for multi-AGV flexible job shop

Abstract

The invention relates to a multi-objective optimization scheduling method and system for a multi-AGV flexible job shop. The method comprises the steps of collecting job shop layout data, AGV operation parameters and production task data, and constructing a scheduling system basic model; performing recurrent neural network training and physical constraint model integration to obtain a hybrid scheduling prediction model; subtask decomposition and path search processing are carried out, and a dual-objective optimization model is obtained through constraint linearization conversion; performing particle swarm optimization solution on the dual-objective optimization model, and obtaining a non-dominated solution set through two-stage iterative calculation of an adaptive weight strategy and constraint conversion; performing multi-dimensional evaluation calculation on the non-dominated solution set to obtain an optimal scheduling execution scheme; and performing model feedback updating on the hybrid scheduling prediction model to obtain a target scheduling prediction model. According to the invention, multi-target balance of task completion time and energy consumption is realized, and the real-time response capability and operation stability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of AGV scheduling, and in particular to a multi-objective optimization scheduling method and system for a multi-AGV flexible workshop. Background Art

[0002] With the rapid development of intelligent manufacturing, multi-AGV scheduling optimization in flexible workshops has become a key technology to improve production efficiency. Traditional AGV scheduling methods mainly rely on static planning and manual experience, which is difficult to cope with dynamic and complex production environments, especially in multi-AGV collaborative operation scenarios. The coupling relationship between task allocation and path planning makes the scheduling optimization problem more complicated.

[0003] Existing AGV scheduling algorithms generally have the problem of insufficient adaptability to real-time dynamic environments. The scheduling method based on filters is overly dependent on prior information, and the filter parameters need to be redesigned for workshops of different sizes; the method based on heuristic algorithms is difficult to balance multiple optimization goals such as task completion time and energy consumption, and is prone to fall into local optimal solutions. Summary of the invention

[0004] The main purpose of the present invention is to provide a multi-objective optimization scheduling method and system for a multi-AGV flexible workshop, which achieves a multi-objective balance between task completion time and energy consumption, and improves the real-time response capability and operational stability of the system.

[0005] To achieve the above object, the present invention provides a multi-objective optimization scheduling method for a multi-AGV flexible workshop, comprising the following steps: Collect the layout data of the job shop, AGV operation parameters and production task data, and build the basic model of the scheduling system; Based on the AGV historical operation data in the basic model of the scheduling system, recursive neural network training and physical constraint model integration are performed to obtain a hybrid scheduling prediction model; Obtaining the AGV state feature vector and task requirement feature output by the hybrid scheduling prediction model, performing subtask decomposition and path search processing, and obtaining a dual-objective optimization model through constraint linearization transformation; The dual-objective optimization model is solved by particle swarm optimization, and a non-dominated solution set is obtained through two-stage iterative calculation of adaptive weight strategy and constraint conversion; Performing multi-dimensional evaluation calculation on the non-dominated solution set to obtain an optimal scheduling execution plan; The optimal scheduling execution plan is input into the AGV control system, and the hybrid scheduling prediction model is updated through model feedback to obtain a target scheduling prediction model.

[0006] The present invention also provides a multi-objective optimization scheduling system for a multi-AGV flexible workshop, comprising: The acquisition module is used to collect the layout data of the job shop, AGV operation parameters and production task data, and build the basic model of the scheduling system; A training module, used to perform recursive neural network training and physical constraint model integration based on the AGV historical operation data in the basic model of the scheduling system to obtain a hybrid scheduling prediction model; A processing module, used to obtain the AGV state feature vector and task requirement features output by the hybrid scheduling prediction model, and perform subtask decomposition and path search processing, and obtain a dual-objective optimization model through constraint linearization conversion; A solution module, used for performing particle swarm optimization on the dual-objective optimization model, and obtaining a non-dominated solution set through two-stage iterative calculation of an adaptive weight strategy and constraint conversion; A calculation module, used for performing multi-dimensional evaluation calculation on the non-dominated solution set to obtain an optimal scheduling execution plan; The updating module is used to input the optimal scheduling execution plan into the AGV control system, perform model feedback update on the hybrid scheduling prediction model, and obtain a target scheduling prediction model.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0009] In summary, the technical solution provided by the present invention constructs a hybrid scheduling prediction model through the deep integration of recursive neural networks and physical constraint models, improves the adaptability of the scheduling system to dynamic environments, and avoids the excessive reliance of traditional methods on prior parameters; adopts a task decomposition method based on similarity and a path planning strategy of the improved A* algorithm to reduce the coupling complexity of task allocation and path planning, and solves the deadlock problem in multi-AGV collaborative scheduling; introduces a two-stage optimization strategy of adaptive particle swarm algorithm and ε-constraint method to achieve a multi-objective balance between task completion time and energy consumption, and overcomes the defect that traditional algorithms are prone to fall into local optimality; designs a multidimensional evaluation system based on time dimension, efficiency dimension and energy consumption dimension, and combines the hierarchical analysis method for comprehensive evaluation to ensure the scientificity and reliability of the scheduling scheme; through online data collection and model dynamic update mechanism, the continuous optimization and adaptive learning of the scheduling system are realized, which significantly improves the real-time response capability and operation stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the steps of a multi-objective optimization scheduling method for a multi-AGV flexible workshop in one embodiment of the present invention; Figure 2 It is a structural block diagram of a multi-objective optimization scheduling system for a multi-AGV flexible workshop in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0011] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0013] Reference Figure 1 This embodiment provides a multi-objective optimization scheduling method for a multi-AGV flexible workshop, including the following steps: S1, collects the layout data of the job shop, AGV operation parameters and production task data, and builds the basic model of the scheduling system; Among them, the workshop layout data is collected, including the coordinate data of the processing equipment and the location data of the AGV charging station. The topological structure modeling of the AGV driving channel is carried out, and the layout topological data of the workshop is obtained by topological analysis of the layout of the channels in the workshop. The motion trajectory of the AGV is mapped according to the workshop layout topological data. By converting the coordinate information of each important position, equipment and channel in the workshop into a unified workshop layout matrix, a matrix model representing the spatial distribution of the workshop is obtained. The operation parameters of the AGV are collected. Including the collection of data such as the driving speed, turning radius, load capacity and battery capacity of the AGV equipment. These parameters are monitored in real time by on-board sensors, and the dynamic parameter model of the AGV is established by combining data fitting techniques such as the least squares method. This model can accurately describe the motion performance and constraints of the AGV under different working conditions. The production tasks are scheduled according to the process flow data in the workshop. The workpiece processing process data in the workshop, the transit time data between processes and the priority data of the task are input into the task scheduler. The task scheduler calculates the optimal production task sequence through a linear programming model. The task sequence will determine the execution order and priority of different tasks to ensure the smoothness and efficiency of the production process. In this process, the linear programming model not only considers the time constraints between tasks, but also combines the priority information of tasks to maximize the overall benefits of workshop production. Based on the workshop layout matrix, the AGV dynamic parameter model and the production task sequence, the kinematic model of AGV is established. In the kinematic model, the velocity vector and position vector of AGV are substituted into the Euler equation as core variables, and the motion constraints of AGV during operation are obtained by solving these equations. Motion constraints refer to the physical restrictions that AGV must meet when performing tasks, such as maximum speed, minimum turning radius, etc. These constraints ensure the safety and efficiency of AGV when performing tasks. Constraint modeling is performed on AGV motion constraints and task time windows to obtain a scheduling constraint model. The time window refers to the time range within which each task needs to be completed. During the scheduling process, it must be ensured that AGV can complete the task within the specified time. At the same time, considering the energy consumption of AGV, the battery capacity of AGV will become a limiting factor in the case of long-term operation. The energy consumption of the speed data and battery capacity data in the AGV dynamic parameter model is integrated to obtain the corresponding energy consumption function, which quantifies the energy demand of AGV when performing different tasks. The motion constraint model, task time window constraint model and energy consumption function of AGV are integrated to obtain the basic model of the scheduling system.

