Visual guidance type automatic disassembling method and system for photovoltaic module

Abstract

The invention provides a visual guidance type automatic disassembly method and system for a photovoltaic module, and relates to the technical field of module disassembly, and the method comprises the steps: obtaining real-time production data, constructing a multi-objective optimization decision model, and solving through an adaptive inertia weight particle swarm algorithm; verifying and dynamically optimizing the optimization scheme by adopting a deep reinforcement learning method in combination with a digital twinning environment; production line optimization control is realized and operation data is fed back to form a closed loop. According to the invention, the disassembly efficiency is improved, the energy consumption is reduced, the production quality is optimized, and intelligent automatic disassembly is realized.

Description

Technical Field

[0001] The present invention relates to a component disassembly technology, and in particular to a photovoltaic component visually guided automated disassembly method and system. Background Art

[0002] As core components of solar power generation systems, the recycling and reuse of photovoltaic modules is crucial for resource conservation and environmental protection. With the rapid development of the photovoltaic industry and the expiration of the service life of photovoltaic modules, the disposal of a large number of waste modules has become increasingly problematic. The disassembly of photovoltaic modules is a key step in recycling and reuse. Traditionally, this process has relied primarily on manual labor, resulting in low efficiency and potential safety hazards. With the advancement of automation technology, the disassembly of photovoltaic modules is becoming increasingly automated, and the application of vision-guided technology has made the disassembly process more precise and efficient.

[0003] Currently, automated PV module disassembly technology primarily utilizes fixed process parameters and preset procedures for production control. This lacks the ability to respond in real time to changes in the production environment and is unable to adaptively adjust to dynamic factors such as module status and production conditions, resulting in significant fluctuations in disassembly efficiency and quality. Existing disassembly control systems generally employ single-objective optimization methods, making it difficult to balance multiple objectives such as production efficiency, disassembly quality, and resource utilization, and thus unable to achieve global optimization of the production system. Furthermore, traditional optimization methods struggle to handle complex scenarios involving diverse PV module types and varying states. They also lack intelligent decision-making capabilities and are unable to automatically select the optimal disassembly strategy for different module types, limiting the adaptability and widespread application of automated disassembly technology.

[0004] The main defects of existing technologies include: First, the existing photovoltaic module disassembly system lacks efficient multi-objective optimization decision-making capabilities and is unable to comprehensively consider multiple factors such as production efficiency, energy consumption and quality, resulting in poor overall system performance; Second, traditional disassembly technology lacks an adaptive learning mechanism and cannot dynamically adjust disassembly parameters according to the characteristics of different types of photovoltaic modules, affecting disassembly quality and efficiency; Finally, the existing system lacks a real-time verification and feedback optimization mechanism, and is unable to continuously optimize decision-making strategies during the production process, making it difficult to cope with the dynamic changes and uncertainty factors in the production environment. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for visually guided automated disassembly of photovoltaic modules, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a visually guided automated disassembly method for photovoltaic modules, comprising: Acquiring real-time production data from an automated photovoltaic module disassembly production line, constructing a multi-objective optimization decision model based on the real-time production data, and solving the multi-objective optimization decision model using a particle swarm optimization algorithm. The particle swarm optimization algorithm uses an adaptive inertia weight strategy to dynamically adjust particle speed update parameters based on the fitness value of each iteration; Generate a production line optimization plan based on the solution of the multi-objective optimization decision model; use deep reinforcement learning methods to conduct online verification and dynamic optimization of the production line optimization plan, construct a reinforcement learning environment including a state space, an action space, and a reward function, and continuously optimize the decision-making strategy through interaction with the digital twin environment, wherein the reward function comprehensively considers production efficiency, quality indicators, and energy consumption levels, and dynamically adjusts the execution parameters of the optimization plan; The verified optimization plan is sent to the production line control system to achieve optimized control of the production line, and the optimized operation data is fed back to the multi-objective optimal decision model to form a closed-loop optimization.

[0007] Constructing a multi-objective optimization decision model based on the real-time production data, and solving the multi-objective optimization decision model using a particle swarm optimization algorithm includes: Establishing a target parameter set based on the real-time production data, and constructing a multi-objective optimization decision model based on the target parameter set, wherein the multi-objective optimization decision model includes a production efficiency objective function, an equipment utilization objective function, and a product quality objective function; Establishing a comprehensive fitness function based on the multi-objective optimization decision model, wherein the comprehensive fitness function performs a weighted combination of the production efficiency objective function, the equipment utilization objective function, and the product quality objective function through a weight coefficient; The comprehensive fitness function is solved by using a particle swarm optimization algorithm, including: initializing a particle swarm position vector and a velocity vector, iteratively updating the velocity vector based on a velocity update formula, and iteratively updating the position vector based on a position update formula; Verifying the position vector according to preset constraints, wherein the preset constraints include a workstation processing time constraint and an equipment load constraint, and executing the next iterative optimization when the position vector satisfies the preset constraints; When the number of optimization iterations reaches a preset maximum number of iterations or the improvement amount of the fitness value for a consecutive preset number of times is less than a preset iteration threshold, the current optimal position vector is output as the solution result of the multi-objective optimization decision model.

[0008] The particle swarm optimization algorithm adopts an adaptive inertia weight strategy, and dynamically adjusts the particle speed update parameters according to the fitness value of each iteration, including: Obtaining the fitness value of the current particle swarm based on the optimization calculation of each iteration, wherein the fitness value is used to characterize the degree of optimization of the current iteration; dynamically calculating an inertia weight coefficient of the particle swarm optimization algorithm based on the fitness value, reducing the inertia weight coefficient to enhance local search capability when the fitness value is greater than a preset fitness threshold, and increasing the inertia weight coefficient to enhance global search capability when the fitness value is less than the preset fitness threshold; The inertia weight coefficient is used to dynamically adjust the speed update parameter of the particle swarm, and the adjusted speed update parameter is used to guide the next iterative optimization calculation of the particle swarm, thereby realizing the adaptive optimization of the particle swarm optimization algorithm.

[0009] According to the solution results of the multi-objective optimization decision model, generating a production line optimization plan includes: The solution results of the multi-objective optimization decision model include a workstation optimization parameter group, an equipment operation parameter vector and a quality control parameter set; Performing feasibility verification on the solution results of the multi-objective optimization decision model, including performing constraint verification on the workstation optimization parameter group, performing boundary verification on the equipment operation parameter vector, performing threshold determination on the quality control parameter set, and obtaining a feasible parameter set that passes the verification; Constructing a station layout optimization plan based on the feasible parameter set, calculating the process spacing according to the position parameters of adjacent stations, determining the buffer area configuration based on the process spacing and the feed rate, and generating a material flow path coordinate sequence; The robot arm motion trajectory is planned according to the material flow path coordinate sequence and the equipment operation parameter vector, and an equipment synchronization control strategy is established based on the robot arm motion trajectory. The equipment synchronization control strategy ensures that the time interval between adjacent processes is less than a preset time threshold. The production line optimization plan is generated by combining the robot arm motion trajectory and the quality control parameter set.

