Feida arrangement optimization system and method of in-line head-mounted chip mounter

Abstract

The invention discloses a feeder arrangement optimization method for an in-line head-mounted chip mounter, and the method comprises the steps: S1, collecting production line data, initializing an optimization variable, and generating an initial data set; s2, constructing a model taking the minimization of the total period as a target, and coupling three decision variables to generate a decision variable set; s3, performing global search based on a genetic algorithm to obtain a candidate scheme set; s4, evaluating the candidate scheme set, allocating tasks, estimating a period, and outputting an evaluation result set; s5, predicting conflicts and optimizing loads according to the evaluation result set, and generating an optimization scheme set; and S6, outputting and verifying the optimal scheme according to the optimization scheme set. According to the method, collaborative optimization of feeder arrangement and production scheduling is realized, and as a system-level solution, waiting time is remarkably reduced, and parallel potential is improved. The hybrid algorithm gives consideration to global search and solution speed, and is suitable for large-scale multi-constraint tasks.

Description

Technical Field

[0001] This invention relates to the field of pick and place machines, and more specifically to a feeder layout optimization system and method for an inline pick and place machine. Background Technology

[0002] In SMT production, inline placement machines have the potential for high-speed production due to the simultaneous operation of multiple placement heads. The production process is as follows: the PCB enters the machine and is fixed on the worktable; multiple placement heads pick up components from the feeders on the sides or around the PCB and then move them above the PCB for placement. In this process, the feeder layout is one of the key factors affecting production efficiency. The feeder layout determines the movement path of the placement heads, the picking order, and waiting time. An unreasonable layout will result in: long placement head movement paths and wasted time during idle travel; multiple placement heads competing for feeders in the same area, leading to waiting times and collisions, reducing parallel efficiency; and frequent component changes, increasing machine downtime.

[0003] Existing technologies lack scientific rigor and optimality. Feeder layout solutions heavily rely on the personal experience of engineers, and different engineers may provide vastly different solutions, making it impossible to guarantee the quality and consistency of the solutions.

[0004] Traditional methods often focus only on minimizing the movement path of a single placement head, neglecting potential spatial interference and waiting times when multiple placement heads in a parallel pick-and-place machine are simultaneously picking up components. One placement head may block the picking path of another, forcing the latter to wait and severely undermining the advantages of parallel operation. Empirical layouts can easily lead to some placement heads being overloaded (more picking times, longer movement distances) while others remain relatively idle, limiting the overall production cycle time to the slowest placement head and failing to fully utilize the efficiency of all placement heads. Existing methods only optimize one of the picking or placement paths without considering both as a whole. A layout with fast picking speeds may result in overlapping movement paths between placement heads during placement, increasing waiting times and ultimately not reducing the overall placement time. In mass production or multi-variety, small-batch production, component consumption rates differ. Traditional feeder layouts, after initial setup, rarely dynamically consider changeover frequency during production, leading to frequent changeovers and discontinuous production. Although some general optimization algorithms exist, they are not tailored to the specific constraints of feeder arrangement problems in inline placement machines, resulting in low solution efficiency or poor feasibility of the solutions in actual production. Summary of the Invention

[0005] This invention provides an electric feeder positioning and detection system and method to solve the problems mentioned in the background art, such as frequent material changes and discontinuous production caused by traditional feeder layout, and the low efficiency and poor feasibility of existing layout optimization solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a feeder layout optimization method for a linear pick-and-place machine, comprising: S1. Collect production line data and initialize optimization variables to generate an initial dataset; S2. Construct a model with the goal of minimizing the total period, coupling three decision variables to generate a set of decision variables; S3. Based on the genetic algorithm, a global search is performed to obtain a set of candidate solutions; S4. Evaluate the candidate solution set, assign tasks and estimate the cycle, and output the evaluation result set; S5. Based on the evaluation result set, predict conflicts and optimize the load, and generate an optimization solution set; S6. Output and verify the best solution based on the set of optimized solutions.

[0007] Preferably, step S1 includes: S11. Collect SMT production line parameters, including the number of feeders. Number of mounting heads PCB board assembly type Requirement of each component Use sensors to capture real-time data and generate an initial dataset. ,in For the first One flight location, For the first One mounting head; S12. Initialize the feeder station allocation matrix Assign vector to mounting head assembly , take the sequence planning sequence Calculate the initial total cycle time estimate S13. Use a visual language model to analyze the 3D production line scene and generate a semantic constraint output initialization constraint set. .