[0014] S2, based on the AGV historical operation data in the basic model of the scheduling system, recursive neural network training and physical constraint model integration are performed to obtain a hybrid scheduling prediction model; Specifically, the AGV running trajectory data, task execution time data and energy consumption data in the basic model of the scheduling system are divided into data sets, and the data are converted into training sample sets suitable for time series analysis through time series segmentation. The training samples are standardized. In order to improve the quality of the data and improve the efficiency of neural network training, the training sample set is input into the batch normalization layer for data standardization to ensure that the mean and variance of the input features are relatively balanced, which helps to accelerate training and avoid the problem of gradient disappearance or explosion. In order to enhance the diversity of training data, data enhancement technology is used to randomly rotate and translate the trajectory data to generate an enhanced training data matrix. In this way, the scale of the training set is effectively expanded, and the adaptability and generalization ability of the model to different scenarios are enhanced. Time series feature extraction is performed through the long short-term memory network unit (LSTM). LSTM can effectively capture the long-term and short-term dependencies in time series data and is suitable for processing the historical trajectory data of AGV. By batch processing the enhanced training data matrix, LSTM can extract the state feature vectors of AGV from it, which will be used as input to build a recursive neural network model. AGV state features include dynamic change features such as position, speed, and acceleration, which can reflect the operating status of AGV in different time periods. The AGV state feature vector is input into a recursive neural network with three hidden layers. The recursive neural network performs deep learning and processing on the input data through multiple layers of neurons to extract more complex rules. In order to optimize the training effect of the network, the parameters of the neural network are optimized using the Adam optimizer. The Adam optimizer is an adaptive optimization algorithm that can effectively adjust the learning rate and accelerate convergence during the training process to avoid the defects of the traditional gradient descent method. After training, the first neural network model can make a preliminary prediction of the state of the AGV. The hidden layer of the first neural network model is subjected to dropout processing. Dropout is a regularization technique that randomly discards some neurons during the training process to avoid overfitting of the model to the training data. When the verification loss does not decrease for multiple consecutive times (for example, N times), the early stopping mechanism is triggered to prevent the model from being over-adjusted during training, resulting in performance degradation. After optimization, the second neural network model obtained has better generalization ability and higher stability. Based on the AGV kinematic equations, time window constraints, and energy consumption functions in the basic model of the scheduling system, these physical constraints are constructed as Lagrangian constraints. Lagrangian constraints effectively introduce physical constraints into the optimization problem, thereby ensuring that scheduling decisions not only meet actual physical limitations, but also balance the needs of multiple objectives. For example, the AGV's kinematic equations ensure that it does not exceed physical limitations during driving, the time window constraints ensure that the task is completed on time, and the energy consumption function limits the AGV's energy consumption, thereby ensuring that the AGV can complete the task efficiently.The physical constraint loss function and the prediction loss of the neural network are weighted and combined to obtain the hybrid optimization objective function. Under the guidance of this objective function, the back propagation algorithm is used for iterative calculation. By gradually adjusting the parameters of the neural network, the hybrid scheduling prediction model can maximize the prediction accuracy on the basis of satisfying the physical constraints. After multiple iterations, the hybrid scheduling prediction model can accurately predict the operating status and scheduling plan of the AGV based on the state characteristics and task requirements of the AGV.

[0015] S3, obtain the AGV state feature vector and task requirement characteristics output by the hybrid scheduling prediction model, perform subtask decomposition and path search processing, and obtain the dual-objective optimization model through constraint linearization transformation; It should be noted that the output of the hybrid scheduling prediction model is processed, and the similarity between the state feature vector of the AGV and the task demand feature is calculated. The similarity between tasks is measured by a weighted combination of cosine distance and Euclidean distance to obtain a task similarity matrix, which reflects the similarity between different tasks. According to the set similarity threshold, the tasks are clustered to obtain the decomposition sequence of the tasks. Each subtask in the task decomposition sequence is mapped and matched with the AGV device. In order to ensure the reasonable pairing of tasks and AGVs, a 0-1 integer programming method is used to construct an M×N-dimensional task allocation matrix, where M is the total number of tasks and N is the total number of AGVs. The matrix clearly represents the allocation relationship between each task and each AGV, and the task allocation scheme is preliminarily obtained. Based on the preliminary allocation scheme, the execution task sequence of each AGV is encoded for optimization. Through the optimization method based on genetic algorithm, a new task optimization sequence is generated using single-point crossover and double-point mutation operations. The single-point crossover operation can explore more task arrangement methods by exchanging some tasks in the task sequence, while the double-point mutation operation increases the diversity of the solution space by randomly changing the order of tasks. These operations help to improve the diversity of task scheduling, thereby improving the optimization effect of scheduling schemes. Perform path planning. According to the workshop layout matrix, the path planning space of the A* algorithm is constructed. The A* algorithm is a graph search algorithm that can find the shortest path in a given grid environment. When constructing the path planning space, the layout space of the workshop is gridded, and the distance between each node is set to the length of an AGV body to ensure the accuracy and practical feasibility of the path calculation. Through the gridding process of the grid method, a path search network is obtained. A heuristic function is set for the path search network to optimize the path search process. The design of the heuristic function is crucial to the performance of the A* algorithm. The evaluation function of the path is constructed by combining the weighted sum of the Manhattan distance and the dynamic obstacle avoidance factor. The Manhattan distance can provide an estimate of the straight-line distance, while the dynamic obstacle avoidance factor considers the impact of real-time obstacles. In this way, the risk of collision of the AGV during the path planning process is effectively avoided. Using the path evaluation function, the cost matrix of the path is obtained, which reflects the cost of different paths and helps optimize the path selection. At the same time, the motion constraints of the AGV are linearized. The kinematic equations, turning constraints, and speed constraints of AGV are usually nonlinear. The first-order Taylor expansion is used to convert nonlinear constraints into a set of linear inequalities. This allows the originally complex motion constraints to be expressed in the form of linear constraints, which is convenient for subsequent optimization calculations. The task completion time function is calculated based on the task optimization sequence. This function can reflect the time requirement for task completion and provide a timeliness basis for scheduling optimization. At the same time, the energy consumption function of each AGV is calculated based on the AGV's dynamic parameter model, which reflects the energy consumption of the AGV during the execution of the task.In order to ensure that the objective functions can be reasonably compared at different scales, the task completion time function and energy consumption function are normalized and converted into standardized objective functions. The standardized objective function, linearized constraint equation and path cost matrix are combined to obtain a dual-objective optimization model.

[0016] S4, solve the dual-objective optimization model by particle swarm optimization, and obtain the non-dominated solution set through two-stage iterative calculation of adaptive weight strategy and constraint transformation; Specifically, the first stage of the dual-objective optimization model is initialized. According to the total number of tasks M and the total number of AGVs N, an M×N-dimensional particle swarm initial solution is generated. The initial solution space represents a preliminary match between tasks and AGVs, covering all possible task allocation methods. In the framework of particle swarm optimization, each particle represents a task allocation scheme, and the solution space is the search range of the particle swarm. The initial solution space is input into the adaptive particle swarm algorithm for optimization. The adaptive particle swarm algorithm controls the exploratory and exploitative nature of the search process by adjusting the inertia weight of the particle swarm. In this process, the inertia weight value is calculated using a linear decreasing method. The adjustment of the inertia weight value helps to balance the local search and global search of particles in the search space, so that particles can maintain a large search range in the early stage, and can finely adjust the accuracy of the solution in the later stage, thereby accelerating the convergence process. The position and velocity of the particle swarm are updated by the adaptive particle swarm algorithm, effectively guiding the particles to move in the direction of the optimal solution. In this process, the movement of particles is not only affected by their own experience, but also optimized by the guidance of the global optimal solution. In order to improve the efficiency of the particle swarm algorithm, the tournament selection strategy is adopted to screen the global optimal solution. This strategy selects the best particle at present through a competition mechanism and gives it a higher selection probability to improve the search quality of the particle swarm. After the first stage of search calculation, a set of first stage search solutions is obtained, which represent the possible solutions found in the preliminary optimization process. The ε-constraint method is applied to the first stage search solution set, and the ε-constraint method is used to transform the objective function, transform the energy consumption target into an inequality constraint condition, and obtain the constraint transformation model. The original multi-objective optimization problem is transformed into a single-objective optimization problem, the energy consumption target is regarded as a constraint condition, and the task completion time is optimized as the main objective. By taking energy consumption as a constraint, it is ensured that while optimizing the task completion time, a certain energy consumption limit is not exceeded, so as to achieve the rational use of resources. Based on the constraint transformation model, the population is evolved by using the binary crossover operator and the Gaussian mutation operator to obtain the initial population of the second stage. The binary crossover operator can exchange some genes of the particles to generate new solutions, while the Gaussian mutation operator increases the diversity of the search space by slightly perturbing the position of the particles, thereby avoiding the risk of falling into the local optimal solution. The initial population of the second stage is locally searched and optimized. The particle position is fine-tuned through the variable neighborhood search strategy, which ensures that a better solution can be found in the local solution space by dynamically adjusting the search range and neighborhood structure. The introduction of variable neighborhood search can effectively improve the local search capability of the particle swarm algorithm, improve the accuracy of the solution and the optimization effect. After local search optimization, a set of local optimal solution sets are obtained, which represent the optimal solutions in the current stage of the search process. In order to select the most representative solution set from the local optimal solution set, non-dominated sorting is performed.Non-dominated sorting is a sorting method based on the Pareto optimization idea. It determines which solutions are optimal by comparing the dominance relationship between solutions. In non-dominated sorting, if a solution is not inferior to another solution in all objectives and is better than another solution in at least one objective, then the solution is considered to dominate the other solution. Through non-dominated sorting, the local optimal solution set is divided into multiple levels to form a priority sorting. The solution set is screened based on the elite retention strategy. The elite retention strategy ensures that the best solution can continue to exist during the evolution process by retaining the current optimal solution and is further optimized in subsequent iterations. The solution is selected based on the crowding distance. The crowding distance is used to measure the density of the solution in the solution space. The smaller the crowding, the solution is relatively sparse in the target space, has more potential, and should be retained first. After non-dominated sorting, crowding distance calculation and elite retention strategy screening, a candidate non-dominated solution set is obtained. By judging the Pareto dominance relationship of these candidate solutions, the comprehensive optimal solution of task completion time and energy consumption is selected. These final non-dominated solution sets represent the solutions that can achieve the best balance between multiple objectives in task scheduling and path planning, namely, non-dominated solution sets.