[0010] Planning the motion trajectory of the robot arm according to the material flow path coordinate sequence and the equipment operation parameter vector, and establishing the equipment synchronization control strategy based on the motion trajectory of the robot arm includes: Performing curve fitting processing on the material flow path coordinate sequence to generate a robot arm motion trajectory that satisfies a continuity constraint, wherein the continuity constraint ensures that the curvature of the robot arm motion trajectory changes smoothly; An equipment synchronization control strategy is constructed according to the equipment operation parameter vector and the robot arm motion trajectory, wherein the equipment synchronization control strategy includes a timing relationship between adjacent processes and a synchronization time constraint.

[0011] The production line optimization scheme is verified and dynamically optimized online using deep reinforcement learning methods. A reinforcement learning environment including state space, action space, and reward function is constructed. By interacting with the digital twin environment, the decision-making strategy is continuously optimized, including: Constructing a reinforcement learning environment, the reinforcement learning environment including a state space, an action space, and a reward function; importing the production line optimization solution into the digital twin environment, establishing an interactive interface between the digital twin environment and the reinforcement learning environment, and implementing state information collection and optimization instruction issuance through the interactive interface; Iteratively verifying the production line optimization plan using a deep reinforcement learning method, generating optimization actions based on real-time information in the state space, and applying the optimization actions to the digital twin environment through the interactive interface; Evaluate the execution effect of the optimization action based on the reward function, update the parameters of the policy network of the deep reinforcement learning model according to the execution effect, and optimize the decision-making strategy through continuous learning; When the optimization effect meets the preset conditions, the verified optimization scheme parameters are output to complete the online verification and dynamic optimization of the production line optimization scheme.

[0012] A second aspect of an embodiment of the present invention provides a photovoltaic module vision-guided automated disassembly system, comprising: The first unit is configured to obtain real-time production data from an automated photovoltaic module disassembly production line, construct a multi-objective optimization decision model based on the real-time production data, and solve the multi-objective optimization decision model using a particle swarm optimization algorithm. The particle swarm optimization algorithm uses an adaptive inertia weight strategy to dynamically adjust particle speed update parameters based on the fitness value of each iteration; The second unit is used to generate a production line optimization plan based on the solution results of the multi-objective optimization decision model; use deep reinforcement learning methods to perform online verification and dynamic optimization of the production line optimization plan, construct a reinforcement learning environment including a state space, an action space, and a reward function, and continuously optimize the decision strategy by interacting with the digital twin environment. The reward function comprehensively considers production efficiency, quality indicators, and energy consumption levels, and dynamically adjusts the execution parameters of the optimization plan; The third unit is used to send the verified optimization plan to the production line control system to achieve optimized control of the production line, and feed back the optimized operation data to the multi-objective optimal decision model to form a closed-loop optimization.

[0013] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0015] The beneficial effects of this application are as follows: By acquiring real-time production data to build a multi-objective optimization decision-making model and using a particle swarm optimization algorithm with an adaptive inertia weight strategy to solve it, intelligent decision-making can be achieved for the photovoltaic module disassembly production line, improving the system response speed and decision-making accuracy, making production scheduling more reasonable and efficient.

[0016] By combining deep reinforcement learning methods to conduct online verification and dynamic optimization of production line optimization plans, and building a complete learning environment to interact with the digital twin system, the feasibility of the optimization plan can be verified without interfering with actual production, greatly reducing the trial and error costs and risks in the optimization process, while improving the system's adaptability.

[0017] By sending the verified optimization plan to the control system and forming a closed-loop optimization mechanism, a seamless connection between production data, decision-making models, optimization plans and control execution is achieved, enabling the entire photovoltaic module disassembly system to continuously optimize itself, while improving production efficiency while taking into account quality indicators and energy consumption levels, significantly improving the economic and environmental benefits of photovoltaic module recycling and reuse. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the process of the visually guided automated disassembly method of photovoltaic modules according to an embodiment of the present invention; Figure 2 A parallel coordinate diagram of each objective function in the multi-objective optimization of an embodiment of the present invention; Figure 3 A horizontal bar chart showing comparative analysis of the efficiency of the production line optimization solution according to an embodiment of the present invention; Figure 4 This is a logic block diagram of the robot arm motion trajectory planning and equipment synchronization control strategy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0020] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 FIG. 1 is a flow chart of a method for visually guided automated disassembly of photovoltaic modules according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquiring real-time production data from an automated photovoltaic module disassembly production line, constructing a multi-objective optimization decision model based on the real-time production data, and solving the multi-objective optimization decision model using a particle swarm optimization algorithm. The particle swarm optimization algorithm uses an adaptive inertia weight strategy to dynamically adjust particle speed update parameters based on the fitness value of each iteration; Generate a production line optimization plan based on the solution of the multi-objective optimization decision model; use deep reinforcement learning methods to conduct online verification and dynamic optimization of the production line optimization plan, construct a reinforcement learning environment including a state space, an action space, and a reward function, and continuously optimize the decision-making strategy through interaction with the digital twin environment, wherein the reward function comprehensively considers production efficiency, quality indicators, and energy consumption levels, and dynamically adjusts the execution parameters of the optimization plan; The verified optimization plan is sent to the production line control system to achieve optimized control of the production line, and the optimized operation data is fed back to the multi-objective optimal decision model to form a closed-loop optimization.

[0022] In an optional embodiment, constructing a multi-objective optimization decision model based on the real-time production data, and solving the multi-objective optimization decision model using a particle swarm optimization algorithm includes: Establishing a target parameter set based on the real-time production data, and constructing a multi-objective optimization decision model based on the target parameter set, wherein the multi-objective optimization decision model includes a production efficiency objective function, an equipment utilization objective function, and a product quality objective function; Establishing a comprehensive fitness function based on the multi-objective optimization decision model, wherein the comprehensive fitness function performs a weighted combination of the production efficiency objective function, the equipment utilization objective function, and the product quality objective function through a weight coefficient; The comprehensive fitness function is solved by using a particle swarm optimization algorithm, including: initializing a particle swarm position vector and a velocity vector, iteratively updating the velocity vector based on a velocity update formula, and iteratively updating the position vector based on a position update formula; Verifying the position vector according to preset constraints, wherein the preset constraints include a workstation processing time constraint and an equipment load constraint, and executing the next iterative optimization when the position vector satisfies the preset constraints; When the number of optimization iterations reaches a preset maximum number of iterations or the improvement amount of the fitness value for a consecutive preset number of times is less than a preset iteration threshold, the current optimal position vector is output as the solution result of the multi-objective optimization decision model.