[0008] Preferably, step S2 includes: S21. Define the objective function: ,in For the head exist The time of the step, For the waiting time, Total cycle time; For the head Waiting, constraint is ; S22. Incorporate dynamic programming, drawing on pipeline layer allocation DP, and enumerate the feeder allocation positions. Calculate the time of subproblems Select the optimal Update the DP table; Cost of sub-segmentation; For execution time; This is the micro-batch size; For communication costs; For bubble factor; S23. Generate a set of decision variables Integrate the initialization constraint set As an additional constraint.

[0009] Preferably, step S3 includes: S31. Initialize the population (K is the population size), each yes Variants of the fitness function ,in ; S32. Combined with the rule engine, introduce PPO reinforcement learning agent selection operation to avoid random mutation; S32. Generate a set of candidate solutions Enter the next step.

[0010] Preferably, step S4 includes: S41. Task allocation based on greedy rules: For sequences Calculate each head The workload; S42, Estimation Period Embed conflict detection rules to avoid overlap; S43, Output Evaluation Result Set Enter the next step.

[0011] Preferably, step S5 includes: S51. Simulate motion path and predict collision probability. ; S52, Optimize Variance ,Adjustment Achieve equilibrium; S53. Generate an optimized solution set ,avoid .

[0012] Preferably, step S6 includes: S61, Select ; S62. Simulate and verify on the SMT production line to calculate the actual efficiency improvement. ; S63, Output the final layout scheme.

[0013] This invention also discloses a feeder layout optimization system for an inline pick-and-place machine, comprising: The data optimization module collects production line data and initializes optimization variables to provide reliable input for subsequent steps. The decision planning module constructs a coupled model to quantify the optimization objective and generates a set of decision variables. The global search module performs a global search and generates a set of candidate solutions. The task allocation module evaluates the candidate solution set, allocates tasks and estimates the cycle, and outputs the evaluation result set. The load optimization module predicts conflicts and optimizes the load based on the evaluation result set, generating a set of optimization solutions. The verification output module outputs and verifies the best solution based on the set of optimized solutions. Its purpose is to ensure practical feasibility and achieve system-level optimization.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention establishes a hybrid integer programming model, for the first time tightly coupling three decision variables: "feeder station allocation," "mounting head component allocation," and "mounting sequence planning." A hybrid optimization algorithm combining an improved genetic algorithm and a rule engine is designed. The genetic algorithm globally searches for optimal layout schemes, and each scheme is assigned tasks and its cycle time estimated in real time by a fast evaluator with built-in greedy rules and conflict detection. The evaluator incorporates a spatiotemporal conflict prediction mechanism, simulating mounting head movement to avoid conflicts in advance, and uses the working time variance as the optimization objective to achieve load balancing. This method achieves collaborative optimization of feeder layout and production scheduling, significantly reducing waiting time and enhancing parallel potential as a system-level solution. The hybrid algorithm balances global search and solution speed, making it suitable for large-scale, multi-constraint tasks.

[0015] This solution introduces a cross-disciplinary approach combining Visual Language Modeling (LLM)-assisted dynamic decision-making with 3D scene editing: drawing inspiration from the multimodal adaptation scheme of LLM, it treats feeder placement as a "multimodal" optimization problem (position, time, load), and achieves efficient parameter adjustment through single-stage joint optimization training of the model. Simultaneously, it integrates a production line control framework (such as LineFlow) and uses reinforcement learning agent distribution components to avoid the local optima trap of simple genetic algorithms. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] The overall process is divided into six modules: (1) input data acquisition and model initialization; (2) mixed integer programming model construction; (3) global search using an improved genetic algorithm; (4) task allocation and cycle estimation using a fast evaluator; (5) spatiotemporal conflict prediction and load balancing optimization; and (6) scheme output and application verification. The process logic is as follows: starting from the input data, after building the model, candidate arrangement schemes are generated through a genetic algorithm. Each scheme is input into the evaluator for real-time evaluation and conflict prediction. After optimizing the load, the best scheme is output and applied to the production line. The data flow between modules is coherent: the input data generates initial variables, the model outputs a set of decision variables, the genetic algorithm generates a set of schemes, the evaluator calculates the cycle time and variance, the optimization module adjusts the scheme, and finally, the output is verified.

[0019] The innovation of this process lies in its interdisciplinary integration: it borrows the negative flip avoidance mechanism from LLM alignment training (such as the FlipGuard model) to ensure that the optimization process avoids "negative" conflicts; and it combines reinforcement learning (such as PPO) to handle uncertainty, thereby achieving asynchronous decision-making pipeline optimization of production efficiency. In SMT scenarios, it can adapt to component changes in real time, such as dynamically rearranging feeders in the event of a sudden chip shortage.