[0017] S5, perform multi-dimensional evaluation calculation on the non-dominated solution set to obtain the optimal scheduling execution plan; Among them, for each solution in the non-dominated solution set, time-related indicators are extracted. Time indicators include the total time for task completion and the waiting time of AGV. By calculating the total time for task completion, the scheduling efficiency of the entire production task is evaluated, while the waiting time of AGV reflects the idleness of the vehicle or the irrationality of task arrangement. All time indicators are standardized so that fair comparisons can be made between different solutions to obtain standardized time indicators. Efficiency indicators are calculated for the non-dominated solution set. Efficiency indicators include the utilization rate of AGV equipment and the completion rate of production tasks. The equipment utilization rate of AGV is obtained by measuring the ratio of the active time of each AGV in the production process to the total working time, reflecting the efficiency of AGV resource utilization in the workshop. The production task completion rate measures the completion of workshop tasks by comparing the ratio between the number of tasks actually completed and the number of planned tasks. These two efficiency indicators are input into the linear programming model and standardized to obtain the corresponding standardized efficiency indicators. Cumulative calculations are performed for the energy consumption data of AGV equipment operation in each non-dominated solution set. By calculating the total energy consumption of each AGV and dividing it by the number of tasks completed by the AGV, the standardized energy consumption index of each AGV is obtained, which reflects the energy efficiency of the AGV under specific task allocation and scheduling. The standardized time index, standardized efficiency index, and standardized energy consumption index are input into the evaluation system of the analytic hierarchy process. The analytic hierarchy process is a comprehensive evaluation method for multi-criteria decision-making, which can rank multiple alternatives according to the weights of each index. In order to establish a hierarchical analysis network, a three-layer evaluation structure is constructed. The three-layer structure includes the target layer, the criterion layer, and the solution layer. The target layer is the final evaluation target of the scheduling execution plan, the criterion layer is the evaluation indexes such as time, efficiency, and energy consumption, and the solution layer is each solution in the non-dominated solution set. By substituting the standardized indexes into this structure, a comprehensive evaluation is performed. The consistency calculation of the judgment matrix in the three-layer evaluation structure is performed. The construction of the judgment matrix is ​​a key step in the analytic hierarchy process, which is used to express the relative importance of each evaluation index. Through the consistency calculation, it is ensured that the weight distribution in the judgment matrix is ​​reasonable and the consistency between different evaluation indicators is high. The consistency calculation is completed by calculating the consistency ratio (CR). If the consistency ratio is lower than the set threshold, it means that the judgment matrix has good logical consistency and the weight distribution is reasonable. Based on the weight matrix obtained by consistency calculation, the standardized time index, efficiency index and energy consumption index are comprehensively calculated to obtain the evaluation score of each non-dominated solution set. This score reflects the comprehensive advantages and disadvantages of each solution under multi-dimensional goals such as time, efficiency and energy consumption. Based on the comprehensive evaluation score sequence, multi-AGV simulation is performed to verify the execution effect of the selected solution in the actual scheduling process. The simulation can simulate the scheduling process of multiple AGVs in the actual operation workshop and evaluate the execution effect of different solutions.Through simulation, the feasibility and effectiveness of the optimization plan are verified to ensure that the selected optimal scheduling execution plan can provide the expected performance in the actual environment.

[0018] S6, input the optimal scheduling execution plan into the AGV control system, perform model feedback update on the hybrid scheduling prediction model, and obtain the target scheduling prediction model.

[0019] Specifically, the optimal scheduling execution plan is decomposed into a time series. The task sequence is converted into a speed control instruction that the AGV can understand and execute. By inputting the task sequence into the kinematic equation of the AGV, the corresponding speed instruction is calculated according to the time node of each task and the interval between tasks, and the real-time control instruction is provided for the AGV. The real-time control sequence of the AGV is state monitored and data filtered. By collecting the position information and task completion status data generated during the real-time operation of the AGV, the operation state matrix of the AGV is obtained. In order to ensure the accuracy and reliability of the data, a third-order Kalman filter is used to filter these data. By setting the noise coefficient to 0.1, the Kalman filter can effectively reduce the influence of external interference and sensor error, and calculate the real position and state change of the AGV. The AGV operation state matrix is ​​input into the recursive neural network model for analysis and prediction. Using the long short-term memory (LSTM) unit, the model can capture the time dependency of the AGV movement process and predict the future movement trajectory of the AGV. By performing collision risk detection on the state prediction sequence, the model can identify potential collision risks. Based on the workshop layout matrix, the minimum safe distance between AGVs is calculated, and the safety threshold is set to 2 times the body length to determine whether a collision occurs. If the predicted trajectory intersects with the motion trajectory of other AGVs and does not meet the safety distance requirements, the system will trigger the obstacle avoidance mechanism and adjust the path or speed to ensure that the safe distance between AGVs is maintained. In order to improve the driving safety of AGVs, the local path of AGVs is replanned according to the obstacle avoidance constraints. In this process, the Bellman equation is used to calculate the path cost, and the path points are updated through continuous iteration to find the optimal path sequence. The Bellman equation helps AGV calculate the path with the lowest cost based on the current state and environmental information, so that AGV can complete the task with the optimal path while avoiding collisions. After the path planning is completed, the AGV's operating state matrix and the optimized path sequence are combined in time sequence to form a new real-time execution data set. These data include important information such as the real-time position of the AGV, task completion status, and energy consumption. Through statistical calculations of these real-time data, the task completion rate and energy consumption rate of the AGV are obtained. The hybrid scheduling prediction model is updated online. The update process is based on real-time data. Through batch processing, every 32 samples are input into the hybrid scheduling prediction model as a batch. Through this operation, the system can optimize and update the model parameters according to the latest operating data, so that the scheduling prediction model can always adapt to the dynamically changing production environment of the workshop. The updated model parameters are weighted and integrated with the physical constraint loss function to ensure that the model can not only accurately predict the motion trajectory and task completion of the AGV, but also fully consider the physical constraints when optimizing the scheduling, such as energy consumption, task time window, etc., to obtain the target scheduling prediction model.

[0020] In one example, the job shop layout data, AGV operation parameters and production task data are collected, and the basic model of the scheduling system is built, including: Collect the coordinate data of processing equipment and the location data of AGV charging stations in the workshop, and perform topological structure modeling on the AGV driving channel to obtain the workshop layout topology data; According to the workshop layout topology data, coordinate mapping of the AGV motion trajectory is performed to obtain the workshop layout matrix; The driving speed, turning radius, load capacity and battery capacity of the AGV equipment are collected by sensors and fitted by the least square method to obtain the AGV dynamic parameter model. The workpiece processing process data, process transfer time data and task priority data are input into the task scheduler for linear programming modeling to obtain the production task sequence. The AGV kinematic model is established based on the workshop layout matrix, AGV dynamic parameter model and production task sequence, and the AGV velocity vector and position vector are substituted into the Euler equation to obtain the AGV motion constraint conditions; Constraint modeling is performed on the AGV motion constraints and task time windows to obtain the scheduling constraint model, and energy consumption integration calculations are performed on the speed data and battery capacity data in the AGV dynamic parameter model to obtain the energy consumption function. The energy consumption function is integrated with the scheduling constraint model to obtain the basic model of the scheduling system.