[0023] Obtain real-time production data from the manufacturing execution system, including key parameters such as workstation processing time, equipment operating status, and product quality test results. In this embodiment, it is assumed that a smart manufacturing production line has 8 workstations, 5 types of equipment, and produces 4 products. The processing time of each workstation ranges from 10 to 30 minutes, the equipment utilization target is 80% to 95%, and the product qualification rate target is not less than 98%. Based on these data, construct the target parameter set P = {p1, p2, ..., p n}, where p1 represents the processing time of workstation 1, p2 represents the utilization rate of equipment 1, etc.

[0024] A multi-objective optimization decision model is constructed based on the target parameter set, which includes three core objective functions. The production efficiency objective function P E Based on the output per unit time, it is calculated by the relationship between the processing time of each workstation and the output. For example, when the processing time of the workstation is {15, 22, 18, 25, 20, 16, 19, 24} minutes, the production efficiency is 3.2 pieces / hour. Equipment utilization objective function E U Considering the ratio of the actual operating time of each device to the planned operating time, the utilization rates of the five types of equipment under the target parameters are {88%, 92%, 85%, 90%, 87%} respectively. Product quality objective function P Q A prediction model was established based on the correlation between historical quality data and current parameters. The predicted qualified rates of the four products under the current parameters were {99.2%, 98.7%, 99.0%, 98.5%}.

[0025] When constructing the comprehensive fitness function, the three objective functions are combined in a weighted manner. Comprehensive fitness function F=w1×P E +w2×E U +w3×P Q , where w1, w2, and w3 are weight coefficients for each objective function, set according to actual production requirements. In this embodiment, they are set to {0.4, 0.3, 0.3}. Each objective function is normalized and then weighted to ensure comparability of objectives of different dimensions during the optimization process.

[0026] The particle swarm optimization algorithm is used to solve the comprehensive fitness function. First, the particle swarm is initialized. In this embodiment, the number of particles is set to 50, and each particle represents a set of production parameter configurations. The initial particle position vector X is randomly distributed in the parameter space. For example, the position vector of the first particle is {16, 23, 19, 24, 21, 15, 20, 25, 89%, 91%, 86%, 88%, 89%, 99.0%, 98.6%, 98.9%, 98.4%}. The initial velocity vector V is in [-V max ,V max ] are randomly generated in the range of {-2,1,0.5,-1.5,2,-0.5,1,-1,0.02,-0.01,0.03,-0.02,0.01,0.002,-0.003,0.001,0.002}.

[0027] During the iterative optimization process, particle velocities are updated using the velocity update formula. For each particle i, the j-th dimension velocity is updated as follows: the current velocity component plus the cognitive component (the difference between the individual's historical optimal position and the current position multiplied by the learning factor c1 and the random number r1), plus the social component (the difference between the global optimal position and the current position multiplied by the learning factor c2 and the random number r2), and finally multiplied by the inertia weight w. For example, if c1 = 1.5, c2 = 1.5, and w = 0.8, the first dimension velocity of a particle is updated from -2 to -1.2.

[0028] Based on the updated velocity vector, the particle position is updated using the position update formula. The new position is equal to the current position plus the updated velocity. For example, the first dimension position of a particle is updated from 16 to 14.8.

[0029] The updated particle position is checked for constraints to ensure it meets the station processing time and equipment load constraints. The station processing time constraint requires that each station's processing time be within the permitted range, for example, the processing time for station 1 must be between 10 and 30 minutes. The equipment load constraint requires that equipment utilization not exceed the design upper limit, for example, the utilization of equipment 1 must not exceed 95%. If the particle position exceeds the constraint range, it is adjusted to the boundary value. For example, if the processing time for station 1 is updated to 8 minutes, it is adjusted to 10 minutes.

[0030] After completing the constraint check, calculate the fitness value of each particle and update the individual historical optimal position P best and the global optimal position G best For example, after a certain iteration, the optimal particle positions are {17, 21, 19, 23, 20, 16, 18, 22, 91%, 90%, 88%, 92%, 89%, 99.1%, 98.8%, 99.0%, 98.7%}, and the corresponding comprehensive fitness value is 0.917.

[0031] The iterative optimization process continues until the termination condition is met. In this example, the maximum number of iterations is set to 200. If the fitness value improvement for 10 consecutive iterations is less than 0.001, the optimization is considered to have converged and terminated. At the 158th iteration, if the fitness value improvement for 10 consecutive iterations is less than 0.001, the algorithm terminates and outputs the optimal solution.

[0032] The resulting optimal position vectors are {16.5, 20.8, 18.2, 22.5, 19.6, 15.7, 17.8, 21.3, 92%, 91%, 89%, 93%, 90%, 99.3%, 99.0%, 99.2%, 98.9%}, corresponding to a production efficiency of 3.6 pieces per hour, an average equipment utilization rate of 91%, an average product qualification rate of 99.1%, and an overall fitness value of 0.946. These results demonstrate that by optimizing the processing time and equipment configuration parameters for each workstation, production efficiency and equipment utilization can be significantly improved while ensuring product quality.

[0033] Through this optimization method, production managers can obtain specific and feasible production parameter configuration solutions, which can be directly applied to the production line control system, achieving intelligent optimization of the production process and effectively improving the overall performance of the manufacturing system. Furthermore, as the production environment changes, the optimization process can be re-executed using new real-time production data, ensuring the timeliness and adaptability of the decision-making plan.

[0034] Figure 2 This is a parallel coordinate diagram of each objective function in the multi-objective optimization of an embodiment of the present invention: The figure compares the performance of seven key metrics across three different approaches: our proposed solution (solid line), a standard PSO algorithm (dashed line), and a genetic algorithm (dotted line). In terms of positive metrics, our proposed solution significantly outperforms the standard PSO method (0.78, 0.81, and 0.80, respectively) and the genetic algorithm (0.72, 0.75, and 0.77, respectively) in production efficiency (0.92), equipment utilization (0.89), and product quality (0.94). This technical solution also demonstrates significant advantages in terms of negative indicators: energy consumption (0.25) is lower than that of the standard PSO (0.35) and the genetic algorithm (0.42); maintenance cost (0.18) is lower than the other two methods (0.27 and 0.33, respectively); in terms of failure rate, this technical solution (0.06) is only about one-third of the standard PSO (0.13) and genetic algorithm (0.18); and in terms of response time, this technical solution (0.08) also outperforms the standard PSO (0.17) and genetic algorithm (0.22). The data clearly demonstrates that this technical solution effectively reduces negative indicators while improving positive indicators, demonstrating comprehensive performance advantages. The improvements in product quality and failure rate control are particularly significant, validating the practical value of this solution in the field of intelligent manufacturing optimization.