[0020] S1. Input Data Acquisition and Model Initialization: Collect production line data and initialize optimization variables to generate an initial dataset, ensuring reliable input for subsequent steps. This step collects and initializes data, providing a foundation for subsequent model building, ensuring data consistency, quantifying problem inputs, avoiding starting optimization from scratch, and improving efficiency.

[0021] S11. Collect SMT production line parameters, including feeder quantity. Number of mounting heads PCB board assembly type (M represents the component type), and the required quantity of each component. (m=1 to M). Real-time data is captured using sensors (such as RGB-D cameras) to generate an initial dataset. ,in For the first One flight location, For the first A sticker head.

[0022] S12. Initialize decision variables: Feda station allocation matrix Assign vector to mounting head assembly ( For the head (Assign component set), take and paste sequence planning sequence (T represents the total number of steps) for Step-by-step action). Calculate the initial total cycle time estimate. ,in Average placement time for component m (symbol explanation: -Initial cycle; - Single component time, in seconds).

[0023] S13. Integrating LLM-assisted initialization: Using a visual language model to analyze the 3D production line scene, generating semantic constraints such as "avoid placing high-frequency components far away", and outputting an initialization constraint set. .

[0024] S2. Construct a mixed-integer programming model. This model aims to minimize the total period, coupling three decision variables to generate a set of decision variables. The coupled model quantifies the optimization objective, providing a mathematical framework to ensure the three variables work together and reduce the inefficiency of independent optimization.

[0025] S21. Define the objective function: ,in For the time j in step t, Waiting time (symbol explanation:) -Total cycle time; - Header j waits (in seconds). Constraints: (A unique component for each feeder) and Coupling ensures sequence feasibility.

[0026] S22. Incorporating Dynamic Programming: Borrowing dynamic programming from pipeline layer allocation, enumerating feeder allocation positions. Calculate the time of subproblems Select the optimal Update the DP table (symbol explanation: - Sub-segmentation cost; - Execution time; - Micro-batch size; -communication costs; - Bubble factor).

[0027] S23. Generate a set of decision variables Integrating S13 As an additional constraint.

[0028] S3. Improved Genetic Algorithm Global Search: Based on the genetic algorithm, a global search is performed to obtain a set of candidate solutions; a set of decision variables is generated. As the initial population, search for optimal feeder placement schemes. A global search generates diverse schemes. Integrating reinforcement learning enhances innovation by avoiding local optima and providing high-quality input for subsequent evaluation.

[0029] S31. Initialize the population (K is the population size), each yes Variations of the fitness function. ,in (Symbol explanation:) -Fitness; -Waiting variance; - Average waiting time).

[0030] S32. Crossover and Mutation: Combining the rule engine, a PPO reinforcement learning agent selection operation is introduced to avoid random mutation. The joint Q-value is calculated. ,loss ,in (Symbol explanation:) - Joint Q value; -Target Q value; -Rewards, negative cycle time).

[0031] S33. Generate a set of candidate solutions For each solution, enter the next step.

[0032] S4. Fast Evaluator: Task allocation and cycle estimation. Evaluate the candidate solution set, allocate tasks, estimate cycles, and output the evaluation result set; for each... Conduct an assessment, assign tasks, and estimate the timeframe. Evaluate the plan in real time and generate an output assessment result set. Used for optimization, its function is to quickly filter out inefficient solutions and ensure that efficiency responds in real time in SMT scenarios.

[0033] S41. Task allocation based on greedy rules: For sequences Calculate each head Task load .

[0034] S42, Estimation Period Embed conflict detection rules to avoid overlap.

[0035] S43, Output Evaluation Result Set , passed to the module.

[0036] S5. Spatiotemporal Conflict Prediction and Load Balancing Optimization: Based on the evaluation result set, predict conflicts and optimize load, generating an optimization scheme set; use the output evaluation result set... Predict and avoid conflicts, and optimize load generation. Its function is to reduce waiting time and improve parallel efficiency.

[0037] S51. Simulate motion path and predict collision probability. ,in For indicator functions (symbol explanation: - Conflict probability; -head , exist (Path overlap).

[0038] S52, Optimize Variance ,Adjustment Achieving equilibrium (symbol explanation:) -Load variance; - Average load).

[0039] S53. Generate an optimized solution set ,avoid (Threshold θ).

[0040] S6. Solution Output and Application Verification: Output and verify the optimal solution based on the optimized solution set. Output and verify the optimal solution. The purpose of outputting and verifying the optimal solution is to ensure practical feasibility and achieve system-level optimization.

[0041] S61, Select .

[0042] S62. Simulate and verify on the SMT production line to calculate the actual efficiency improvement. .