[0021] In this example, the coordinate data of the processing equipment in the workshop and the location data of the AGV charging station are collected, and the topological structure of the AGV driving channel is modeled. By collecting the physical layout information of the workshop, a complete workshop layout topology data model is constructed to help the system understand the spatial position relationship between various equipment and charging stations in the workshop. The positions of all processing equipment, charging stations, and AGV driving channels in the workshop are measured to form the coordinate information of each location point. For example, assuming the location coordinates of the processing equipment are ,in The location of the charging station is marked as , and the AGV driving channel forms different topological structures according to the actual layout, and each channel forms a graphical structure through physical or logical connections. According to the workshop layout topology data, the motion trajectory of the AGV is mapped to obtain the workshop layout matrix. The workshop layout matrix is ​​a discretized spatial representation. It divides the area where the AGV may travel into grids and maps each grid to a specific coordinate point to represent the layout of the entire workshop. Assume that the workshop layout matrix is ,in Indicates Row, No. The state of the column grid (such as whether it is an obstacle, a pass area, etc.). Through matrix modeling, the motion trajectory of the AGV is combined with the actual layout of the workshop to provide support for path planning. In the process of AGV dynamic modeling, the sensor collects the driving speed, turning radius, load capacity and battery capacity of the AGV equipment. At the same time, in order to fit these parameters, the least squares method is used for data analysis, and a set of optimal parameters is solved by minimizing the fitting error. Let the speed of the AGV be , the turning radius is , load capacity is , the battery capacity is , then these parameters are fitted by the following equations: ; in, , , , is a function of the fitted speed, turning radius, load capacity and battery capacity, is the corresponding sampling point, is the number of sampling points. In this way, an accurate AGV dynamic parameter model is obtained. The workpiece processing process data, process transfer time data and task priority data are input into the task scheduler for linear programming modeling to determine the execution order of production tasks. The linear programming model is modeled based on multiple factors such as the process transfer time of the task, the task priority, and the operation restrictions of the AGV. Assume that the process transfer time of the task is , the task priority is , then the task scheduling is optimized through the following linear programming objective function: ; in, It is The priority of each task, It is The objective function is to minimize the total time to complete the task while giving priority to high-priority tasks. The kinematic model of AGV is established based on the workshop layout matrix, the AGV dynamic parameter model and the production task sequence. The AGV kinematic model predicts the future motion trajectory by calculating the AGV's speed and position vector. Let the AGV's position vector be , the velocity vector is , then the kinematic model of AGV is described by Euler equation: ; in, Indicates the rate of change of the AGV position, Represents the velocity vector of the AGV. By substituting the velocity vector and position vector of the AGV into the above equation, the motion constraints of the AGV are obtained. These constraints reflect the spatial and temporal constraints of the AGV when performing tasks, such as speed limits, acceleration limits, maximum turning radius, etc. When constructing the scheduling constraint model, the motion constraints of the AGV and the time window constraints of the task are considered. The time window constraints of the task are expressed by the following inequality: ; in, It's a task The start time of It's a task The completion time, It's a task By modeling these constraints, the scheduling system can not only meet the physical constraints of the AGV, but also meet the timeliness requirements of the task. The energy consumption function is obtained by integrating the speed data and battery capacity data in the AGV's dynamic parameter model. The energy consumption function is related to factors such as the AGV's speed, acceleration, travel distance, and load. Suppose the energy consumption of the AGV is , then the energy consumption function is expressed as: ; in, , , γ is the coefficient related to the specific model of AGV, It's speed. is the acceleration, is the load, is the total time of task execution. By integrating the energy consumption, the total energy consumption of AGV in the process of task execution is obtained. The energy consumption function is integrated with the scheduling constraint model to obtain the basic model of the scheduling system.

[0022] In one example, recursive neural network training and physical constraint model integration are performed based on the AGV historical operation data in the basic model of the scheduling system to obtain a hybrid scheduling prediction model, including: The AGV running trajectory data, task execution time data and energy consumption data in the basic model of the scheduling system are divided into data sets, and time series segmentation is performed to obtain a training sample set; The training sample set is input into the batch normalization layer for data standardization, and the trajectory data is enhanced by random rotation and translation to obtain an enhanced training data matrix; The enhanced training data matrix is ​​divided into batches, and the time series features are extracted through the long short-term memory network unit to obtain the AGV state feature vector; The AGV state feature vector is input into a recursive neural network with three hidden layers, and the parameters are optimized by the Adam optimizer to obtain the first neural network model; Perform dropout processing on the hidden layer of the first neural network model, and trigger the early stopping mechanism when the verification loss does not decrease for N consecutive times, to obtain the second neural network model; Based on the basic model of the scheduling system, the AGV kinematic equation, time window constraint and energy consumption function are constructed as Lagrangian constraint terms to obtain the physical constraint loss function. The physical constraint loss function and the neural network prediction loss are weightedly combined to obtain a hybrid optimization objective function, which is then input into the second neural network model, and the hybrid scheduling prediction model is obtained through iterative calculation using the back propagation algorithm.

[0023] In this example, the AGV operation trajectory data, task execution time data, and energy consumption data in the basic model of the scheduling system are divided into data sets, and time series segmentation is performed to obtain a training sample set. The AGV operation trajectory data contains the coordinate information of each position passed by each AGV during the execution of the task, the task execution time data contains the time required for each task from start to completion, and the energy consumption data includes the real-time energy consumption information of the AGV when performing the task. Through time series segmentation, these continuous data are cut according to a fixed time step to obtain a set of training samples. Assume that the trajectory data of the AGV when performing the task is ,in For AGV at all times The location coordinates of the task execution time data are , the energy consumption data is By dividing these data into multiple time windows, multiple training sample sets are obtained. The training sample sets are input into the batch normalization layer for data standardization, so that the data ranges of different features tend to be consistent, avoiding some features from having too much impact on model training due to the large amount of data. The data is processed with zero mean and unit variance. Suppose the training sample set is ,in For the samples, after standardization, the standardized data of the samples Obtained by the following formula: ; in, and are the mean and standard deviation of the training set respectively. In order to enhance the diversity and generalization ability of the data, the trajectory data is randomly rotated and translated for data enhancement. Assume that the trajectory data set is , through rotation transformation, the trajectory data is rotated along a certain angle to obtain enhanced data , and its rotation angle is randomly generated. During the translation process, the trajectory data is randomly translated to obtain a new trajectory data set Through these data enhancement techniques, multiple variants of training samples are generated to increase the robustness of the model. The enhanced training data matrix is ​​input into the long short-term memory network (LSTM) unit for time series feature extraction. LSTM is a recurrent neural network variant that can effectively capture long-term dependencies in time series data. In the LSTM network, the input data at each time step It will be processed by a memory unit and output the hidden state of the next time step The core of LSTM includes input gate, forget gate and output gate, which control the flow of information and effectively avoid the gradient vanishing and exploding problems in traditional RNN. Assume that the input data is , after LSTM processing, the state feature vector is obtained , these feature vectors can effectively represent the state information of the AGV. The AGV state feature vector is input into a recursive neural network (RNN) with 3 hidden layers. RNN is a type of neural network used to process sequence data. By sharing weights, it can propagate information between multiple time steps and capture long-term dependencies in the sequence. In this model, the output of each hidden layer is used as the input of the next layer, and finally the prediction result is obtained through an output layer. During the training process, the Adam optimizer is used to optimize the parameters of the network. The Adam optimizer dynamically adjusts the learning rate by calculating the mean and variance of the gradient, thereby accelerating the convergence process. The update formula of the Adam optimization algorithm is: ; ; ; ; in, and are the first-order and second-order moment estimates, respectively, and is the attenuation rate, is a small constant to prevent division by zero errors, is the learning rate. Through this process, the parameters of the neural network can be effectively optimized to obtain the first neural network model. After obtaining the first neural network model, in order to improve the generalization ability of the model, the hidden layer of the network is subjected to dropout processing. Dropout is a technology to prevent overfitting. By randomly discarding the output of some neurons during the training process, the network uses different sub-networks at each update to improve the robustness of the model. During the training process, an early stopping mechanism is set. When the loss function on the validation set is continuously When there is no significant decrease in the iteration, stop training in advance to prevent overfitting. On this basis, based on the basic model of the scheduling system, the AGV kinematic equation, time window constraint and energy consumption function are constructed as Lagrangian constraint terms to obtain the physical constraint loss function. The Lagrangian method is an optimization method that incorporates constraints into the objective function and handles constraint problems by introducing Lagrangian multipliers. Assume that the objective function is , the constraints are , then the Lagrangian function is expressed as: ; in, is the Lagrange multiplier, is the constraint condition. The optimal solution is obtained by minimizing the Lagrangian function. The physical constraint loss function and the neural network prediction loss are weighted to obtain the hybrid optimization objective function. The hybrid optimization objective function combines the errors of physical constraints and neural network predictions, and can optimize the scheduling results based on system constraints. The expression of the hybrid optimization objective function is: ; in, is the prediction loss for the neural network, is the physical constraint loss, and is the weight coefficient, which controls the relative importance of the two. The hybrid optimized self-standard function is input into the second warp network model, and the hybrid scheduling prediction model is obtained by iterative calculation through the anti-co-propagation algorithm.