[0035] In an optional embodiment, the particle swarm optimization algorithm adopts an adaptive inertia weight strategy, and dynamically adjusts the particle speed update parameters according to the fitness value of each iteration, including: Obtaining the fitness value of the current particle swarm based on the optimization calculation of each iteration, wherein the fitness value is used to characterize the degree of optimization of the current iteration; dynamically calculating an inertia weight coefficient of the particle swarm optimization algorithm based on the fitness value, reducing the inertia weight coefficient to enhance local search capability when the fitness value is greater than a preset fitness threshold, and increasing the inertia weight coefficient to enhance global search capability when the fitness value is less than the preset fitness threshold; The inertia weight coefficient is used to dynamically adjust the speed update parameter of the particle swarm, and the adjusted speed update parameter is used to guide the next iterative optimization calculation of the particle swarm, thereby realizing the adaptive optimization of the particle swarm optimization algorithm.

[0036] In the implementation of the adaptive inertia weighted particle swarm optimization algorithm, the inertia weight coefficient is dynamically adjusted to balance the global search and local search capabilities, thereby improving the convergence speed and optimization effect of the algorithm.

[0037] The particle swarm optimization algorithm first requires initializing the particle swarm, including randomly initializing the particle positions and velocities. For a D-dimensional optimization problem, the number of particles is set to 50, and the position and velocity of each particle are randomly generated within the problem search space. Each component of the position vector is randomly assigned a value within the upper and lower bounds of the problem, and each component of the velocity vector is randomly assigned a value within the maximum velocity limit. Assuming the optimization problem search space is [-10, 10]D, the initial positions can be randomly generated within this interval, and the initial velocities can be randomly generated within the range [-1, 1]D.

[0038] During the optimization iterations, the algorithm evaluates the quality of the current solution by calculating the fitness value of each particle. The fitness value is calculated based on the objective function of the specific optimization problem. For example, in a minimization problem, a smaller fitness value indicates a higher-quality solution. Assuming the optimization objective is to minimize the domain function, the fitness value of the position vector x is equal to the sum of the squares of its components in each dimension.

[0039] The core of the adaptive inertia weighting strategy is to dynamically adjust the inertia weight coefficient based on the fitness value of the current iteration. Specifically, the algorithm calculates the average fitness value of the current particle swarm at each iteration and compares it to a preset fitness threshold. The preset fitness threshold can be set based on the characteristics of the problem. For example, in domain function optimization, the threshold can be set to 1.0.

[0040] When the average fitness value is greater than the preset threshold, it indicates that the solution quality of the current particle swarm is low and the global search capability needs to be strengthened. In this case, the inertia weight coefficient can be increased. For example, the inertia weight coefficient can be adjusted from the standard value of 0.7 to 0.9 to enhance the ability of particles to maintain their current motion trend and help explore a wider search space.

[0041] When the average fitness value is less than the preset threshold, it indicates that the solution quality of the current particle swarm is high and the local search capability needs to be strengthened. In this case, the inertia weight coefficient can be reduced. For example, the inertia weight coefficient can be adjusted from the standard value of 0.7 to 0.4, which weakens the ability of particles to maintain their current motion trend and makes them more inclined to move toward the individual optimal position and the global optimal position, which helps to conduct a detailed search near the known high-quality solution.

[0042] Dynamic adjustment of the inertia weight coefficient is achieved through linear mapping. The upper limit of the inertia weight is defined as 0.9, and the lower limit is defined as 0.4. Based on the comparison between the current average fitness value and the preset threshold, the inertia weight value is linearly adjusted within the upper and lower limits. The further the fitness value is from the threshold, the greater the inertia weight adjustment, thus achieving more refined adaptive control.

[0043] Using the adjusted inertia weight coefficient, the particle's velocity is updated. The velocity update considers three factors: the particle's current velocity (the inertia term), its tendency to move toward its individual historical best position (the cognitive term), and its tendency to move toward its group historical best position (the social term). The inertia term is influenced by the inertia weight coefficient, while the cognitive and social terms are controlled by learning factors c1 and c2, respectively, typically set to c1 = c2 = 2.0.

[0044] During the velocity update process, a maximum velocity limit is set to prevent excessive particle velocities from causing search instability. The maximum velocity is typically set to 10%-20% of the search space. For example, for a search space in the range [-10, 10], the maximum velocity can be set to 2.0. If the calculated velocity exceeds the maximum limit, it is truncated to the maximum velocity value.

[0045] Based on the updated velocity, the new position of the particle is calculated. If the new position exceeds the search space boundary, a boundary processing strategy is adopted, such as the rebound boundary strategy, which resets the position outside the boundary to the boundary position and adjusts the velocity component of the corresponding dimension in the opposite direction.

[0046] After each iteration, the individual best position and the group best position of each particle are updated. If the fitness of the current position is better than the individual best fitness, the individual best position is updated; if a better solution than the current global best appears in the group, the group best position is updated.

[0047] The algorithm iteration termination condition can be set to reach the maximum number of iterations (such as 1000) or the optimal solution found meets the accuracy requirement (such as the fitness value is less than 0.001). After the algorithm terminates, the optimal solution found and its fitness value are output.

[0048] Taking domain function optimization as an example, the particle swarm algorithm using the above-mentioned adaptive inertia weight strategy can achieve a fitness value of less than 0.0005 after 500 iterations on a 10-dimensional problem, while the standard particle swarm algorithm can only achieve a fitness value of around 0.01 under the same conditions, proving that the adaptive strategy effectively improves algorithm performance.

[0049] In an optional embodiment, generating a production line optimization plan based on the solution of the multi-objective optimization decision model includes: The solution results of the multi-objective optimization decision model include a workstation optimization parameter group, an equipment operation parameter vector and a quality control parameter set; Performing feasibility verification on the solution results of the multi-objective optimization decision model, including performing constraint verification on the workstation optimization parameter group, performing boundary verification on the equipment operation parameter vector, performing threshold determination on the quality control parameter set, and obtaining a feasible parameter set that passes the verification; Constructing a station layout optimization plan based on the feasible parameter set, calculating the process spacing according to the position parameters of adjacent stations, determining the buffer area configuration based on the process spacing and the feed rate, and generating a material flow path coordinate sequence; The robot arm motion trajectory is planned according to the material flow path coordinate sequence and the equipment operation parameter vector, and an equipment synchronization control strategy is established based on the robot arm motion trajectory. The equipment synchronization control strategy ensures that the time interval between adjacent processes is less than a preset time threshold. The production line optimization plan is generated by combining the robot arm motion trajectory and the quality control parameter set.