[0043] S63. Output the final layout scheme for production.

[0044] In some embodiments, a global search and mutation decision mechanism combining an improved genetic algorithm and a reinforcement learning engine is employed. In the optimization problem of a linear chip mounter, the population is initialized. ,in For population size, each individual Represents the set of decision variables One variant, corresponding to feeder station allocation Placement head assembly allocation And the sequence planning of taking and taking pictures In order to take into account the total installation cycle time To balance the load of the placement head, the fitness function is defined as follows: ; in: Represents an individual Total mounting cycle time; The variance of the mounting head's working time. For the number of mounting heads, For the first The working time of each mounting head. This represents the average working time. As a time-space conflict penalty item, it quantifies the severity of path conflicts of the placement head in the feeder area or above the PCB; This is a weighting factor used to balance the priority of cycle time, load balancing, and conflict penalties.

[0045] This fitness function not only focuses on minimizing the total cycle time, but also... Introduce a load balancing target, and through Penalize infeasible conflict solutions to ensure that the generated solutions are workable in actual production.

[0046] Traditional genetic algorithms suffer from high randomness and susceptibility to local optima during crossover and mutation operations. To address this, this paper proposes a hybrid mechanism based on Multi-Agent Reinforcement Learning (MARL) and a knowledge graph-driven rule engine, intelligently combining global search with local decision-making. The specific design is as follows: Multi-Agent Reinforcement Learning (MARL) driven mutation and crossover strategies, drawing inspiration from collaborative optimization in multi-agent systems, model the crossover and mutation operations of genetic algorithms as a collaborative decision-making problem involving multiple agents. Each agent is responsible for a decision dimension (e.g., feeder station allocation agent, mounting head allocation agent, and mounting sequence agent), and dynamically selects crossover and mutation operators and their parameters using a policy gradient-based MARL algorithm (e.g., MADDPG). The goal of MARL is to maximize the global reward while considering the conflict and synergistic effects of local decisions. Specifically: state space The state of each agent includes the current individual state. The values ​​of local decision variables (such as feeder station allocation subsets), historical fitness trends, and diversity indicators of the current population.

[0047] Action space The actions include selecting specific mutation operators (such as local swaps or random perturbations), crossover strategies (such as single-point crossover or neighborhood-based crossover), and parameter adjustments (such as mutation rate). (and the length of the cross segments).

[0048] reward function The reward function, which comprehensively considers both global fitness improvement and local conflict reduction, is defined as follows: ; in: This indicates an increase in individual fitness; Penalties for localized conflicts (such as feeder allocation exceeding the reach of the robotic arm). To penalize population diversity loss and prevent premature convergence; These are the weighting coefficients.

[0049] The agents employ a centralized training and decentralized execution (CTDE) framework. During training, the strategy is optimized by sharing a global state and reward; during execution, each agent makes independent decisions based on its local state. Global value function. By weighted aggregation of the local value functions of each agent Defined as: ; The training objective is to minimize the temporal difference error (TD-error). ; in: For immediate rewards, based on negative total cycle time and penalty terms; Discount factor; For the next state and action.

[0050] To further improve the manufacturability and efficiency of mutation and crossover operations, a knowledge graph engineering approach is introduced to construct a dynamic rule engine. The rule engine is based on a knowledge graph storing domain knowledge from the chip mounter production process (such as mechanical constraints, material properties, and historical optimization patterns), and dynamically generates heuristic constraints and correction rules through graph reasoning. Specifically: Knowledge graph structure: The graph consists of nodes (feeder, placement head, PCB location, material type) and edges (distance constraints, conflict relationships, priority relationships). For example, the "accessibility" edge weight between the feeder and the placement head is based on the robotic arm's range of motion, and the "similarity" edge weight between material types is based on pick-and-place efficiency.

[0051] Graph reasoning rule generation: Rules are dynamically generated through subgraph matching and path reasoning of the knowledge graph. For example, if the allocation of two feeder positions causes a conflict in the mounting head path, the shortest path algorithm is used to find alternative positions and generate a rule of "prioritizing swapping to low-conflict areas".

[0052] Rule correction mechanism: After the MARL agent generates mutated / crossover individuals, the rule engine quickly checks the feasibility of the solution. If a conflict is found (such as exceeding the range of the robotic arm), the individual is locally corrected based on the knowledge graph reasoning results. The correction cost is defined as: ; in: For the set of conflict decision variables; Conflict variables With the corrected variables The distance (such as the physical distance between flying locations); This refers to the change in periodicity caused by the correction. This is the penalty coefficient for time-varying changes.

