Industrial flexible assembly process planning method and system based on artificial intelligence

Abstract

The invention relates to the technical field of industrial production and manufacturing, and discloses an industrial flexible assembly process planning method and system based on artificial intelligence. According to the method, multi-source heterogeneous data of an assembly line is collected in real time, and a structured input vector is generated through multi-modal data fusion processing; constructing a dynamic assembly path planning model based on deep reinforcement learning; establishing a multi-objective optimization function, generating a Pareto frontier solution set in combination with a constraint condition, and performing optimization by using an improved genetic algorithm; and constructing a virtual assembly simulation environment verification scheme and adjusting the model by means of a digital twinning technology. In addition, federal learning is adopted to realize multi-production-line collaborative optimization, a knowledge graph is constructed to assist decision making, and an online learning mechanism is applied to update model parameters. The assembly period can be effectively shortened, the equipment energy consumption and the material waste rate are reduced, the intelligent level and the production efficiency of industrial assembly are improved, and the flexible production requirement of the modern industry is met.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial production and manufacturing, and in particular to an industrial flexible assembly process planning method and system based on artificial intelligence. Background Art

[0002] In today's industrial production field, the assembly link is a key stage in product manufacturing, and the rationality of its process planning directly affects production efficiency, cost control and product quality. With the increasing diversification and personalization of market demand, industrial production is moving towards flexibility and intelligence, and traditional assembly process planning methods are gradually exposing many limitations.

[0003] Traditional assembly process planning is often based on fixed production models and experience, and lacks effective use of real-time production data. In actual production, factors such as material properties, equipment operating status, environmental parameters, and order requirements are constantly changing. For example, material quality fluctuations, sudden equipment failures, the impact of ambient temperature and humidity on assembly accuracy, and customers' temporary changes in order priorities frequently occur. Traditional methods are difficult to respond to these changes quickly, resulting in low assembly efficiency, frequent adjustments to production plans, and increased production costs.

[0004] At the same time, in terms of multi-objective optimization, it is difficult for traditional methods to take into account multiple key indicators such as assembly cycle, equipment energy consumption and material waste rate. In order to shorten the assembly cycle, the equipment may be overused, resulting in a significant increase in energy consumption; or in order to reduce material waste, a more conservative assembly strategy is adopted, which prolongs the assembly cycle. This situation of losing one thing while gaining another makes it difficult to achieve optimal production efficiency.

[0005] Traditional planning methods are also unable to cope with complex constraints. Factors such as workstation capacity limitations and irreversibility of process sequences pose many challenges to assembly process planning. Moreover, when there is a need to change orders in real time, traditional methods cannot optimize and adjust the assembly process in a timely manner, which can easily cause production delays and waste of resources.

[0006] In addition, as the scale of industrial production continues to expand, companies often have multiple assembly lines. However, most existing planning methods are aimed at a single assembly line and lack the ability to coordinate and optimize multiple production lines, making it impossible to achieve efficient operation of the entire production system. In addition, the management and utilization of historical process data and empirical knowledge are not sufficient, and it is impossible to provide effective decision support for assembly process planning. Summary of the invention

[0007] The purpose of the present invention is to provide an industrial flexible assembly process planning method and system based on artificial intelligence to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions: An industrial flexible assembly process planning method based on artificial intelligence, the method comprising:

[0009] Step S1: Real-time collect multi-source heterogeneous data of the assembly line, including material attributes, equipment operating status, environmental parameters, and historical process data, and perform feature extraction and normalization processing through a multi-modal data fusion module to generate a structured input vector;

[0010] Step S2: Build a dynamic assembly path planning model based on deep reinforcement learning, learn the long-term benefits of different assembly sequences and resource allocation strategies through the Q network, define the state space as the workstation load, material supply status, and order priority, and the action as the assembly path selection and equipment parameter adjustment;

[0011] Step S3: Establish a multi-objective optimization function, with minimizing the assembly cycle time, equipment energy consumption, and material waste rate as the optimization objectives, and generate a Pareto front solution set in combination with constraint conditions, where the constraint conditions include workstation capacity limitations, process sequence irreversibility, and real-time order change requirements;

[0012] Step S4: Use an improved genetic algorithm to iteratively optimize the Pareto front solution set, introduce an adaptive crossover probability and mutation operator, and screen the optimal assembly process plan through an elite retention strategy;

[0013] Step S5: Build a virtual assembly simulation environment based on digital twin technology, verify the feasibility of the optimization plan in real time, and dynamically adjust the reward function weight of the Q network according to the simulation results.