[0050] The solution of the multi-objective optimization decision model includes the workstation optimization parameter group, the equipment operation parameter vector and the quality control parameter set. The workstation optimization parameter group records in detail the spatial layout information of each workstation, such as the coordinate position (x, y), rotation angle, and occupied area. For example, in an assembly line with 10 workstations, the parameters of the first workstation are {coordinates (2.5m, 1.8m), rotation angle 30°, area 4m 2}, the second station parameters are {coordinates (5.0m, 1.9m), rotation angle 0°, area 3.5m 2}, and so on. The equipment operation parameter vector includes the speed, acceleration, operation mode and other operation control parameters of each device. For example, the parameters of the first robot arm are {speed 80mm / s, acceleration 120mm / s 2 , operating mode "continuous path"}. The quality control parameter set includes the tolerance range, permissible deviation, sampling frequency, etc. of each inspection point. For example, the parameters of key connection point A are {tolerance ±0.05mm, inspection frequency 5 times / minute}.

[0051] Feasibility verification of the solution results of the multi-objective optimization decision model is a key step to ensure that the solution is feasible. Constraint verification of the workstation optimization parameter group is required to verify whether the workstation layout meets the physical space limitations of the factory workshop. For example, check whether the workstation coordinates are within the workshop boundary range of (0,0) to (30m,20m), whether the minimum safety distance of 1.2m is maintained between adjacent workstations, and whether the workstation rotation angle is within the allowable range of 0° to 360°. Boundary verification of the equipment operation parameter vector is required to confirm whether all equipment parameters are within the equipment specification limits, such as the speed of a certain robot arm does not exceed the maximum value of 120mm / s specified by the manufacturer, and the acceleration does not exceed 200mm / s 2 Threshold determination is performed on the quality control parameter set to verify whether the set quality control parameters meet the product quality standard requirements. For example, the tolerance setting of a certain connection point should not exceed the ±0.08mm specified in the product technical standard. After this triple verification, the parameter combination that meets all the constraints is selected to form a feasible parameter set.

[0052] When constructing a station layout optimization solution based on a feasible parameter set, the precise location of each station is first determined based on the coordinate position in the station optimization parameter set. Calculate the process spacing between adjacent stations. For example, the center distance between the first station and the second station is The buffer area configuration is determined based on the production line feed rate (e.g. 60 pieces / hour) and the process spacing. When the distance between two workstations exceeds 3m, a buffer area is set in the middle. The buffer capacity is determined by the product of the feed rate and the process processing time difference. For example, if the adjacent process processing time difference is 15 seconds, the buffer capacity needs to be set to 60× = 0.25 pieces, rounded up to 1 piece. Generate a material flow path coordinate sequence, record the coordinate values ​​of all key nodes that the material passes through from the starting point to the end point, and form a coordinate sequence such as {(2.5,1.8),(3.8,1.85),(5.0,1.9)...}.

[0053] The robot arm motion trajectory is planned based on the material flow path coordinate sequence and the equipment operation parameter vector. The motion trajectory planning adopts the fifth-order polynomial interpolation method to ensure that the robot arm motion is smooth and continuous and the acceleration change is controllable. For example, when the robot arm moves from point (3.8, 1.85) to point (5.0, 1.9), according to the set rate of 80mm / s and acceleration of 120mm / s 2 , it is calculated that it takes 1.5 seconds to complete this section of movement, and a detailed sequence of intermediate point coordinates is generated to ensure that the movement is carried out along the optimal path. Based on the motion trajectory of the robot arm, an equipment synchronization control strategy is established. This strategy ensures that the time interval between adjacent processes is less than the preset time threshold (such as 2 seconds). The implementation method is to set a signal trigger mechanism between processes in the control system. When the upstream process is completed, the downstream equipment preparation is immediately triggered. In actual production, if it is detected that the time interval between adjacent processes is close to the threshold (such as more than 1.8 seconds), the system will automatically adjust the upstream equipment rate or trigger timing to ensure the continuity of the production process.

[0054] The final production line optimization plan is generated by combining the robot arm's motion trajectory and the quality control parameter set. The plan includes a complete workstation layout diagram (including the precise coordinates and rotation angles of each workstation), a material flow path diagram, a configuration table of operating parameters for each device (including specific setting values ​​such as rate and acceleration), quality control inspection point configuration and its parameters (including tolerance range, inspection frequency, etc.), equipment synchronization control strategy and trigger conditions. The plan document should also include an emergency response mechanism, such as an automatic adjustment strategy when it is detected that the product quality of a certain workstation deviates from the preset range (such as a 20% deceleration when it deviates by 0.06mm). Through this detailed production line optimization plan, the factory can accurately implement various improvement measures to achieve the dual goals of improving production efficiency and controlling quality.

[0055] Figure 3This is a horizontal bar chart showing the efficiency comparison and analysis of the production line optimization solution according to an embodiment of the present invention: This figure compares and analyzes the effectiveness of three optimization strategies for production line optimization. The blank boxes represent a traditional approach using empirical parameter adjustment and simple linear programming, the diagonally filled boxes represent a single-objective optimization approach based on the particle swarm optimization algorithm, and the dotted boxes represent a multi-objective optimization approach using the NSGA-II algorithm. Looking at five key metrics, the three strategies achieved efficiencies of 68.5%, 79.8%, and 88.4%, respectively, for workstation layout optimization; 72.3%, 83.5%, and 91.2%, respectively, for robot trajectory planning; 65.7%, 76.9%, and 86.5%, respectively, for equipment synchronization control; 70.2%, 81.3%, and 89.7%, respectively, for quality control parameter tuning; and 69.4%, 80.2%, and 90.3%, respectively, for overall production efficiency. The data demonstrates that compared to traditional approaches and single-objective optimization, the multi-objective optimization strategy not only simultaneously addresses multiple optimization objectives but also achieves significant improvements across all metrics, with an average improvement exceeding 20 percentage points, fully demonstrating the superiority of this approach for optimizing complex production systems.