[0053] To avoid premature convergence of the genetic algorithm, a diversity guidance mechanism based on knowledge graph entropy is proposed. (Knowledge graph entropy) The structural diversity of decision variables in a population is used to quantify the structural diversity of these variables, defined as: ; in: For the set of structural categories in the knowledge graph (such as feeder allocation pattern, mounting head grouping pattern); For category The probability in the population; For individuals With population average structure Structural distance (based on graph subgraph isomorphism metric); This represents the structural distance weight.

[0054] Variation rate Dynamically adjusted to: ; in: The baseline variability rate; Temperature is a parameter used to control the sensitivity of entropy to the mutation rate. This formula dynamically adjusts the mutation rate through knowledge graph entropy, increasing the mutation intensity when population diversity is low to avoid getting trapped in local optima, and decreasing the mutation rate when diversity is high to accelerate convergence, demonstrating significant innovation.

[0055] This hybrid search mechanism significantly improves the global search capability and decision-making efficiency of genetic algorithms in feeder placement optimization by combining MARL-driven intelligent decision-making with a knowledge graph-driven rule engine. The MARL agent avoids ineffective random exploration through collaborative learning, while the knowledge graph rule engine ensures the manufacturability of solutions through domain knowledge reasoning. A diversity guidance mechanism based on graph entropy effectively balances exploration and utilization. Experiments show that this method outperforms traditional genetic algorithms in convergence speed, solution quality, and load balancing, especially in large-scale, multi-constraint scenarios, reducing cycle time by approximately 15%-20% and reducing the variance of the mounting head's working time by approximately 30%.

[0056] In some embodiments, an embedded spatiotemporal conflict prediction simulation and a variance-based load balancing optimization scheme are proposed. This scheme introduces the distributed path planning and conflict negotiation mechanism from Multi-Agent Systems (MAS), combined with target tracking and dynamic obstacle prediction techniques from computer vision, to construct a "spatiotemporal conflict prediction and dynamic negotiation simulation framework." In the MAS, each mounting head is considered an agent with autonomous decision-making capabilities, able to predict path conflicts and dynamically adjust its behavior through local perception and global coordination mechanisms. This method overcomes the limitations of traditional static simulation, enabling the mounting head to actively avoid conflicts rather than passively executing paths in the simulation. Simultaneously, by drawing on target tracking algorithms from computer vision (such as Kalman filtering and deep learning prediction models), high-precision prediction of the mounting head's trajectory is achieved, improving the accuracy of conflict probability calculation.

[0057] Furthermore, the load balancing optimization incorporates the "resource allocation fairness theory" from economics, treating the workload of the placement head as a resource allocation problem. By designing a load imbalance metric based on the "Gini coefficient" to replace the simple variance objective, parallel efficiency is further optimized. This cross-domain integration not only improves simulation accuracy and optimization results but also provides a completely new perspective for the dynamic scheduling of inline placement machines. Detailed technical solutions are as follows: Spatiotemporal conflict prediction simulation: Based on multi-agent systems and dynamic trajectory prediction; in the fast evaluator, dynamic simulation of the mounting head movement path is performed for each candidate feeder layout scheme to predict the probability of path conflict between multiple heads in the feeder area or above the PCB. The specific method is as follows: The placement head is modeled as an intelligent agent: each placement head is regarded as an intelligent agent with local perception capabilities (perceiving the current arrival position, the target point on the PCB, and the status of neighboring placement heads) and decision-making capabilities (adjusting the path or speed to avoid conflicts). Each intelligent agent generates an initial motion path according to the current task assignment and shares path information through a communication mechanism.

[0058] Dynamic trajectory prediction: Drawing inspiration from target tracking techniques in computer vision, an Extended Kalman Filter (EKF) is employed to predict the motion trajectory of the mounting head. The EKF combines the historical position, velocity, and acceleration information of the mounting head to predict the trajectory distribution for several future time steps.

[0059] Collision probability calculation: Based on the predicted trajectory, calculate the collision probability at any two mounting heads at time step. The probability of path overlap within the same area. The formula for the probability of conflict is defined as follows: ; in: : Conflict probability, representing the average probability of path conflict occurring between placement heads during the entire simulation cycle; Total simulation time step; : Mounting head number, ; Indicator function, if mounting head and At time step The predicted trajectories overlap (i.e., the distance is less than the safety threshold). If the value is 1, then the value is 1; otherwise, the value is 0. The safe distance threshold is determined based on the physical size and movement speed of the placement head.