[0024] In one example, the AGV state feature vector and task requirement feature output by the hybrid scheduling prediction model are obtained, and subtask decomposition and path search processing are performed, and a dual-objective optimization model is obtained through constraint linearization conversion, including: The similarity between the AGV state feature vector and the task requirement feature output by the hybrid scheduling prediction model is calculated, and the task similarity matrix is ​​obtained through the weighted combination of cosine distance and Euclidean distance. Tasks are clustered according to the similarity threshold to obtain the task decomposition sequence. Map and match each subtask in the task decomposition sequence with the AGV equipment, and construct an M×N dimensional task allocation matrix through 0-1 integer programming, where M is the total number of tasks and N is the total number of AGVs, to obtain the initial allocation plan; Encode the execution task sequence of each AGV in the initial allocation plan, and obtain the task optimization sequence through single-point crossover and double-point mutation operations; The A* algorithm path planning space is constructed according to the workshop layout matrix, the distance between nodes is set to the length of one AGV body, and the space is gridded by the grid method to obtain the path search network; A heuristic function is set for the path search network, and the weighted sum of the Manhattan distance and the dynamic obstacle avoidance factor is used as the path evaluation function to obtain the path cost matrix; Perform first-order Taylor expansion on the AGV motion constraints, transform the AGV kinematic equations, turning constraints and speed constraints into a set of linear inequalities, and obtain the linearized constraint equations. The task completion time function is calculated based on the task optimization sequence, the energy consumption function is calculated for the AGV dynamic parameter model, and the standardized objective function is obtained through normalization. The standardized objective function, linearized constraint equation and path cost matrix are combined to obtain a dual-objective optimization model.

[0025] In this example, the similarity between the AGV state feature vector output by the hybrid scheduling prediction model and the task requirement feature is calculated, and the task similarity matrix is ​​obtained by weighted combination of cosine distance and Euclidean distance. The calculation of task similarity is based on the feature similarity between tasks, which includes the execution time of tasks, the resource priority required for tasks, etc. The similarity between different tasks is measured by weighted combination of cosine distance and Euclidean distance. Let the task feature vector be and ,in is the dimension of the task feature, and the cosine distance and Euclidean distance are calculated by the following formulas: Cosine distance: ; in, Representation Task and tasks The inner product of eigenvectors, and The tasks are and tasks The modulus of the eigenvector. Euclidean distance: ; in, and The tasks are and tasks In the The final task similarity is obtained by weighted combination of cosine distance and Euclidean distance: ; in, and is the weighting coefficient, . According to the set similarity threshold, cluster the tasks, classify the tasks with high similarity into the same category, and obtain the task decomposition sequence. Map and match the subtasks after each task decomposition with the AGV equipment, and construct a task allocation matrix through 0-1 integer programming. Assume that there is tasks and AGV, the matching between task and AGV is represented as a The matrix ,in Indicates the task Assigned to AGV ,otherwise This task allocation matrix is ​​optimized through a 0-1 integer programming model. The goal is to minimize the total task completion time or maximize the efficiency of task execution. The specific objective function is expressed as: ; in, It's a task In AGV The execution cost (e.g., task execution time) on each AGV is calculated, and the initial task allocation scheme is obtained by solving the optimization problem. The initial task allocation scheme is optimized. The task execution sequence of each AGV is encoded, and the single-point crossover and double-point mutation operations in the genetic algorithm are used to search for the optimized task sequence. Assume that the task execution sequence of each AGV is represented as a permutation vector ,in is the number of tasks that the AGV needs to perform. Through the single-point crossover operation, the task sequences of the two parent individuals are randomly exchanged to generate offspring individuals: ; ; The double-point mutation operation generates new individuals by randomly exchanging two task positions in the sequence. In the process of continuous iterative optimization, the genetic algorithm can find the optimal combination of task sequences through operations such as selection, crossover, and mutation. After the task allocation is optimized, path planning is performed. According to the workshop layout matrix, the path planning space of the A* algorithm is constructed. The A* algorithm is a heuristic search algorithm that estimates the cost from the current node to the target node and selects the path with the minimum cost for search. Assuming that the workshop layout is a two-dimensional grid, each grid cell represents a location in the workshop, and the path cost is composed of two factors: actual cost and heuristic estimate. The actual cost is the path cost from the starting point to the current node, and the heuristic cost is the estimated cost from the current node to the target node.

[0026] ; in, From the starting point to the current node The actual cost of From the current node The heuristic estimated cost to the target node. The result of path planning is to find the shortest path from the starting point to the target node. In the path planning process, the workshop layout matrix is ​​rasterized by the grid method, and the workshop space is divided into several small grids. Each grid represents a node, and the connection between the nodes represents the feasible path of the AGV. In order to increase the feasibility of the path, a dynamic obstacle avoidance factor is set to incorporate the influence of obstacles into the path evaluation function. In the path evaluation process, the weighted sum of the Manhattan distance and the obstacle avoidance factor is used as the path cost: ; in, is the Manhattan distance, obstacle is the effect of obstacles on the path, and is the weight coefficient. In order to meet the motion constraints of AGV, the AGV kinematic equation, turning constraint and speed constraint are linearized. Assume that the kinematic equation of AGV is: ; in, is the speed of the AGV, is the distance traveled by the AGV, is the travel time. Through Taylor expansion, these nonlinear constraints are transformed into linear inequality forms. For example, the turning radius constraint of AGV is obtained through Taylor expansion: ; in, is the turning angle, is the turning angular velocity, is the turning radius. Based on the task optimization sequence, the task completion time function is calculated, and the energy consumption function is calculated in combination with the AGV dynamics model. The energy consumption function is related to factors such as speed, driving distance, and load. Assuming that energy consumption is proportional to speed, it can be expressed as: ; in, For energy consumption, is the energy consumption coefficient, is the speed. After obtaining the standardized objective function through normalization, the task completion time function, the linearized constraint equation and the path cost matrix are combined to finally obtain a dual-objective optimization model. This model aims to optimize the execution time and energy consumption of the task at the same time.

[0027] In one example, a dual-objective optimization model is solved by particle swarm optimization. Through two-stage iterative calculations of adaptive weight strategy and constraint transformation, a non-dominated solution set is obtained, including: The first stage of the dual-objective optimization model is initialized, and an M×N-dimensional particle swarm initial solution is generated according to the total number of tasks M and the total number of AGVs N, and the initial solution space is obtained; The initial solution space is input into the adaptive particle swarm algorithm, and the inertia weight value is calculated by the linear reduction method to obtain the adaptive weight parameter; The position and velocity of the particle swarm are updated based on the adaptive weight parameters, and the global optimal solution is screened using the tournament selection strategy to obtain the first-stage search solution set; Applying the ε-constraint method to the first-stage search solution set, the energy consumption target is transformed into inequality constraints, and the constraint transformation model is obtained; Based on the constrained transformation model, the population is evolved through binary crossover operator and Gaussian mutation operator to obtain the initial population of the second stage; The initial population of the second stage is optimized by local search, and the particle position is fine-tuned by using the variable neighborhood search strategy to obtain the local optimal solution set; The local optimal solution set is stratified according to the non-dominated sorting method, the congestion distance is calculated and the solution set is screened in combination with the elite retention strategy to obtain the candidate non-dominated solution set. The Pareto dominance relationship of the candidate non-dominated solution set is then determined, and the solution with the best comprehensive performance in terms of task completion time and energy consumption is selected to obtain the non-dominated solution set.