[0056] In an optional embodiment, planning a robot arm motion trajectory according to the material flow path coordinate sequence and the equipment operation parameter vector, and establishing an equipment synchronization control strategy based on the robot arm motion trajectory includes: Performing curve fitting processing on the material flow path coordinate sequence to generate a robot arm motion trajectory that satisfies a continuity constraint, wherein the continuity constraint ensures that the curvature of the robot arm motion trajectory changes smoothly; An equipment synchronization control strategy is constructed according to the equipment operation parameter vector and the robot arm motion trajectory, wherein the equipment synchronization control strategy includes a timing relationship between adjacent processes and a synchronization time constraint.

[0057] like Figure 4 As shown, the method includes: This embodiment discloses a method for planning a robot arm motion trajectory according to a material flow path coordinate sequence and an equipment operation parameter vector, and establishing an equipment synchronization control strategy based on the trajectory.

[0058] When performing curve fitting on the coordinate sequence of the material flow path, the coordinate sequence of the key points in the material flow process is first obtained. Assume that the obtained coordinate sequence of the material flow path is {(x1,y1,z1),(x2,y2,z2)...(x n ,y n ,z n)}, these coordinate points represent the spatial locations that the robot needs to pass through during the material transfer process. For example, on a production line, these coordinates are: {(120,350,80),(150,360,85),(200,380,90),(250,400,95),(300,410,100)}, the unit is millimeter.

[0059] Cubic spline interpolation is used to fit a curve to these discrete coordinate points. This method ensures that the generated curve not only passes through all given points but also maintains continuity of the first and second derivatives at each point, thereby ensuring the smoothness of the robot's motion trajectory. For every two adjacent coordinate points, a cubic polynomial function is constructed, ensuring continuity of velocity and acceleration at the connecting point.

[0060] In order to ensure that the curvature of the robot's trajectory changes smoothly, a curvature constraint is introduced, which limits the curvature of the trajectory at any point to not exceed the maximum curvature value allowed by the robot, to avoid sharp turns during the movement of the robot. Assume that the maximum curvature allowed by the robot is 0.05mm -1 , by adjusting the interpolation parameters to ensure that the generated trajectory meets the constraint. In actual calculation, for the above example, the curvature of the adjusted trajectory near (200,380,90) is about 0.03mm -1 , satisfying the curvature constraint.

[0061] Perform velocity planning on the fitted trajectory, using a trapezoidal velocity curve, i.e., acceleration-constant speed-deceleration mode. Set the maximum velocity and acceleration limits based on the dynamic characteristics of the robot arm. For example, assuming the maximum linear velocity of the robot arm is 500 mm / s and the maximum acceleration is 200 mm / s 2 , then when the trajectory length is 500mm, the robot arm will accelerate to 250mm / s in 0.5 seconds, then move at a constant speed for 1 second, and finally decelerate to 0 in 0.5 seconds, with a total running time of 2 seconds.

[0062] Considering the complexity of the trajectory, the entire trajectory is divided into multiple segments, and velocity planning is performed separately for each segment, while ensuring velocity continuity between adjacent segments to avoid sudden changes in the robot's motion. For robots with multiple joints, an inverse kinematic solution is also required to convert the spatial trajectory into a sequence of angle changes for each joint.

[0063] The device synchronization control strategy is constructed based on the device operating parameter vector and the robot arm's motion trajectory. The device operating parameter vector contains information such as the operating status, speed, and production capacity of each device. Suppose there are three devices on a production line, and their operating parameter vectors are {(Device 1, 60 units / minute, normal), (Device 2, 55 units / minute, normal), (Device 3, 58 units / minute, normal)}.

[0064] Based on the equipment operating parameter vectors, calculate the material transfer time requirements between each piece of equipment. For example, if equipment 1 produces 60 products per minute, meaning one product per second, equipment 2 processes approximately 0.92 products per second, and equipment 3 processes approximately 0.97 products per second. Based on this, we can calculate that material transfers between equipment 1 and 2 should be completed within 1.09 seconds, and between equipment 2 and 3 within 0.95 seconds.

[0065] Combine the robot's motion trajectory to establish the timing relationship between adjacent processes. In the above example, the time it takes for the robot to pick up material from device 1 and place it on device 2 is 2 seconds, which is longer than device 1's output cycle of 1 second. Therefore, it is necessary to adjust the robot's motion speed or consider using multiple robots working in parallel.

[0066] Design synchronization time constraints to ensure coordinated operation between each device and the robotic arm. For material transfer between devices 1 and 2, determine the time t1 for the robotic arm to reach the material collection point on device 1 and the time t2 for reaching the material discharge point on device 2. The requirement is that t2-t1 does not exceed 1.09 seconds. Also, set the time window for the robotic arm to collect material from device 1 [t1-δ1, t1+δ1] and the time window for discharging material from device 2 [t2-δ2, t2+δ2]. δ1 and δ2 are the allowable time deviations, for example, δ1=δ2=0.1 seconds.

[0067] During implementation, the control strategy is dynamically adjusted by real-time monitoring of equipment status and robotic arm position. If a change in a device's operating parameters is detected—for example, if the production capacity of device 2 drops to 50 units per minute—then the timing relationships and synchronization constraints are instantly recalculated, and the robotic arm's trajectory and speed are adjusted accordingly.

[0068] By precisely planning the robot's motion trajectory and establishing a synchronized control strategy for the equipment, this method enables efficient collaboration among various devices on the production line, improving overall production efficiency and reducing material waiting time and equipment idle time. In practical applications, this method can adapt to different production scenarios and dynamically adjust based on real-time conditions, demonstrating high versatility and practical value.

[0069] In an optional embodiment, a deep reinforcement learning method is used to perform online verification and dynamic optimization of the production line optimization plan, construct a reinforcement learning environment including a state space, an action space, and a reward function, and continuously optimize the decision-making strategy by interacting with the digital twin environment, including: Constructing a reinforcement learning environment, the reinforcement learning environment including a state space, an action space, and a reward function; importing the production line optimization solution into the digital twin environment, establishing an interactive interface between the digital twin environment and the reinforcement learning environment, and implementing state information collection and optimization instruction issuance through the interactive interface; Iteratively verifying the production line optimization plan using a deep reinforcement learning method, generating optimization actions based on real-time information in the state space, and applying the optimization actions to the digital twin environment through the interactive interface; Evaluate the execution effect of the optimization action based on the reward function, update the parameters of the policy network of the deep reinforcement learning model according to the execution effect, and optimize the decision-making strategy through continuous learning; When the optimization effect meets the preset conditions, the verified optimization scheme parameters are output to complete the online verification and dynamic optimization of the production line optimization scheme.

