Robot cable manufacturing process optimization method and system based on deep reinforcement learning

Abstract

The invention provides a robot cable manufacturing process optimization method and system based on deep reinforcement learning, and relates to the technical field of intelligent manufacturing, and the method comprises the steps: collecting cable manufacturing data, constructing a digital twin model, and achieving the process flow simulation through a graph network; building a deep reinforcement learning environment by taking the manufacturing data and the simulation data as state input; analyzing a causal dependency relationship of process parameter adjustment; constructing an industrial knowledge graph expert system to generate an optimization strategy; and a process optimization closed loop is formed. According to the invention, self-adaptive optimization of the cable manufacturing process is realized, and the manufacturing efficiency and quality are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and particularly to a method and system for optimizing the robot cable manufacturing process based on deep reinforcement learning. Background Art

[0002] Robot cable manufacturing is an important production link in the electronic product and automotive industries. Traditional cable manufacturing processes mainly rely on manual experience to adjust process parameters, including key factors such as temperature, pressure, and speed. With the development of industry and intelligent manufacturing, the level of intelligence and automation in cable manufacturing processes has been continuously improved, and digital technologies such as digital twins and artificial intelligence have been widely applied in the manufacturing process.

[0003] In the current industrial field, deep reinforcement learning technology has achieved certain results in the field of process parameter optimization. By establishing a state-action mapping relationship, automatic adjustment of parameters in the manufacturing process can be realized. Digital twin technology realizes high-precision simulation and prediction of the actual manufacturing process by constructing a virtual model of a physical entity. At the same time, industrial knowledge graph technology provides knowledge support for manufacturing decisions by establishing the association relationship between domain concepts.

[0004] However, the existing methods for optimizing the robot cable manufacturing process have obvious deficiencies; existing methods often regard the adjustment of process parameters as an independent decision-making problem, ignoring the causal dependence relationship between parameters, resulting in the lack of systematicness and coordination of the optimization strategy; traditional optimization methods are difficult to effectively integrate the learning results driven by data and expert experience knowledge, and often have problems of insufficient robustness and poor adaptability when facing complex and changeable manufacturing environments; existing methods lack a closed-loop optimization mechanism and cannot continuously adjust and improve process parameters according to real-time production data, resulting in the difficulty of maintaining the optimization effect in the long term, especially when the equipment is aging or the material characteristics change. Summary of the Invention

[0005] An embodiment of the present invention provides a method and system for optimizing the robot cable manufacturing process based on deep reinforcement learning, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention, a method for optimizing the robot cable manufacturing process based on deep reinforcement learning is provided, including: Collect cable manufacturing data; construct a digital twin model, construct the execution components of the manufacturing equipment as graph network nodes, construct the process flow between components as graph network edges, and obtain manufacturing process simulation data through the graph dynamics network and state iteration optimization of thermal coupling and fluid-structure coupling; Use the cable manufacturing data and the manufacturing process simulation data as state inputs, and use the process parameter adjustment instruction as the action output to establish a deep reinforcement learning environment; In the reinforcement learning training environment, analyze the causal dependencies of process parameter adjustment instructions, establish a parameter adjustment dependency graph, and generate an initial process control strategy based on the parameter adjustment dependency graph; Construct an industrial knowledge graph expert system in the form of triples, fuse the initial process optimization strategy with the expert rule base based on the credibility weight, and generate an improved process optimization strategy through a hybrid intelligent method that combines symbolic reasoning chains and deep learning feature extraction; Adjust the process parameters of the robot cable manufacturing equipment according to the improved process optimization strategy, and feedback the cable manufacturing data collected after adjustment to the digital twin model to form a process optimization closed loop.