[0014] Preferably, the step S2 further includes:

[0015] Step S21: Define the state space vector as S = (L i , M j , P k ), where L i represents the real-time load rate of the i-th workstation, M j represents the inventory balance of the j-th type of material, and P k is the priority coefficient of the k-th order;

[0016] Step S22: Build a double deep Q network architecture, including a main network and a target network. The main network outputs the action value function Q(s, a; θ), where s is the current state, a is the action, and θ is the network parameter. The target network is used to calculate the target value Q'(s', a'; θ'), where s' is the next state, a' is the next action, and θ' is the target network parameter;

[0017] Step S23: Design the reward function R = α·T reduce +β·Esave -γ·W waste , where T reduce is the reduction amount of the assembly cycle, E save is the energy consumption reduction rate, W waste is the material waste rate, and α, β, and γ are dynamic weight coefficients, and satisfy α + β + γ = 1.

[0018] Preferably, in the step S3, the multi-objective optimization function is defined as:

[0019] minF(x) = [f1(x), f2(x), f3(x)]

[0020] where where T i (x) represents the time of the i-th assembly link, where E j (x) represents the energy consumption of the j-th device, where W k (x) represents the waste rate of the k-th type of material;

[0021] The constraint conditions include: g1(x) = C max -C i (x) ≥ 0, where C max is the maximum capacity of the workstation, and C i (x) is the actual load; g2(x) = O(x) - O min ≥ 0, where O(x) is the actual order completion rate, and O min is the minimum requirement.

[0022] Preferably, the operations of the improved genetic algorithm in the step S4 include:

[0023] Step S41: The chromosome coding uses an integer sequence to represent the assembly path order, and the gene positions correspond to the workstation numbers and material distribution identifiers;

[0024] Step S42: The adaptive crossover probability P c is dynamically adjusted according to the population diversity, and the calculation formula is P c = 0.8 - 0.3·(G current / G max ), where G current is the current iteration number, and G max is the maximum iteration number;

[0025] Step S43: The mutation operator adopts a Gaussian perturbation strategy, and applies a normal distribution random offset with a mean of 0 and a variance decreasing with the iteration number to the selected gene positions, where the variance σ 2 = 1 / (G current + 1).

[0026] Preferably, step S5 further includes:

[0027] Step S51: Construct a high-fidelity digital twin model based on the physical assembly line, and synchronize real-time data to drive the dynamic evolution of the virtual environment;

[0028] Step S52: Inject random perturbation events into the simulation environment, including equipment failure simulation and material supply delay, to test the robustness of the optimization scheme;

[0029] Step S53: According to the conflict detection information in the simulation results, reversely correct the exploration strategy of the Q network and the fitness function of the genetic algorithm.

[0030] Preferably, it further includes:

[0031] Step S6: Implement multi-production line collaborative optimization using a federated learning framework. Each local model updates the global model parameters through encrypted gradient aggregation while preserving data privacy;

[0032] Step S7: Define a knowledge graph to store historical optimization schemes and process rules, and mine implicit constraint relationships through a graph neural network to assist the decision-making and reasoning of the Q network.

[0033] Preferably, the construction of the knowledge graph in step S7 includes:

[0034] Step S71: The entity nodes are divided into three categories: equipment, materials, and processes, and the relationship edges are defined as "dependency", "conflict", and "substitution";

[0035] Step S72: Use the TransE algorithm for graph embedding to minimize the loss function L of the triple (h, r, t) = ‖h + r - t‖ 2 , where h is the head entity embedding vector, r is the relationship vector, and t is the tail entity embedding vector;

[0036] Step S73: Based on the message passing mechanism of the graph neural network, aggregate the neighborhood node features to generate the vector representation of the process rules.

[0037] Preferably, in step S1, the multi-modal data fusion module uses an attention mechanism to weight the features of different data sources.

[0038] Preferably, it further includes:

[0039] Step S8: Use an online learning mechanism to dynamically update the model parameters. When a change in the assembly line configuration or a new process is introduced is detected, trigger an incremental training process and only fine-tune the network layer for the new data.