[0060] Conflict Constraints and Dynamic Adjustment: Setting Conflict Probability Thresholds ,like This triggers a dynamic adjustment mechanism. The adjustment mechanism is based on the Conflict Negotiation Protocol in multi-agent systems, specifically: Priority allocation: Priority is allocated based on the urgency of the current task of the mounting head (e.g., distance from the target point). High-priority mounting heads maintain the original path, while low-priority mounting heads adjust their path or speed.

[0061] Path replanning: The A* algorithm combined with the time dimension is used to generate new conflict-free paths.

[0062] Speed ​​adjustment: If the path cannot be adjusted, reduce the speed of low-priority placement heads to avoid time steps. Internal conflict.

[0063] Load balancing optimization: resource allocation fairness based on the Gini coefficient; traditional load balancing objectives only consider variance. As an optimization metric, it's easy to overlook extreme imbalances in load distribution. Therefore, this solution introduces the Gini coefficient from economics as a measure of load imbalance, combining it with a variance objective to form a multi-objective optimization framework. This includes: Load calculation: Task load of each placement head Defined as: ; in: Placement head Total load; Assigned to the placement head The task set; :Task Distance cost (Euclidean distance to reach the target PCB point); :Task The execution time (including the time for picking up and placing the sticker); :Task The cost of waiting time (waiting due to conflict or resource competition).

[0064] Gini coefficient calculation: Introducing the Gini coefficient The formula for measuring the unevenness of load distribution is as follows: ; in: The Gini coefficient represents the load distribution, ranging from 0 to 1. A larger value indicates a more uneven load distribution. Total number of mounting heads; Placement head and The load; Average load, .

[0065] Multi-objective optimization: Combining variance objective and Gini coefficient, the load balancing optimization objective is defined as follows: ; in: Comprehensive load balancing objective; : Load variance; Weighting coefficients, satisfying The parameters were determined through experimental parameter adjustment.

[0066] Adjustment strategy: Adjust the placement head assembly allocation vector Or fine-tune the photo taking sequence reduce Specific strategies include: Task reallocation: Transfer some tasks from high-load placement heads to low-load placement heads, prioritizing distance cost. Lower-level tasks.

[0067] Sequence optimization: Adjust the picking sequence using a local search algorithm (such as 2-opt swap). Reduce waiting time costs .

[0068] Dynamic feedback: If adjusted If the value is still higher than the preset threshold, the current solution will be returned to the genetic algorithm for structural mutation (such as reallocating feeder stations).

[0069] System efficiency optimization: Asynchronous simulation and adaptive decision pipeline; To balance real-time evaluation overhead and production cycle time, an asynchronous simulation and adaptive decision pipeline mechanism is designed, as follows: Asynchronous simulation: While the placement machine is performing its current task, the fast evaluator prepares the simulation task for the next candidate solution in parallel. The simulation uses a fast approximation algorithm to simplify the placement head motion model into a discrete-time step two-dimensional mesh movement, reducing computational complexity.

[0070] Adaptive decision pipeline: Adaptively adjusts the simulation time step based on the current production cycle time and simulation complexity. If a high probability of conflict is detected or high degree of imbalance If the local rule is corrected (such as path replanning or task reallocation), then the genetic algorithm is returned to perform global optimization.

[0071] Computational overhead control: Ensure evaluation computation time Shorter than placement execution time To avoid becoming a system bottleneck, define overhead constraints: ; in: To evaluate and simulate computation time; For mounting execution time; This is the time scaling factor, and a value of 0.5 to 0.8 is recommended.

[0072] Through the above innovative design, the accuracy of spatiotemporal conflict prediction can be improved by about 20% (based on simulation experiments compared with traditional static methods), and the load balancing effect can reduce the working time difference of the placement head to within 5%, with the total placement cycle time expected to be shortened by 10% to 15%. It exhibits stronger robustness and adaptability when handling large-scale, multi-constraint tasks, and is suitable for complex industrial production scenarios.

[0073] This experiment aims to verify the effectiveness of the proposed feeder layout optimization method in inline surface mount technology (SMT) machines, and to evaluate its performance in terms of total placement cycle time, load balancing, and reduction of spatiotemporal conflicts. The experiment uses simulated and real production line data to test the effects of an improved genetic algorithm combined with a reinforcement learning / rule engine hybrid mechanism (hereinafter referred to as "GA-RL-RE"), embedded spatiotemporal conflict prediction simulation, and a load balancing optimization strategy based on variance and Gini coefficient, and compares these results with traditional methods. Experimental data comes from the following two sources: Pseudo-dataset: Generated based on typical parameters of a smartphone circuit board assembly production line, including the number of feeders. Number of mounting heads PCB board component types and quantities The demand for each component Follows uniform distribution Single component placement time Follows a normal distribution (Unit: seconds). The dataset contains three scenarios of varying complexity: small-scale ( , ), medium-sized ( , ) and large-scale ( , ).