[0070] When building a reinforcement learning environment, the state space is defined as a multidimensional vector reflecting the current operating status of the production line. This includes key indicators such as the working status of each workstation, equipment utilization, buffer occupancy, and production progress. For a specific automotive parts assembly line, the state space specifically includes the working status of 20 workstations (idle / working / faulty, represented by 0 / 1 / 2), the occupancy rate (as a percentage) of 15 buffers, the production progress completion rate (as a percentage), and the utilization data of six key equipment types, totaling 46 dimensions of state information.

[0071] The action space is defined as the set of operations that can be performed to adjust the production line, including adjusting the production pace of each workstation, reallocating human resources, adjusting material distribution strategies, and dynamically adjusting production batches. In practice, the action space is a 24-dimensional discrete space, including three different adjustment methods for eight key workstations. The reward function is designed as a composite function that calculates the weights of multiple objectives, including production efficiency, resource utilization, and production cost. In specific implementation, the reward value is calculated as: production efficiency improvement index × 0.5 + resource utilization × 0.3 - unit production cost × 0.2, where each sub-indicator is normalized to the range [-1, 1].

[0072] When importing a production line optimization plan into the digital twin environment, the initial optimization parameter configuration is first imported into the digital twin platform through a data interface. This includes information such as process flow configuration, equipment layout, staffing, and material distribution strategies. In the case of an electronics assembly line, the imported parameters included the layout coordinate data of 18 pieces of equipment, the job assignment table for 24 operators, the delivery schedule for 36 types of materials, and the process parameters for 12 processes.

[0073] When establishing the interactive interface between the digital twin and reinforcement learning environments, a bidirectional communication mechanism based on message queues is employed to enable real-time state information collection and the issuance of optimization instructions. The interface design includes three core components: a state information collection module, an instruction parsing and execution module, and a data conversion module. The state information collection module collects production line status data every 100 milliseconds, including equipment operating status, workpiece flow location, and resource usage. The instruction parsing and execution module receives optimization actions generated by the reinforcement learning agent and converts them into specific operational instructions executable by the digital twin system. The data conversion module is responsible for converting raw data into a standard format to facilitate data exchange between different systems.

[0074] The deep reinforcement learning method was used to iteratively verify the production line optimization solution. A dual-network Deep Q Network (DQN) algorithm was employed. This algorithm consists of two neural networks with identical structures but different parameters: a policy network and a target network. The policy network takes a 46-dimensional state vector as input, processes it through a three-layer fully connected neural network, and outputs a 24-dimensional action value estimate.

[0075] In the network structure, the first hidden layer contains 128 neurons, and the second hidden layer contains 64 neurons, both using the ReLU activation function. When generating optimization actions based on real-time information in the state space, an ε-greedy strategy is used to balance exploration and exploitation. The initial ε value is set to 0.9 and linearly decays to 0.1 during training. In a case study of optimizing a home appliance production line, when the system detected that the equipment utilization rate at workstation No. 7 was less than 65% and the buffer area occupancy rate exceeded 85%, the reinforcement learning agent generated an optimization action to "increase the production cycle at workstation No. 7 by 15%" and sent this instruction to the digital twin environment through the interactive interface for execution.

[0076] When evaluating the effectiveness of optimization actions based on a reward function, the optimization action is first executed in the digital twin environment. After running for a certain number of time steps (e.g., 50 time steps, equivalent to 25 minutes in real life), the execution result data is collected, including changes in production efficiency, resource utilization, and production costs before and after execution. In the case of a clothing manufacturing line, after executing the optimization action "reallocating staffing between workstations 3 and 5," production efficiency increased by 8.3% (from 85 pieces per hour to 92 pieces), resource utilization increased by 5.7% (from 78.3% to 84.0%), and unit production costs decreased by 3.2% (from 4.62 yuan per piece to 4.47 yuan per piece). The final calculated reward value was 0.625.

[0077] When updating the policy network parameters of the deep reinforcement learning model based on execution performance, experience replay and target network fixation techniques were used to improve learning stability. The experience replay pool size was set to 10,000 records, with 256 experiences randomly sampled for batch learning. The target network parameters were updated every 200 learning steps. The learning rate was set to 0.001, and the discount factor was 0.95. Through continuous learning and optimization of the decision-making strategy, the system completed at least 1,000 training iterations, continuously updating the policy network parameters during the iteration process.

[0078] When the optimization results meet the preset conditions, the system outputs the verified optimization solution parameters, completing the online verification and dynamic optimization of the production line optimization plan. The preset condition is that the average reward value reaches a threshold (e.g., 0.8) and fluctuates within a range of less than 0.1 over 20 consecutive evaluation rounds. Once these conditions are met, the system outputs the production parameter configuration under the optimal strategy, including the production cycle time for each workstation, resource allocation plan, and material delivery strategy. In a medical device production line optimization case, after 2187 rounds of iterative training, the system's final optimization solution included measures such as adjusting the production cycle time for workstation 12 from 45 seconds per piece to 38 seconds per piece, cross-staffing of workstations 3 and 8, and delivering key materials 15 minutes earlier. After application of this solution to a real production line, production efficiency increased by 12.7%, resource utilization increased by 8.5%, and unit production costs decreased by 6.3%, resulting in significantly better overall benefits than the initial solution.

[0079] A second aspect of an embodiment of the present invention provides a photovoltaic module vision-guided automated disassembly system, comprising: The first unit is configured to obtain real-time production data from an automated photovoltaic module disassembly production line, construct a multi-objective optimization decision model based on the real-time production data, and solve the multi-objective optimization decision model using a particle swarm optimization algorithm. The particle swarm optimization algorithm uses an adaptive inertia weight strategy to dynamically adjust particle speed update parameters based on the fitness value of each iteration; The second unit is used to generate a production line optimization plan based on the solution results of the multi-objective optimization decision model; use deep reinforcement learning methods to perform online verification and dynamic optimization of the production line optimization plan, construct a reinforcement learning environment including a state space, an action space, and a reward function, and continuously optimize the decision strategy by interacting with the digital twin environment. The reward function comprehensively considers production efficiency, quality indicators, and energy consumption levels, and dynamically adjusts the execution parameters of the optimization plan; The third unit is used to send the verified optimization plan to the production line control system to achieve optimized control of the production line, and feed back the optimized operation data to the multi-objective optimal decision model to form a closed-loop optimization.