[0040] Preferably, the present invention further includes an industrial flexible assembly process planning system based on artificial intelligence, and the system includes:

[0041] A multi-modal data acquisition and processing module for real-time acquisition of multi-source heterogeneous data on the assembly line, including material attributes, equipment operating status, environmental parameters, and historical process data, and performing feature extraction and normalization processing through a multi-modal data fusion module to generate a structured input vector;

[0042] A dynamic assembly path planning module that constructs a dynamic assembly path planning model based on deep reinforcement learning, learns the long-term benefits of different assembly sequences and resource allocation strategies through a Q-network, defines the state space as workstation load, material supply status, and order priority, and the actions as assembly path selection and equipment parameter adjustment;

[0043] A multi-objective optimization module that establishes a multi-objective optimization function with the optimization objectives of minimizing the assembly cycle time, equipment energy consumption, and material waste rate, and generates a Pareto front solution set in combination with constraint conditions, where the constraint conditions include workstation capacity limitations, process sequence irreversibility, and real-time order change requirements;

[0044] An iterative optimization module that iteratively optimizes the Pareto front solution set using an improved genetic algorithm, introduces an adaptive crossover probability and mutation operator, and screens the optimal assembly process plan through an elitist retention strategy;

[0045] A virtual assembly simulation verification module that constructs a virtual assembly simulation environment based on digital twin technology, real-time verifies the feasibility of the optimization plan, and dynamically adjusts the reward function weights of the Q-network according to the simulation results.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] In terms of data processing and utilization, it real-time acquires multi-source heterogeneous data on the assembly line, and performs feature extraction and normalization processing through a multi-modal data fusion module to generate a structured input vector. This enables the system to fully utilize various types of information in the production process, including material attributes, equipment operating status, environmental parameters, and historical process data, etc. Through in-depth analysis of these data, the system can more accurately grasp the production situation, provide a solid data foundation for subsequent assembly process planning, and avoid decision-making mistakes caused by data loss or inaccuracy.

[0048] In terms of assembly path planning, a dynamic assembly path planning model is constructed based on deep reinforcement learning. By learning the long-term benefits of different assembly sequences and resource allocation strategies through the Q-network, the workstation load, material supply status, and order priority are incorporated into the state space, and the assembly path selection and equipment parameter adjustment are regarded as actions. This intelligent planning method can dynamically adjust the assembly strategy according to the real-time state, effectively improving the assembly efficiency. For example, when the material supply is insufficient, the model can automatically adjust the assembly path, prioritize the orders with less dependence on the material, and avoid assembly stagnation caused by material shortage, thus greatly shortening the overall assembly cycle.

[0049] From the perspective of multi-objective optimization, a multi-objective optimization function is established, with the optimization objectives of minimizing the assembly cycle time, equipment energy consumption, and material waste rate, and combined with constraints such as workstation capacity limitation, process sequence irreversibility, and real-time order change requirements to generate the Pareto front solution set. This enables the system to find a balance among multiple objectives and achieve the maximization of overall benefits. For example, through the optimization algorithm, the equipment operation time and material usage can be reasonably arranged on the premise of ensuring the assembly quality, which not only reduces the equipment energy consumption and material waste, but also meets the order delivery time requirements.

[0050] The application of the improved genetic algorithm further enhances the optimization effect. An integer sequence is used for chromosome coding to represent the assembly path sequence, and the adaptive crossover probability and mutation operator can be dynamically adjusted according to the population diversity. Combining the elite retention strategy to screen the optimal assembly process plan. This algorithm can quickly search for a better assembly plan in the huge solution space, improving the convergence speed and optimization accuracy of the algorithm, and ensuring that the system can continuously provide high-quality assembly planning.

[0051] Constructing a virtual assembly simulation environment based on digital twin technology is a major highlight of the present invention. By verifying the feasibility of the optimization plan in real time and dynamically adjusting the reward function weight of the Q-network according to the simulation results, problems that may occur during the assembly process, such as assembly conflicts and equipment interference, can be discovered in advance. This not only avoids the cost increase caused by errors in actual production, but also further improves the system performance by continuously optimizing the reward function to guide the system to learn better assembly strategies.

[0052] In terms of multi-production line collaborative optimization, a federated learning framework is adopted to achieve multi-production line collaborative optimization. Each local model updates the global model parameters through encrypted gradient aggregation, which not only protects data privacy but also realizes information sharing and collaborative work among multiple production lines. This helps to improve the efficiency of the entire enterprise production system, achieve reasonable resource allocation, and enhance the market competitiveness of the enterprise.

[0053] The construction of the knowledge graph and the application of graph neural networks provide powerful knowledge support for the system. By storing historical optimization solutions and process rules, mining implicit constraint relationships, and assisting the decision-making reasoning of the Q network, the system can draw on past experience, make more reasonable decisions, and further improve the accuracy and reliability of the assembly process planning.