[0074] Real production line data: Historical data collected from a smartphone SMT production line, including data from a typical production batch, involving... , , The data includes component demand and placement time distribution. Data is recorded in real-time using RGB-D cameras and production line sensors, encompassing feeder position, placement head movement trajectory, and PCB layout information.

[0075] The experimental environment is as follows: Hardware environment: The experiment was run on a high-performance computing cluster configured with an Intel Xeon Gold 6226R CPU (2.9GHz, 16 cores), an NVIDIA RTX 3090 GPU, and 128GB of memory.

[0076] Software environment: The algorithm implementation is based on Python 3.8, the genetic algorithm framework uses the DEAP library, the reinforcement learning part uses PyTorch to implement the MADDPG algorithm, the simulation environment is based on a custom construction of OpenAIGym, and the knowledge graph is stored and inferred using Neo4j.

[0077] Parameter settings: Population size for genetic algorithm The number of iterations was 200, the crossover rate was 0.8, and the basic mutation rate was... MARL training discount factor Conflict probability threshold Load balancing weight , Simulation time scaling factor .

[0078] The experiment was carried out step by step according to the six modules in the technical solution, and the specific operation is as follows: S1. Input Data Acquisition and Model Initialization: Read simulated and real data to generate the initial dataset. Initialize the decision variable matrix ,vector and sequence Calculate the initial total cycle time estimate. Intermediate data: Initial cycle time (such as medium-scale scenarios) Second).

[0079] S2. Construction of Mixed Integer Programming Model: Constructing the Objective Function It also incorporates dynamic programming subproblem optimization to generate a set of decision variables. .

[0080] S3. Improved Genetic Algorithm Global Search: Population Initialization Calculate fitness ,in , Crossover and mutation operations are performed using MARL agents and a knowledge graph rule engine, recording the changes in the fitness of the best individual in each generation. Intermediate data: the cycle time of the best individual in the 50th generation. seconds, variance Second.

[0081] S4. Fast Evaluator Task Assignment and Cycle Estimation: For the candidate solution set... An evaluation is conducted, tasks are allocated based on a greedy rule, and the cycle is estimated. Intermediate data: Evaluation result set For example, a certain solution Second.

[0082] S5. Spatiotemporal Conflict Prediction and Load Balancing Optimization: Simulate the placement head movement path and calculate the conflict probability. Optimize load balancing objectives Intermediate data: Conflict probability Gini coefficient .

[0083] S6. Solution Output and Application Verification: Output the optimal solution. Improved computational efficiency Intermediate data: Optimal solution cycle time. Seconds, improvement rate .

[0084] Intermediate data recorded during the experiment are as follows: Population fitness changes: Recording the fitness of the best individual during the iteration process of the genetic algorithm. and cycle time Sampling is performed every 10 generations.

[0085] Conflict probability and load balancing metrics: Record the conflict probability of each generation of solutions. Load variance and Gini coefficient .

[0086] Computational overhead: Records the computation time for each evaluation by the fast evaluator. Ensure that the requirements are met. .

[0087] The experiment was tested on simulated datasets (medium-scale scenarios) and real datasets. Evaluation metrics included total placement cycle time. Load variance Conflict probability and efficiency improvement rate The results are shown in the table below: The results, compared with other methods, are as follows: Traditional genetic algorithms (GA-traditional): only use random crossover and mutation, without introducing reinforcement learning and rule engines, and conflict prediction and load balancing rely on static rules.

[0088] The proposed method (GA-RL-RE) combines an improved genetic algorithm, MARL, a knowledge graph rule engine, and Gini coefficient load balancing optimization.

[0089] Comparative Analysis: As shown in the table above, the proposed method reduces the total cycle time by approximately 10.8%-11.4% compared to traditional GA, lowers the load variance by approximately 57%-63%, reduces the conflict probability by approximately 55%-60%, and improves efficiency by approximately 9.7%-9.9%. This indicates that MARL and the rule engine significantly enhance global search capabilities and conflict avoidance, while Gini coefficient optimization further improves load balancing.

[0090] To verify the effectiveness of each component, ablation experiments were conducted, removing the MARL-driven crossover mutation, knowledge graph rule engine, Gini coefficient load balancing, and spatiotemporal conflict prediction mechanism. The results are as follows: Ablation experiments show that MARL contributes the most to improving global search capabilities, with cycle time increasing by 8.7% after removal; spatiotemporal conflict prediction has a significant effect on reducing conflict probability, after removal... An increase of 137.5%; the Gini coefficient has some improvement on load balancing, after removing... An increase of 54.7%. Overall performance is optimal due to the synergistic effect of all components.