[0028] In this example, during the initialization of the first phase, the total number of tasks And the total number of AGVs Generate a dimensional particle swarm initial solution space. These particles represent possible task allocation schemes, and each particle represents a task allocation matrix. Assume that the position of each particle is represented by To indicate that is a binary variable representing the task Is it assigned to AGV? .Right now: ; The position of each particle corresponds to a task allocation scheme, and the solution space in the entire particle swarm is all possible task allocation schemes. The initial solution space is input into the adaptive particle swarm algorithm, which is an optimization method that simulates the foraging behavior of bird flocks in nature. In this process, particles adjust their positions based on their personal experience and global experience and approach the optimal solution. The inertia weight value is calculated by the linear decreasing method to update the speed and position of the particle. Assume that the speed update formula of the particle swarm algorithm is: ; in, It is a particle In Dimension The speed on is the inertia weight, is the learning factor, is a randomly generated value between [0,1]. It is a particle In Dimension The best historical position on is the dimension of the global optimal solution By continuously updating the speed and position of particles, the particle swarm algorithm can make particles gradually approach the global optimal solution. In this process, the tournament selection strategy is used to screen out the global optimal solution in the particle swarm and obtain the search solution set of the first stage. The tournament selection strategy refers to randomly selecting a number of particles in the particle swarm and then selecting the best particle as the representative of the next generation. Through this strategy, the particle swarm can focus on exploring the optimal solution area and avoid falling into the local optimum. Apply the first stage search solution set -Constraint method, which transforms the energy consumption target into inequality constraints. -Constraint method is an effective method to deal with multi-objective optimization problems. It takes one objective function as the optimization goal and other objective functions as constraints to ensure that other constraints are met under the optimization goal. For example, suppose the original dual-objective optimization model is as follows: ; in, Indicates the task completion time. Represents energy consumption. Transformed into constraints, we get a new optimization model: ; in, is the maximum energy consumption value set in advance. The constraint transformation method ensures that the optimization process not only focuses on the task completion time, but also takes energy consumption into consideration. Based on the constraint transformation model, population evolution is performed through binary crossover operator and Gaussian mutation operator. The binary crossover operator generates a new solution by exchanging partial solutions of two particles, while the Gaussian mutation operator fine-tunes the solution of the particle by adding Gaussian noise. The binary crossover operator is expressed as: ; The update method of Gaussian mutation operator is: ; in, The mean is 0 and the variance is Gaussian noise. Through these operations, the particle swarm can generate new solutions in the solution space. In the second stage, the initial population is optimized for local search. Local search optimization adopts a variable neighborhood search strategy to improve the quality of the solution by finding the optimal solution near the current solution. Variable neighborhood search enables the algorithm to avoid local optimality and find the global optimal solution by gradually adjusting the neighborhood structure. Each time a local search is performed, the particle solution is fine-tuned within a certain range. This process takes into account the specific constraints of task scheduling, such as the task completion time and energy consumption of each AGV. Through local search optimization, the particle solution will gradually approach the optimal solution of the problem. After the local search optimization is completed, the particle swarm is non-dominated sorting. The non-dominated sorting method is a method of stratifying the solution set and determining the quality of the solution by comparing the dominance relationship between the solutions. Let solution A dominate solution B, that is, if solution A is not inferior to solution B in all objectives and is better than solution B in at least one objective, then solution A dominates solution B. The result of non-dominated sorting is that the solution set is divided into multiple levels, and the higher the ranking, the closer the solution is to the optimal solution. After the non-dominated sorting is completed, the crowding distance of each solution is calculated to measure the density of each solution in the solution set. Solutions with smaller crowding distances are often located at the edge of the solution space and represent the global optimal solution. By combining crowding distance and elite retention strategy, the most potential solutions are screened out to obtain the candidate non-dominated solution set. The Pareto dominance relationship of the candidate non-dominated solution set is determined, and the solution with the best comprehensive task completion time and energy consumption is selected as the final solution. These final solutions represent the best scheduling solutions that can effectively balance task completion time and energy consumption in multi-AGV workshops. The non-dominated solution set is obtained through the particle swarm optimization algorithm process.

[0029] In one example, a multi-dimensional evaluation calculation is performed on the non-dominated solution set to obtain an optimal scheduling execution plan, including: Extract the time index for each solution in the non-dominated solution set, including the total task completion time and AGV waiting time, and obtain the standardized time index; The efficiency index is calculated for the non-dominated solution set, and the AGV equipment utilization rate and production task completion rate are substituted into the linear programming model to obtain the standardized efficiency index. The running energy consumption data of each AGV in the non-dominated solution set is accumulated and calculated, and the total energy consumption is divided by the number of completed tasks to obtain the standardized energy consumption index; The standardized time index, standardized efficiency index, and standardized energy consumption index are input into the hierarchical analysis network to establish a three-layer evaluation structure, and the consistency calculation of the judgment matrix in the three-layer evaluation structure is performed to obtain the index weight matrix; Based on the indicator weight matrix, the standardized time indicator, standardized efficiency indicator, and standardized energy consumption indicator are comprehensively calculated to obtain an evaluation score sequence. Multi-AGV simulation is performed based on the evaluation score sequence to select the optimal scheduling execution plan.

[0030] In this example, time indicators are extracted for each solution in the non-dominated solution set, including the total task completion time and the AGV waiting time. The total task completion time is expressed as ,in It is The total time to complete a task is the time from the start of the task to the completion of the last task. It refers to the idle time spent by each AGV during the entire task execution process, except for the execution of tasks. It is calculated by the following formula: ; in, It refers to the task Is it assigned to AGV? , if assigned, 1, otherwise It's a task The execution time, is the total time to complete all tasks. For each solution, the calculated total task completion time and AGV waiting time are normalized. Normalized time index Calculated by the following formula: ; in, is the time index of the current solution, and are the minimum and maximum time indicators of all solutions respectively. Through standardization, the influence of scale differences between different solutions is eliminated, so that the time indicators can be compared on a unified scale. The efficiency index is calculated for the non-dominated solution set. The efficiency index is measured by the utilization rate of AGV equipment and the completion rate of production tasks. AGV equipment utilization rate It refers to the ratio of the actual working time of the AGV during the entire operation process to the total available time, which is calculated by the following formula: ; in, It is AGV The time actually involved in the work during the task execution, Is the total time to complete the task. Production task completion rate It refers to the ratio of the number of tasks completed within the specified time to the total number of tasks, expressed as: ; in, It's a task The indicator variable of whether the task is completed is 1 if completed, otherwise it is 0. Substitute the AGV equipment utilization rate and production task completion rate into the linear programming model to obtain the standardized efficiency index. The efficiency index is normalized by comparing it with the minimum and maximum values ​​among all solutions, similar to the normalization of the time index: ; in, is the efficiency index of the current solution, and are the minimum and maximum efficiency indicators of all solutions respectively. The running energy consumption data of each AGV in the non-dominated solution set is accumulated and the total energy consumption is divided by the number of completed tasks to obtain the standardized energy consumption index. The energy consumption of AGV is related to its working time, load, and motion trajectory. Assume that the energy consumption of each AGV is It is expressed as: ; in, Is AGV in time The energy consumption function, Is AGV in time speed, Is AGV in time The load, It is AGV Working hours. Total energy consumption Is the sum of the energy consumption of all AGVs: ; Divide the total energy consumption by the number of completed tasks to get the standardized energy consumption index The calculation formula is: ; Similarly, through standardization, the energy consumption indicators can be compared on the same scale. The standardized time indicator, standardized efficiency indicator and standardized energy consumption indicator are input into the hierarchical analysis network to establish a three-layer evaluation structure. The hierarchical analysis network is a multi-criteria decision-making method, including the target layer, the criterion layer and the plan layer. The target layer is the optimal scheduling execution plan, the criterion layer includes three indicators: time, efficiency and energy consumption, and the plan layer includes all possible scheduling plans. The key step of the hierarchical analysis network is to establish a judgment matrix and perform consistency calculations on it. Assuming that the relative importance of time, efficiency and energy consumption is to be compared, a judgment matrix is ​​constructed. , where each element Representation Guidelines Relative to the criteria According to the principle of hierarchical analysis network, the judgment matrix should meet the consistency requirement. To verify whether the consistency is met, the formula is: ; in, is a consistency indicator, is a random consistency indicator. If If it is less than 0.1, the judgment matrix is ​​considered to have good consistency, otherwise the judgment matrix needs to be adjusted. Based on the judgment matrix, the weight matrix of each standard is calculated ,in They are the weights of time, efficiency and energy consumption respectively. The weight matrix is ​​obtained by the eigenvalue method or by directly solving the linear equations. Based on the indicator weight matrix, the standardized time indicator, the standardized efficiency indicator and the standardized energy consumption indicator are comprehensively calculated. The comprehensive calculation formula is: ; in, is the comprehensive score of each scheduling solution, They are the standardized time, efficiency and energy consumption indicators, are their weights. According to the comprehensive score sequence All scheduling plans are sorted and the actual performance of different scheduling plans is verified through multi-AGV simulation. The operation of different scheduling plans in the actual workshop environment is simulated to evaluate the task completion, AGV operation efficiency and energy consumption performance. Based on the simulation results, the scheduling plan with the highest comprehensive score is selected as the optimal scheduling execution plan.