[0054] The introduction of the online learning mechanism enables the system to adapt to the dynamic changes of the assembly line. When a change in the assembly line configuration or the introduction of a new process is detected, an incremental training process is triggered, and only the network layer is fine-tuned for the newly added data, ensuring that the system can be adjusted and optimized in a timely manner in the face of production environment changes and continuously maintain good performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is the working principle diagram of the industrial flexible assembly process planning method described in the present invention;

[0056] Figure 2 is the construction step diagram of the multi-objective optimization function;

[0057] Figure 3 is the optimization step diagram of the improved genetic algorithm;

[0058] Figure 4 is the flowchart of the dynamic update monitoring of the assembly line model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] Please refer to Figures 1-4 , the present invention provides a technical solution: an industrial flexible assembly process planning method based on artificial intelligence, the method includes:

[0061] Step S1: Real-time collect multi-source heterogeneous data of the assembly line, including material attributes, equipment operating status, environmental parameters, and historical process data. Use the multi-modal data fusion module to extract features and normalize the collected data, and finally generate a structured input vector. By collecting rich and diverse data, a comprehensive information basis is provided for subsequent analysis and decision-making, ensuring the accuracy and reliability of the assembly process planning. The multi-modal data fusion module uses an attention mechanism to weight the features of different data sources, which can highlight important data features and improve the data processing effect.

[0062] Step S2: Build a dynamic assembly path planning model based on deep reinforcement learning. Learn the long-term benefits of different assembly sequences and resource allocation strategies through the Q-network. Define the state space as the workstation load, material supply status, and order priority, and set the actions as the assembly path selection and equipment parameter adjustment. This model can intelligently select the optimal assembly path and equipment parameters based on the current state information to optimize the assembly process.

[0063] Step S3: Establish a multi-objective optimization function with the optimization objectives of minimizing the assembly cycle time, equipment energy consumption, and material waste rate. Generate the Pareto front solution set in combination with the constraint conditions, where the constraint conditions cover the workstation capacity limit, process sequence irreversibility, and real-time order change requirements. By comprehensively considering multiple optimization objectives and actual constraint conditions, the optimal assembly plan that meets various requirements can be found.

[0064] Step S4: Use an improved genetic algorithm to iteratively optimize the Pareto front solution set. Introduce an adaptive crossover probability and mutation operator, and screen the optimal assembly process plan through the elite retention strategy. This algorithm can efficiently search for the optimal solution in the solution space and continuously optimize the assembly process plan.

[0065] Step S5: Build a virtual assembly simulation environment based on digital twin technology. In this environment, verify the feasibility of the optimization plan in real time and dynamically adjust the reward function weight of the Q-network according to the simulation results. The virtual assembly simulation environment provides a safe and efficient platform for plan verification, which can detect problems in advance and optimize them.

[0066] The present invention will be further described below in conjunction with Embodiments 1 to 5:

[0067] Embodiment 1:

[0068] This embodiment mainly elaborates on the detailed construction process of the dynamic assembly path planning model in Step S2 and the specific implementation method of the multi-modal data fusion module in Step S1.

[0069] In Step S1, the multi-modal data fusion module uses the attention mechanism to weight the features of different data sources. For example, among the collected material attributes, equipment operating status, environmental parameters, and historical process data, for some key processes, the quality attributes of materials and the key operating parameters of equipment may be more important for the assembly process planning. Through the attention mechanism, higher weights can be assigned to these important data, so that these key information can be more fully reflected in the feature extraction and normalization processes. Specifically, a neural network can be used to learn the importance of different data sources and then dynamically adjust the weights.

[0070] In Step S2, further refine its implementation process. Define the state space vector as S = (Li , M j , P k ), where L i represents the real-time load rate of the i-th workstation. During the actual assembly process, the working time, the number of tasks, etc. of the workstation are monitored in real time through sensors installed on the workstation, and then the real-time load rate is calculated. M j represents the inventory balance of the j-th type of material, which can be obtained by interacting with the material inventory management system. P k is the priority coefficient of the k-th order. The order priority can be preset according to factors such as the delivery time of the order and the importance of the customer.

[0071] Construct a double deep Q-network architecture, including a main network and a target network. The main network outputs the action value function Q(s, a; θ), where s is the current state, a is the action, and θ is the network parameter. The main network predicts the values corresponding to different actions according to the current state information. The target network is used to calculate the target value Q′(s′, a′; θ′), where s′ is the next state, a′ is the next action, and θ′ is the target network parameter. The existence of the target network can make the learning process more stable and reduce the fluctuations during the update of the main network. By regularly copying the parameters of the main network to the target network, it is ensured that the target network can reflect the current state of the main network.