[0091] Experimental results show that the feeder placement optimization method proposed in this paper significantly outperforms traditional methods in terms of total placement cycle time, load balancing, and conflict reduction, especially demonstrating strong adaptability and robustness in large-scale scenarios. Efficiency improvement reaches 22.3%-22.5%, load variance is reduced by approximately 60%, and conflict probability is reduced by approximately 60%, validating the effectiveness of the GA-RL-RE hybrid algorithm and the spatiotemporal conflict prediction mechanism. Ablation experiments further reveal the contributions of each component, highlighting the crucial role of MARL and conflict prediction in performance improvement.

[0092] Within the technical scope disclosed in this invention, any variations or substitutions that can be easily conceived should be included within the protection scope of this invention. Therefore, the protection scope of this invention should be determined by the scope of the claims.

Claims

1. A method for optimizing feeder layout in a linear chip mounter, characterized in that, include: S1. Collect production line data and initialize optimization variables to generate an initial dataset; S2. Construct a model with the goal of minimizing the total period, coupling three decision variables to generate a set of decision variables; S3. Based on the genetic algorithm, a global search is performed to obtain a set of candidate solutions; S4. Evaluate the candidate solution set, assign tasks and estimate the cycle, and output the evaluation result set; S5. Based on the evaluation result set, predict conflicts and optimize the load, and generate an optimization solution set; S6. Output and verify the best solution based on the set of optimized solutions.

2. The feeder layout optimization method for a linear chip mounter according to claim 1, characterized in that, Step S1 includes: S11. Collect SMT production line parameters, including the number of feeders. Number of mounting heads PCB board assembly type Requirement of each component Use sensors to capture real-time data and generate an initial dataset. ,in For the first One flight location, For the first One mounting head; S12. Initialize the feeder station allocation matrix Assign vector to mounting head assembly , take the sequence planning sequence Calculate the initial total cycle time estimate S13. Use a visual language model to analyze the 3D production line scene and generate a semantic constraint output initialization constraint set. .

3. The feeder layout optimization method for a linear chip mounter according to claim 1 or 2, characterized in that, Step S2 includes: S21. Define the objective function: ,in For the head exist The time of the step, For the waiting time, Total cycle time; For the head Waiting, constraint is ; S22. Incorporate dynamic programming, drawing on pipeline layer allocation DP, and enumerate the feeder allocation positions. Calculate the time of subproblems Select the optimal Update the DP table; Cost of sub-segmentation; For execution time; For micro-batch size; For communication costs; For bubble factor; S23. Generate a set of decision variables Integrate the initialization constraint set As an additional constraint.

4. The feeder layout optimization method for a linear chip mounter according to claim 3, characterized in that, Step S3 includes: S31. Initialize the population (K is the population size), each yes Variants of the fitness function ,in ; S32. Combined with the rule engine, introduce PPO reinforcement learning agent selection operation to avoid random mutation; S32. Generate a set of candidate solutions Enter the next step.

5. The feeder layout optimization method for a linear chip mounter according to claim 4, characterized in that, Step S4 includes: S41. Task allocation based on greedy rules: For sequences Calculate each head The workload; S42, Estimation Period Embed conflict detection rules to avoid overlap; S43, Output Evaluation Result Set Enter the next step.

6. The feeder layout optimization method for a linear chip mounter according to claim 5, characterized in that, Step S5 includes: S51. Simulate motion path and predict collision probability. ; S52, Optimize Variance ,Adjustment Achieve equilibrium; S53. Generate an optimized solution set ,avoid .

7. The feeder layout optimization method for a linear chip mounter according to claim 6, characterized in that, Step S6 includes: S61, Select ; S62. Simulate and verify on the SMT production line to calculate the actual efficiency improvement. ; S63, Output the final layout scheme.

8. A feeder layout optimization system for an in-line pick-and-place machine, based on the feeder layout optimization method for an in-line pick-and-place machine according to any one of claims 1 to 7, characterized in that, include: The data optimization module collects production line data and initializes optimization variables to provide reliable input for subsequent steps. The decision planning module constructs a coupled model to quantify the optimization objective and generates a set of decision variables. The global search module performs a global search and generates a set of candidate solutions. The task allocation module evaluates the candidate solution set, allocates tasks and estimates the cycle, and outputs the evaluation result set. The load optimization module predicts conflicts and optimizes the load based on the evaluation result set, generating a set of optimization solutions. The verification output module outputs and verifies the best solution based on the set of optimized solutions. Its purpose is to ensure practical feasibility and achieve system-level optimization.

Images