[0031] In one example, the optimal scheduling execution plan is input into the AGV control system, and the hybrid scheduling prediction model is updated by model feedback to obtain the target scheduling prediction model, including: The optimal scheduling execution plan is decomposed into a time series, and the task sequence is converted into speed control instructions through the AGV kinematic equation to obtain the AGV real-time control sequence; The position coordinate data and task completion status data in the AGV real-time control sequence are sampled, and the data is filtered by a third-order Kalman filter with a noise coefficient of 0.1 to obtain the AGV operation state matrix; The AGV running state matrix is ​​input into the recursive neural network model, and the motion trajectory of the AGV is calculated through the long short-term memory unit to obtain the state prediction sequence. The state prediction sequence is then subjected to collision risk detection. The minimum safe distance between AGVs is calculated based on the workshop layout matrix, and the safety threshold is set to 2 times the vehicle body length to obtain the obstacle avoidance constraint condition. Replan the AGV local path according to the obstacle avoidance constraints, calculate the optimal path cost through the Bellman equation, and iteratively update the path points to obtain the optimized path sequence; The AGV operation state matrix and the optimized path sequence are combined in time sequence, and the task completion rate and energy consumption rate are statistically calculated to obtain a real-time execution data set; The hybrid scheduling prediction model is updated online. The real-time execution data set is input into the hybrid scheduling prediction model in batches of 32 samples to update the parameters, and the updated model parameters are obtained. The updated model parameters are weightedly integrated with the physical constraint loss function to obtain the target scheduling prediction model.

[0032] In this example, the optimal scheduling execution plan is decomposed into a time series, and the task sequence is converted into the speed control instruction of the AGV. The information such as the start and end time of the task, the task priority, etc. is converted into the speed and position instructions that the AGV should execute. Assume there is a task sequence , each task Including start time , End time , Task location In order to convert the task sequence into speed control instructions, the kinematic equation of AGV is combined, namely: ; in, It is the AGV that performs the task The speed of and The tasks are and location, and is the start and end time of the task. In this way, the task sequence is converted into a series of speed control instructions to generate the real-time control sequence of the AGV. The real-time control sequence of the AGV is sampled and filtered using a third-order Kalman filter. The purpose of the Kalman filter is to remove noise and improve the accuracy of the system state. The general form of the Kalman filter is expressed as: ; in, is the filtered state estimate, is the current measured value, is the Kalman gain, is the measurement matrix. Through this process, we get the AGV operation state matrix , which contains the location information, speed information and other important control parameters of the AGV. The noise coefficient in the third-order Kalman filter is set to 0.1 to ensure good noise suppression during the filtering process. The filtered AGV operation state matrix is ​​input into the recursive neural network model, and the motion trajectory of the AGV is calculated using the long short-term memory unit (LSTM). The LSTM network is a model for processing time series data, which can remember long-term dependent information and is suitable for predicting the motion trajectory of the AGV. Assume that the state matrix of the AGV is ,in Indicates time AGV status data (such as position, speed, etc.), the input of the LSTM network is , whose output is the state prediction for the next moment In this way, the LSTM network can predict the future motion trajectory of the AGV. Perform collision risk detection on the state prediction sequence. Calculate the minimum safe distance between AGVs based on the workshop layout matrix. Assume that the body length of each AGV is , the minimum safe distance between each pair of AGVs in the workshop Must meet: ; If the actual distance between AGVs is less than the minimum safety distance, there is a risk of collision. Collision risk detection is achieved by calculating the distance between AGVs and comparing it with the set safety distance to ensure that AGVs can avoid collisions during operation. Under obstacle avoidance constraints, the path planning of AGVs needs to be readjusted to avoid collisions while ensuring the optimality of the path. The Bellman equation is used to calculate the path cost. The basic form of the Bellman equation is: ; in, Yes Status The best price under is the control input, Is the current state and control input The price of is the discount factor, is the state transfer function. On this basis, the AGV iteratively updates the path cost through the Bellman equation to obtain the optimized path sequence. After the path planning is completed, the AGV's operating state matrix is ​​combined with the optimized path sequence to form a time series combination. Through the statistical calculation of the task completion rate and energy consumption rate, the real-time execution data set is obtained. Task completion rate It is expressed as: ; in, It's a task An indicator variable indicating whether the task is completed. is the total number of tasks. And the energy consumption rate Then: ; in, It is AGV Energy consumption, is the total time to complete the task. Based on the real-time execution data, the hybrid scheduling prediction model is updated online. Assuming that 32 samples are used as a batch for training, the real-time execution data set is input into the hybrid scheduling prediction model for parameter update. The updated model parameters are calculated by the back propagation algorithm. In this process, the loss function of the model is a weighted combination of the prediction error and the physical constraint loss function. Assuming the physical constraint loss function Contains physical constraints such as path constraints and speed constraints, and the prediction loss function Then is the prediction error of the model, and the weighted combination loss function of the two is: ; in, and is the weight coefficient. By adjusting these two coefficients, the influence of physical constraints and prediction errors is balanced. The parameters of the target scheduling prediction model are optimized by the back propagation algorithm to obtain the updated hybrid scheduling prediction model. This model can dynamically adjust the scheduling strategy according to real-time data to ensure that AGV can perform tasks efficiently and stably in complex operating environments.

[0033] Reference Figure 2 This embodiment provides a multi-objective optimization scheduling system for a multi-AGV flexible workshop, including: Collection module 1 is used to collect the layout data of the operation workshop, AGV operation parameters and production task data, and build the basic model of the scheduling system; Training module 2 is used to perform recursive neural network training and physical constraint model integration based on the AGV historical operation data in the basic model of the scheduling system to obtain a hybrid scheduling prediction model; Processing module 3 is used to obtain the AGV state feature vector and task requirement characteristics output by the hybrid scheduling prediction model, and perform subtask decomposition and path search processing, and obtain a dual-objective optimization model through constraint linearization conversion; Solving module 4 is used to solve the dual-objective optimization model by particle swarm optimization, and obtain a non-dominated solution set through two-stage iterative calculation of adaptive weight strategy and constraint conversion; The calculation module 5 is used to perform multi-dimensional evaluation calculation on the non-dominated solution set to obtain the optimal scheduling execution plan; The updating module 6 is used to input the optimal scheduling execution plan into the AGV control system, perform model feedback update on the hybrid scheduling prediction model, and obtain the target scheduling prediction model.

[0034] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0035] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0036] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0037] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0038] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0039] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0040] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A multi-objective optimization scheduling method for a multi-AGV flexible workshop, characterized in that: The following steps are involved: Collect the layout data of the job shop, AGV operation parameters and production task data, and build the basic model of the scheduling system; Based on the AGV historical operation data in the basic model of the scheduling system, recursive neural network training and physical constraint model integration are performed to obtain a hybrid scheduling prediction model; Obtaining the AGV state feature vector and task requirement feature output by the hybrid scheduling prediction model, performing subtask decomposition and path search processing, and obtaining a dual-objective optimization model through constraint linearization transformation; The dual-objective optimization model is solved by particle swarm optimization, and a non-dominated solution set is obtained through two-stage iterative calculation of adaptive weight strategy and constraint conversion; Performing multi-dimensional evaluation calculation on the non-dominated solution set to obtain an optimal scheduling execution plan; The optimal scheduling execution plan is input into the AGV control system, and the hybrid scheduling prediction model is updated through model feedback to obtain a target scheduling prediction model.

2. The multi-objective optimization scheduling method for a multi-AGV flexible workshop according to claim 1 is characterized in that: The collection of job shop layout data, AGV operating parameters and production task data, and the construction of a basic model of the scheduling system include: Collect the coordinate data of processing equipment and the location data of AGV charging stations in the workshop, and perform topological structure modeling on the AGV driving channel to obtain the workshop layout topology data; Coordinate mapping of the AGV motion trajectory is performed according to the workshop layout topology data to obtain a workshop layout matrix; The driving speed, turning radius, load capacity and battery capacity of the AGV equipment are collected by sensors and fitted by the least square method to obtain the AGV dynamic parameter model. The workpiece processing process data, process transfer time data and task priority data are input into the task scheduler for linear programming modeling to obtain the production task sequence. An AGV kinematic model is established based on the workshop layout matrix, the AGV dynamic parameter model and the production task sequence, and the AGV velocity vector and position vector are substituted into the Euler equation to obtain the AGV motion constraint conditions; The AGV motion constraint conditions and task time windows are constrained and modeled to obtain a scheduling constraint model. The speed data and battery capacity data in the AGV dynamic parameter model are subjected to energy consumption integration operations to obtain an energy consumption function. The energy consumption function is integrated with the scheduling constraint model to obtain a scheduling system basic model.