[0072] Design the reward function R = α·T reduce + β·E save - γ·W waste , where T reduce is the reduction amount of the assembly cycle, E save is the reduction rate of energy consumption, W waste is the material waste rate, and α, β, γ are dynamic weight coefficients, and satisfy α + β + γ = 1. In practical applications, these weight coefficients can be dynamically adjusted according to different assembly scenarios and enterprise requirements. If the enterprise currently pays more attention to energy conservation, then the value of β can be appropriately increased and other coefficients can be reduced. By reasonably designing the reward function, the Q-network is guided to learn an assembly strategy that better meets the actual needs.

[0073] Example 2:

[0074] This example details the specific application and calculation method of the multi-objective optimization function and constraint conditions in step S3. In step S3, the multi-objective optimization function is defined as:

[0075] minF(x) = [f1(x), f2(x), f3(x)]

[0076] where Here, T i(x) represents the time of the i-th assembly process. In an actual assembly line, each assembly process has its specific operation procedure and time consumption. Through the analysis of historical process data and on-site actual measurement, the average time of each assembly process can be obtained. For example, for a certain assembly process, its average assembly time is obtained as 5 minutes through multiple measurements. When calculating the multi-objective optimization function, the time of this process can be substituted into T i (x). By accumulating the times of all assembly processes, the objective function value of the entire assembly cycle time is obtained, and this is used to measure the length of the assembly cycle.

[0077] Among them, E j (x) represents the energy consumption of the j-th device. The energy consumption of the device can be obtained through the energy consumption monitoring device installed on the device. For example, a certain device consumes 10 degrees of electricity in an assembly task. Substitute this value into E j (x), calculate the sum of the energy consumption of all devices, and use it as the objective function value of the energy consumption of the device to evaluate the energy consumption situation.

[0078] Among them, W k (x) represents the waste rate of the k-th type of material. The waste rate of the material can be calculated by comparing the actual quantity of the material used and the standard quantity of the material. For example, the standard quantity of a certain type of material is 100, and the actual quantity used is 110. Then its waste rate is (110 - 100)÷100 = 10%. Substitute this value into W k (x), calculate the sum of the waste rates of all materials, and use it as the objective function value of the waste rate of the material.

[0079] The constraint conditions include: g1(x) = C max -C i (x) ≥ 0, where C max is the maximum capacity of the workstation, and C i (x) is the actual load. In the actual assembly process, the bearing capacity of the workstation is limited. For example, the maximum capacity of a certain workstation is to be able to process 10 assembly tasks simultaneously. Through real-time monitoring, the actual load of the current workstation is obtained as 8 tasks. Substitute these values into the constraint condition formula for judgment. If the actual load exceeds the maximum capacity, then this assembly plan does not meet the requirements and needs to be adjusted again. g2(x) = O(x) - O min ≥ 0, where O(x) is the actual order completion rate, and O minThis is the minimum requirement. The actual order completion rate can be obtained by calculating the ratio of the number of actually completed orders to the total number of orders. Suppose the total number of orders is 100 and 90 orders are actually completed, then the actual order completion rate is 90%. If the minimum required order completion rate set by the enterprise is 85%, substitute these values into the constraint condition formula to determine whether the current assembly plan meets the order completion rate requirement.

[0080] Embodiment 3:

[0081] In step S4, the chromosome encoding of the improved genetic algorithm uses an integer sequence to represent the assembly path order, and the gene positions correspond to the workstation numbers and material distribution identifiers. For example, for a simple assembly system with 3 workstations and 2 types of materials, the chromosome encoding is [1, 2, 3, 1, 2], where the first three numbers 1, 2, 3 represent the workstation numbers passed through in sequence in the assembly path, and the last two numbers 1, 2 represent the material identifiers assigned in the corresponding assembly links. Through this encoding method, the assembly process and material distribution can be intuitively represented, facilitating subsequent genetic algorithm operations.

[0082] Adaptive crossover probability P c Is dynamically adjusted according to the population diversity, and the calculation formula is P c = 0.8 - 0.3·(G current / G max ), where G current Is the current iteration number, and G max Is the maximum iteration number. In the initial stage of the genetic algorithm, the population diversity is relatively high, and at this time the crossover probability is relatively high. For example, when G current Is 0, according to the formula, the calculated crossover probability P c Is 0.8. A higher crossover probability can promote gene exchange between individuals, accelerate the search speed of the algorithm, and explore a wider solution space. As the number of iterations increases, the population gradually converges, the diversity decreases, and the crossover probability also decreases. When G current Is close to G max , the crossover probability will become relatively low, avoiding excessive search of the algorithm and preventing the destruction of the relatively good solutions that have been found.