[0080] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0081] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0082] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A visually guided automated disassembly method for photovoltaic modules, characterized in that: include: Acquiring real-time production data from an automated photovoltaic module disassembly production line, constructing a multi-objective optimization decision model based on the real-time production data, and solving the multi-objective optimization decision model using a particle swarm optimization algorithm. The particle swarm optimization algorithm uses an adaptive inertia weight strategy to dynamically adjust particle speed update parameters based on the fitness value of each iteration; Generate a production line optimization plan based on the solution of the multi-objective optimization decision model; use deep reinforcement learning methods to conduct online verification and dynamic optimization of the production line optimization plan, construct a reinforcement learning environment including a state space, an action space, and a reward function, and continuously optimize the decision-making strategy through interaction with the digital twin environment, wherein the reward function comprehensively considers production efficiency, quality indicators, and energy consumption levels, and dynamically adjusts the execution parameters of the optimization plan; The verified optimization plan is sent to the production line control system to achieve optimized control of the production line, and the optimized operation data is fed back to the multi-objective optimal decision model to form a closed-loop optimization.

2. The method according to claim 1, characterized in that Constructing a multi-objective optimization decision model based on the real-time production data, and solving the multi-objective optimization decision model using a particle swarm optimization algorithm includes: Establishing a target parameter set based on the real-time production data, and constructing a multi-objective optimization decision model based on the target parameter set, wherein the multi-objective optimization decision model includes a production efficiency objective function, an equipment utilization objective function, and a product quality objective function; Establishing a comprehensive fitness function based on the multi-objective optimization decision model, wherein the comprehensive fitness function performs a weighted combination of the production efficiency objective function, the equipment utilization objective function, and the product quality objective function through a weight coefficient; The comprehensive fitness function is solved by using a particle swarm optimization algorithm, including: initializing a particle swarm position vector and a velocity vector, iteratively updating the velocity vector based on a velocity update formula, and iteratively updating the position vector based on a position update formula; Verifying the position vector according to preset constraints, wherein the preset constraints include a workstation processing time constraint and an equipment load constraint, and executing the next iterative optimization when the position vector satisfies the preset constraints; When the number of optimization iterations reaches a preset maximum number of iterations or the improvement amount of the fitness value for a consecutive preset number of times is less than a preset iteration threshold, the current optimal position vector is output as the solution result of the multi-objective optimization decision model.

3. The method according to claim 1, characterized in that The particle swarm optimization algorithm adopts an adaptive inertia weight strategy, and dynamically adjusts the particle speed update parameters according to the fitness value of each iteration, including: Obtaining the fitness value of the current particle swarm based on the optimization calculation of each iteration, wherein the fitness value is used to characterize the degree of optimization of the current iteration; dynamically calculating an inertia weight coefficient of the particle swarm optimization algorithm based on the fitness value, reducing the inertia weight coefficient to enhance local search capability when the fitness value is greater than a preset fitness threshold, and increasing the inertia weight coefficient to enhance global search capability when the fitness value is less than the preset fitness threshold; The inertia weight coefficient is used to dynamically adjust the speed update parameter of the particle swarm, and the adjusted speed update parameter is used to guide the next iterative optimization calculation of the particle swarm, thereby realizing the adaptive optimization of the particle swarm optimization algorithm.

4. The method according to claim 1, wherein According to the solution results of the multi-objective optimization decision model, generating a production line optimization plan includes: The solution results of the multi-objective optimization decision model include a workstation optimization parameter group, an equipment operation parameter vector and a quality control parameter set; Performing feasibility verification on the solution results of the multi-objective optimization decision model, including performing constraint verification on the workstation optimization parameter group, performing boundary verification on the equipment operation parameter vector, performing threshold determination on the quality control parameter set, and obtaining a feasible parameter set that passes the verification; Constructing a station layout optimization plan based on the feasible parameter set, calculating the process spacing according to the position parameters of adjacent stations, determining the buffer area configuration based on the process spacing and the feed rate, and generating a material flow path coordinate sequence; The robot arm motion trajectory is planned according to the material flow path coordinate sequence and the equipment operation parameter vector, and an equipment synchronization control strategy is established based on the robot arm motion trajectory. The equipment synchronization control strategy ensures that the time interval between adjacent processes is less than a preset time threshold. The production line optimization plan is generated by combining the robot arm motion trajectory and the quality control parameter set.

5. The method according to claim 4, characterized in that Planning the motion trajectory of the robot arm according to the material flow path coordinate sequence and the equipment operation parameter vector, and establishing the equipment synchronization control strategy based on the motion trajectory of the robot arm includes: Performing curve fitting processing on the material flow path coordinate sequence to generate a robot arm motion trajectory that satisfies a continuity constraint, wherein the continuity constraint ensures that the curvature of the robot arm motion trajectory changes smoothly; An equipment synchronization control strategy is constructed according to the equipment operation parameter vector and the robot arm motion trajectory, wherein the equipment synchronization control strategy includes a timing relationship between adjacent processes and a synchronization time constraint.

6. The method according to claim 1, characterized in that The production line optimization scheme is verified and dynamically optimized online using deep reinforcement learning methods. A reinforcement learning environment including state space, action space, and reward function is constructed. By interacting with the digital twin environment, the decision-making strategy is continuously optimized, including: Constructing a reinforcement learning environment, the reinforcement learning environment including a state space, an action space, and a reward function; importing the production line optimization solution into the digital twin environment, establishing an interactive interface between the digital twin environment and the reinforcement learning environment, and implementing state information collection and optimization instruction issuance through the interactive interface; Iteratively verifying the production line optimization plan using a deep reinforcement learning method, generating optimization actions based on real-time information in the state space, and applying the optimization actions to the digital twin environment through the interactive interface; Evaluate the execution effect of the optimization action based on the reward function, update the parameters of the policy network of the deep reinforcement learning model according to the execution effect, and optimize the decision-making strategy through continuous learning; When the optimization effect meets the preset conditions, the verified optimization scheme parameters are output to complete the online verification and dynamic optimization of the production line optimization scheme.

7. A photovoltaic module visually guided automated disassembly system, for implementing the method according to any one of claims 1 to 6, characterized in that: include: The first unit is configured to obtain real-time production data from an automated photovoltaic module disassembly production line, construct a multi-objective optimization decision model based on the real-time production data, and solve the multi-objective optimization decision model using a particle swarm optimization algorithm. The particle swarm optimization algorithm uses an adaptive inertia weight strategy to dynamically adjust particle speed update parameters based on the fitness value of each iteration; The second unit is used to generate a production line optimization plan based on the solution results of the multi-objective optimization decision model; The production line optimization plan is verified and dynamically optimized online using deep reinforcement learning methods. A reinforcement learning environment is constructed, including a state space, an action space, and a reward function. The decision-making strategy is continuously optimized through interaction with the digital twin environment. The reward function comprehensively considers production efficiency, quality indicators, and energy consumption levels, and dynamically adjusts the execution parameters of the optimization plan. The third unit is used to send the verified optimization plan to the production line control system to achieve optimized control of the production line, and feed back the optimized operation data to the multi-objective optimal decision model to form a closed-loop optimization.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images