3. The multi-objective optimization scheduling method for a multi-AGV flexible workshop according to claim 2 is characterized in that: The hybrid scheduling prediction model is obtained by performing recursive neural network training and physical constraint model integration based on the AGV historical operation data in the scheduling system basic model, including: The AGV running trajectory data, task execution time data and energy consumption data in the basic model of the scheduling system are divided into data sets, and time series segmented to obtain a training sample set; The training sample set is input into a batch normalization layer for data standardization, and the trajectory data is enhanced by random rotation and translation to obtain an enhanced training data matrix; The enhanced training data matrix is ​​divided into batches, and time series features are extracted through long short-term memory network units to obtain AGV state feature vectors; Inputting the AGV state feature vector into a recursive neural network including three hidden layers, performing parameter optimization through an Adam optimizer, and obtaining a first neural network model; Performing dropout processing on the hidden layer of the first neural network model, and triggering an early stopping mechanism when the verification loss does not decrease for N consecutive times, to obtain a second neural network model; Based on the basic model of the scheduling system, the AGV kinematic equation, time window constraint and energy consumption function are constructed as Lagrangian constraint terms to obtain the physical constraint loss function; The physical constraint loss function and the neural network prediction loss are weightedly combined to obtain a hybrid optimization objective function, and the hybrid optimization objective function is input into the second neural network model, and a hybrid scheduling prediction model is obtained by iterative calculation using a back propagation algorithm.

4. The multi-objective optimization scheduling method for a multi-AGV flexible workshop according to claim 3 is characterized in that: The step of obtaining the AGV state feature vector and task requirement feature output by the hybrid scheduling prediction model, performing subtask decomposition and path search processing, and obtaining a dual-objective optimization model through constraint linearization conversion includes: The similarity between the AGV state feature vector and the task requirement feature output by the hybrid scheduling prediction model is calculated, and the task similarity matrix is ​​obtained by weighted combination of cosine distance and Euclidean distance. Tasks are clustered according to the similarity threshold to obtain a task decomposition sequence. Map and match each subtask in the task decomposition sequence with the AGV device, construct an M×N dimensional task allocation matrix through 0-1 integer programming, where M is the total number of tasks and N is the total number of AGVs, and obtain an initial allocation plan; Encoding the execution task sequence of each AGV in the initial allocation scheme, and obtaining the task optimization sequence through single-point crossover and double-point mutation operations; Constructing the A* algorithm path planning space according to the workshop layout matrix, setting the distance between nodes to the length of one AGV body, and rasterizing the space by the grid method to obtain a path search network; A heuristic function is set for the path search network, and a weighted sum of the Manhattan distance and the dynamic obstacle avoidance factor is used as a path evaluation function to obtain a path cost matrix; Performing a first-order Taylor expansion on the AGV motion constraint condition, converting the AGV kinematic equation, turning constraint and speed constraint into a set of linear inequalities, and obtaining a linearized constraint equation; The task completion time function is calculated based on the task optimization sequence, the energy consumption function is calculated for the AGV dynamic parameter model, and a standardized objective function is obtained through normalization processing. The standardized objective function, the linearized constraint equation and the path cost matrix are combined to obtain a dual-objective optimization model.

5. The multi-objective optimization scheduling method for a multi-AGV flexible workshop according to claim 4 is characterized in that: The dual-objective optimization model is solved by particle swarm optimization, and a non-dominated solution set is obtained through two-stage iterative calculation of adaptive weight strategy and constraint conversion, including: Initializing the dual-objective optimization model in the first stage, generating an M×N-dimensional particle swarm initial solution according to the total number of tasks M and the total number of AGVs N, and obtaining an initial solution space; The initial solution space is input into the adaptive particle swarm algorithm, and the inertia weight value is calculated by a linear decreasing method to obtain an adaptive weight parameter; Based on the adaptive weight parameters, the position and speed of the particle swarm are updated, and the global optimal solution is screened using the tournament selection strategy to obtain the first-stage search solution set; Applying the ε-constraint method to the solution set searched in the first stage, the energy consumption target is converted into inequality constraints to obtain a constraint conversion model; Based on the constraint transformation model, the population is evolved by using a binary crossover operator and a Gaussian mutation operator to obtain an initial population of the second stage; The initial population of the second stage is optimized by local search, and the particle position is fine-tuned by using the variable neighborhood search strategy to obtain the local optimal solution set; The local optimal solution set is stratified according to the non-dominated sorting method, the congestion distance is calculated and the solution set is screened in combination with the elite retention strategy to obtain a candidate non-dominated solution set, and the Pareto dominance relationship of the candidate non-dominated solution set is determined, and the optimal solution in terms of task completion time and energy consumption is selected to obtain a non-dominated solution set.

6. The multi-objective optimization scheduling method for a multi-AGV flexible workshop according to claim 5 is characterized in that: The performing multi-dimensional evaluation calculation on the non-dominated solution set to obtain the optimal scheduling execution plan includes: Extracting time indicators for each solution in the non-dominated solution set, including the total task completion time and the AGV waiting time, to obtain a standardized time indicator; Calculate the efficiency index of the non-dominated solution set, substitute the AGV equipment utilization rate and the production task completion rate into the linear programming model, and obtain the standardized efficiency index; The operation energy consumption data of each AGV in the non-dominated solution set is accumulated and calculated, and the total energy consumption is divided by the number of completed tasks to obtain a standardized energy consumption index; Input the standardized time index, the standardized efficiency index, and the standardized energy consumption index into a hierarchical analysis network to establish a three-layer evaluation structure, and perform consistency calculation on the judgment matrix in the three-layer evaluation structure to obtain an index weight matrix; Based on the indicator weight matrix, the standardized time indicator, the standardized efficiency indicator, and the standardized energy consumption indicator are comprehensively calculated to obtain an evaluation score sequence, and multi-AGV simulation is performed based on the evaluation score sequence to select the optimal scheduling execution plan.

7. The multi-objective optimization scheduling method for a multi-AGV flexible workshop according to claim 6 is characterized in that: The optimal scheduling execution plan is input into the AGV control system, and the hybrid scheduling prediction model is updated by model feedback to obtain a target scheduling prediction model, including: Decomposing the optimal scheduling execution plan into a time series, converting the task sequence into a speed control instruction through the AGV kinematic equation, and obtaining the AGV real-time control sequence; The position coordinate data and task completion status data in the AGV real-time control sequence are sampled, and the data is filtered by a third-order Kalman filter with a noise coefficient of 0.1 to obtain an AGV operation state matrix; The AGV operation state matrix is ​​input into the recursive neural network model, the motion trajectory of the AGV is calculated through the long short-term memory unit, and a state prediction sequence is obtained. The state prediction sequence is subjected to collision risk detection, and the minimum safe distance between AGVs is calculated based on the workshop layout matrix. The safety threshold is set to 2 times the vehicle body length to obtain the obstacle avoidance constraint condition. Replanning the AGV local path according to the obstacle avoidance constraint condition, calculating the optimal path cost by Bellman equation, and iteratively updating the path points to obtain an optimized path sequence; The AGV operation state matrix and the optimized path sequence are combined in time sequence, and the task completion rate and energy consumption rate are statistically calculated to obtain a real-time execution data set; The hybrid scheduling prediction model is updated online, and the real-time execution data set is input into the hybrid scheduling prediction model in batches of 32 samples to update the parameters to obtain updated model parameters, and the updated model parameters are weightedly integrated with the physical constraint loss function to obtain the target scheduling prediction model.

8. A multi-objective optimization scheduling system for a multi-AGV flexible workshop, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the system comprises: The acquisition module is used to collect the layout data of the job shop, AGV operation parameters and production task data, and build the basic model of the scheduling system; A training module, used to perform recursive neural network training and physical constraint model integration based on the AGV historical operation data in the basic model of the scheduling system to obtain a hybrid scheduling prediction model; A processing module, used to obtain the AGV state feature vector and task requirement features output by the hybrid scheduling prediction model, and perform subtask decomposition and path search processing, and obtain a dual-objective optimization model through constraint linearization conversion; A solution module, used for performing particle swarm optimization on the dual-objective optimization model, and obtaining a non-dominated solution set through two-stage iterative calculation of an adaptive weight strategy and constraint conversion; A calculation module, used for performing multi-dimensional evaluation calculation on the non-dominated solution set to obtain an optimal scheduling execution plan; The updating module is used to input the optimal scheduling execution plan into the AGV control system, perform model feedback update on the hybrid scheduling prediction model, and obtain a target scheduling prediction model.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images