[0083] The mutation operator adopts a Gaussian perturbation strategy, applying a normal distribution random offset with a mean of 0 and a variance that decreases with the number of iterations to the selected gene positions, where the variance σ 2 = 1 / (G current + 1). In the initial stage of iteration, the variance is relatively large. For example, when G current Is 0, the variance σ 2It is 1. At this time, the mutation amplitude is relatively large, enabling the search for new solutions within a large range and increasing the diversity of the population. As the number of iterations increases, the variance gradually decreases, and the mutation amplitude also decreases accordingly. This allows the algorithm to optimize the solution more precisely in the later stage and avoid being far from the optimal solution due to too large a mutation amplitude. Through the synergistic effect of the adaptive crossover probability and the mutation operator, combined with the elite retention strategy, the optimal assembly process plan can be effectively screened out.

[0084] Example 4:

[0085] This example details the construction of a virtual assembly simulation environment and related verification and adjustment operations based on digital twin technology in step S5.

[0086] In step S5, a high-fidelity digital twin model is constructed based on the physical assembly line. First, detailed modeling is carried out on each component of the physical assembly line, including workstations, equipment, material transfer systems, etc. For example, for a workstation, its geometric model can be established to simulate its spatial layout and operation process; for equipment, its dynamic model can be established to simulate its operating state and performance parameters. By collecting the real-time operation data of the physical assembly line, such as the temperature and rotation speed of equipment, the position of materials, etc., the dynamic evolution of the virtual environment is synchronously driven, enabling the virtual assembly simulation environment to truly reflect the actual situation of the physical assembly line.

[0087] Random perturbation events are injected into the simulation environment, including equipment failure simulation and material supply delay. For example, simulate that a certain piece of equipment suddenly fails during the assembly process and stops running for a period of time; or simulate the supply delay of a certain type of material, resulting in the assembly line waiting for materials. Through these random perturbation events, the robustness of the optimization scheme is tested. Observe whether the assembly process can continue under these abnormal conditions, whether it can return to normal within a certain time, and the impact on indicators such as the assembly cycle, equipment energy consumption, and material waste rate.

[0088] According to the conflict detection information in the simulation results, the exploration strategy of the Q-network and the fitness function of the genetic algorithm are corrected in reverse. If a conflict is found in the assembly path during the simulation, such as two assembly tasks simultaneously requiring the use of the same piece of equipment, then according to the severity and occurrence frequency of the conflict, the exploration strategy of the Q-network is adjusted. For example, the exploration reward for avoiding conflict paths can be increased, and the exploration probability of paths prone to conflicts can be reduced. At the same time, the fitness function of the genetic algorithm is adjusted so that the fitness value of chromosomes containing conflict paths decreases, thereby reducing the probability of such chromosomes being selected in subsequent iterations and guiding the algorithm to search for a better conflict-free assembly scheme.

[0089] Example 5:

[0090] In step S6, a federated learning framework is adopted to achieve collaborative optimization of multiple production lines. When an enterprise has multiple assembly lines, the production situations and data of each assembly line are unique. To achieve collaborative optimization, local models are deployed on each assembly line. The local models are trained using multi-source heterogeneous data collected from their respective assembly lines, which include the unique material attributes, equipment operating status, environmental parameters, and historical process data of that assembly line.

[0091] During the federated learning process, the local models do not directly share the original data but calculate the gradients generated during model training. For example, through the calculation of locally collected data, gradient information about relevant model parameters such as assembly path planning and resource allocation strategies is obtained. Then, encryption technology is used to encrypt these gradients, and the encrypted gradient information is uploaded to the central server. The central server is responsible for aggregating these encrypted gradients, and through a specific algorithm, the gradients of each local model are weighted and fused to update the global model parameters. In this way, the collaborative utilization of multi-production line data is achieved, the global model is optimized, and the leakage of data privacy of each assembly line is avoided.

[0092] In step S7, a knowledge graph is constructed. First, it is clear that the entity nodes are divided into three categories: equipment, materials, and processes. For equipment nodes, information such as the model, function, manufacturer, and maintenance cycle of the equipment is recorded in detail; material nodes record attributes such as the name, specification, quantity, and supplier of the materials; process nodes record the name, operation steps, preceding process, and subsequent process of each assembly process.

[0093] The relationship edges are defined as "dependency", "conflict", and "substitution". For example, certain processes have a "dependency" relationship with specific equipment, that is, the process must be completed on a specific equipment; different materials may have a "conflict" relationship during the assembly process, such as two materials cannot be used in the same assembly link at the same time; there is a "substitution" relationship between certain materials or processes, and they can be replaced with each other under certain conditions.

[0094] The TransE algorithm is used for graph embedding, minimizing the loss function L = ‖h + r - t‖ of the triple (h, r, t) 2 , where h is the head entity embedding vector, r is the relationship vector, and t is the tail entity embedding vector. Through this algorithm, the entities and relationships in the knowledge graph are mapped to a low-dimensional vector space for subsequent calculation and analysis. Based on the message passing mechanism of the graph neural network, the features of neighboring nodes are aggregated to generate a vector representation of the process rules. For example, for a certain process node, through the message passing mechanism, the feature information of the adjacent equipment nodes, material nodes, and other process nodes is aggregated, so as to obtain the comprehensive feature vector of the process in the entire assembly process, providing richer knowledge support for the decision-making and reasoning of the Q network.

[0095] In step S9, an online learning mechanism is adopted to dynamically update the model parameters. During the operation of the assembly line, configuration changes or the introduction of new processes may occur. For example, a batch of advanced equipment is newly purchased and connected to the assembly line, or a new assembly process is adopted. At this time, the system detects these changes through methods such as sensor monitoring and data interface information interaction.

[0096] Once a configuration change of the assembly line or the introduction of a new process is detected, an incremental training process is triggered. During the incremental training process, only the network layer is fine-tuned for the newly added data. For example, when new equipment is introduced, the relevant operation data and process parameters of the new equipment are obtained, and the network layer parameters related to the equipment are fine-tuned, rather than retraining the entire model. This can ensure that the model adapts to the new changes while greatly reducing the consumption of training time and computing resources, enabling the model to promptly adapt to the dynamic changes of the assembly line, continuously maintain good performance, and provide more accurate and efficient support for industrial flexible assembly process planning.

[0097] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0098] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made therein without departing from the principle and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An industrial flexible assembly process planning method based on artificial intelligence, characterized in that, It includes the following steps: Step S1: Real-time collect multi-source heterogeneous data of the assembly line, including material attributes, equipment operation status, environmental parameters, and historical process data, and perform feature extraction and normalization processing through a multi-modal data fusion module to generate a structured input vector; Step S2: Build a dynamic assembly path planning model based on deep reinforcement learning. Through the Q-network, learn the long-term benefits of different assembly sequences and resource allocation strategies. Define the state space as workstation load, material supply status, and order priority, and the actions as assembly path selection and equipment parameter adjustment; Step S3: Establish a multi-objective optimization function, with the optimization objectives of minimizing the assembly cycle time, equipment energy consumption, and material waste rate. Combine the constraint conditions to generate a Pareto front solution set, where the constraint conditions include workstation capacity limitations, process sequence irreversibility, and real-time order change requirements; Step S4: Use an improved genetic algorithm to iteratively optimize the Pareto front solution set, introduce an adaptive crossover probability and mutation operator, and screen the optimal assembly process plan through an elite retention strategy; Step S5: Build a virtual assembly simulation environment based on digital twin technology, verify the feasibility of the optimization plan in real time, and dynamically adjust the reward function weight of the Q-network according to the simulation results.

2. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 1, wherein The said Step S2 further includes: Step S21: Define the state space vector as S = (L i , M j , P k ), where L i represents the real-time load rate of the i-th workstation, M j represents the inventory balance of the j-th type of material, and P k is the priority coefficient of the k-th order; Step S22: Construct a double deep Q - network architecture, including a main network and a target network. The main network outputs the action - value function Q(s, a; θ), where s is the current state, a is the action, and θ is the network parameter. The target network is used to calculate the target value Q ′ (s ′ , a ′ ; θ ′ ), where s ′ is the next state, a ′ is the next action, and θ ′ is the target network parameter; Step S23: Design the reward function \(R = \alpha\cdot T\) reduce +\(\beta\cdot E\) save -\(\gamma\cdot W\) waste , where \(T\) reduce is the reduction in assembly cycle, \(E\) save is the reduction rate of energy consumption, \(W\) waste is the material waste rate, and \(\alpha\), \(\beta\), \(\gamma\) are dynamic weight coefficients, and satisfy \(\alpha+\beta+\gamma = 1\).

3. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 1, wherein In the said Step S3, the multi-objective optimization function is defined as: minF(x)=[f1(x),f2(x),f3(x)] Among them Among them, T i (x) represents the time of the i-th assembly process Among them, E j (x) represents the energy consumption of the j-th device Among them, W k (x) represents the waste rate of the k-th type of material The constraints include: g1(x) = C max -C i (x) ≥ 0, where C max is the maximum capacity of the workstation, and C i (x) is the actual load; g2(x) = O(x) - O min ≥ 0, where O(x) is the actual order completion rate and O min is the minimum requirement.

4. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 1, wherein The operations of the improved genetic algorithm in the said Step S4 include: Step S41: The chromosome encoding uses an integer sequence to represent the assembly path order, and the gene positions correspond to the workstation numbers and material distribution identifiers; Step S42: Adaptive crossover probability P c It is dynamically adjusted according to the population diversity, and the calculation formula is P c = 0.8 - 0.3·(G current / G max ), where G current is the current iteration number, and G max is the maximum iteration number; Step S43: The mutation operator adopts a Gaussian perturbation strategy to apply a random offset of a normal distribution with a mean of 0 and a variance that decreases with the number of iterations to the selected gene positions, where the variance σ 2 = 1 / (G current + 1).

5. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 1, wherein, The said Step S5 further includes: Step S51: Build a high-fidelity digital twin model based on the physical assembly line, and synchronize real-time data to drive the dynamic evolution of the virtual environment; Step S52: Inject random perturbation events into the simulation environment, including equipment failure simulation and material supply delay, to test the robustness of the optimization plan; Step S53: According to the conflict detection information in the simulation results, reverse-correct the exploration strategy of the Q-network and the fitness function of the genetic algorithm.

6. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 1, characterized in that It also includes: Step S6: Use a federated learning framework to achieve multi-production line collaborative optimization. Each local model updates the global model parameters through encrypted gradient aggregation while preserving data privacy; Step S7: Define a knowledge graph to store historical optimization plans and process rules, and mine implicit constraint relationships through a graph neural network to assist the decision-making and reasoning of the Q-network.

7. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 6, wherein The construction of the knowledge graph in the said Step S7 includes: Step S71: The entity nodes are divided into three categories: equipment, materials, and processes, and the relationship edges are defined as "dependency", "conflict", and "substitution"; Step S72: Perform graph embedding using the TransE algorithm to minimize the loss function L = ‖h + r - t‖ of the triple (h, r, t), where h is the head entity embedding vector, r is the relation vector, and t is the tail entity embedding vector; 2 , where h is the head entity embedding vector, r is the relation vector, and t is the tail entity embedding vector; Step S73: Based on the message passing mechanism of the graph neural network, aggregate the neighborhood node features to generate a vector representation of the process rules.

8. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 1, wherein In the said Step S1, the multi-modal data fusion module uses an attention mechanism to weight the features of different data sources.

9. The method for planning an industrial flexible assembly process based on artificial intelligence according to claim 1, characterized in that It also includes: Step S8: Use an online learning mechanism to dynamically update the model parameters. When a change in the assembly line configuration or a new process is introduced, trigger an incremental training process and only fine-tune the network layer for the new data.

10. An industrial flexible assembly process planning system based on artificial intelligence, characterized in that, It includes: The multi-modal data acquisition and processing module is used to collect multi-source heterogeneous data of the assembly line in real time, including material attributes, equipment operation status, environmental parameters and historical process data, and perform feature extraction and normalization processing through the multi-modal data fusion module to generate a structured input vector; The dynamic assembly path planning module constructs a dynamic assembly path planning model based on deep reinforcement learning, learns the long-term benefits of different assembly sequences and resource allocation strategies through the Q-network, defines the state space as the workstation load, material supply status and order priority, and the action as the assembly path selection and equipment parameter adjustment; The multi-objective optimization module establishes a multi-objective optimization function, with the minimization of the assembly cycle time, equipment energy consumption and material waste rate as the optimization objectives, and generates a Pareto front solution set in combination with the constraint conditions, where the constraint conditions include workstation capacity limitation, process sequence irreversibility and real-time order change requirements; The iterative optimization module uses an improved genetic algorithm to iteratively optimize the Pareto front solution set, introduces an adaptive crossover probability and mutation operator, and screens the optimal assembly process plan through the elitist retention strategy; The virtual assembly simulation verification module constructs a virtual assembly simulation environment based on digital twin technology, verifies the feasibility of the optimization plan in real time, and dynamically adjusts the reward function weight of the Q-network according to the simulation results.

Images