[0007] In an optional embodiment, construct a digital twin model, construct the execution components of the manufacturing equipment as graph network nodes, construct the process flow between components as graph network edges, and obtain manufacturing process simulation data through graph dynamics network and state iterative optimization of thermal coupling and fluid-structure coupling, including: Form a state vector by combining the position vector, velocity vector, force vector, and temperature field distribution of the execution components of the cable manufacturing equipment, and generate a node feature vector through a feature encoder; Collect the heat flux density, mass flow rate, and stress tensor of the process flow between the execution components, and form a process flow feature vector; Based on the node feature vector and the process flow feature vector, construct a thermal coupling field equation, calculate the temperature field distribution and stress tensor, establish a fluid-structure interface coupling model, and generate fluid-structure coupling data; Input the node feature vector and the fluid-structure coupling data into a gated recurrent unit, use a message passing matrix to fuse adjacent node information, and generate a node feature update vector; Calculate the change in edge attributes based on the node feature update vector and the process flow feature vector, and generate a process flow feature update vector; Input the node feature update vector and the process flow feature update vector into a state transition function, generate a predicted state vector, collect measurement data and perform error correction through a Kalman gain matrix to generate a state correction vector; Use the state correction vector as the new state vector and execute it in a loop until the simulation accuracy meets the preset accuracy threshold.

[0008] In an optional embodiment, based on the node feature vector and the process flow feature vector, construct a thermal coupling field equation, calculate the temperature field distribution and stress tensor, establish a fluid-structure interface coupling model, and generate fluid-structure coupling data including: Decompose the node feature vector into a thermal field feature component and a force field feature component, and decompose the process flow feature vector into a flow feature component and a heat transfer feature component; Construct a temperature field control equation based on the thermal field characteristic components and the heat transfer characteristic components. The temperature field control equation includes terms of material density, specific heat capacity, thermal conductivity, as well as an external heat source term and a deformation heat source term determined by the node eigenvector, where the deformation heat source term is obtained through the coupling calculation of stress and strain rate; Establish a thermal stress field equation according to the force field characteristic components and the flow characteristic components, and perform a coupling calculation on the elastic strain term, the thermal expansion strain term, and the viscous stress term determined by the process flow eigenvector; Establish a fluid-structure interface coupling model, associate the temperature gradients of the solid domain and the fluid domain through the interface heat flux continuity equation, and associate the stresses of the solid domain and the fluid domain through the interface stress balance equation; Form a coupled matrix equation from the temperature field control equation and the thermal stress field equation, and use block iterative solution to obtain the temperature field distribution and stress tensor, and generate fluid-structure coupling data.

[0009] In an alternative embodiment, in the reinforcement learning training environment, analyze the causal dependence relationship of process parameter adjustment instructions, establish a parameter adjustment dependence graph, and generate an initial process control strategy based on the parameter adjustment dependence graph, including: Collect historical data of process parameter adjustment instructions, establish the joint probability distribution and marginal probability distribution between process parameter adjustment instructions, and calculate the mutual information amount and partial correlation coefficient between process parameter adjustment instructions using the joint probability distribution and marginal probability distribution; Construct a process parameter adjustment dependence graph based on the mutual information amount and partial correlation coefficient, and assign weight values to the causal dependence relationships in the process parameter adjustment dependence graph using the sensitivity analysis results of process parameter adjustment instructions; Form a state vector from the current value and change rate of process parameter adjustment instructions, and construct a process control strategy according to the hierarchical structure of the process parameter adjustment dependence graph, so that the results of upper-level process parameter adjustment instructions guide the decision-making of lower-level process parameter adjustment instructions; Import the process control strategy into the reinforcement learning training environment for testing, record the state transition results and causal effects of process parameter adjustment instructions, and construct a process control strategy evaluation function based on the causal effects and operation costs; Use the process control strategy evaluation function to calculate the stability index of the state vector, and optimize the process control strategy in combination with the evaluation results under the perturbation of process parameter adjustment instructions to generate an initial process control strategy.

[0010] In an alternative embodiment, form a state vector from the current value and change rate of process parameter adjustment instructions, and construct a process control strategy according to the hierarchical structure of the process parameter adjustment dependence graph, so that the results of upper-level process parameter adjustment instructions guide the decision-making of lower-level process parameter adjustment instructions, including: The current value and the change rate of the process parameter adjustment instruction are formed into an initial state vector, the mean and standard deviation of the process parameter adjustment instruction are calculated, and the initial state vector is dynamically normalized to generate a standardized state vector; Based on the process parameter adjustment dependency graph, calculate the path weights of adjacent nodes, substitute the path weights into the node level determination function, calculate the level values of the nodes corresponding to each process parameter adjustment instruction, and calculate the association strength between the nodes within the layer according to the level values and path weights; Taking the standardized state vector as the input, construct a multi-layer conditional policy function according to the level values, use the level reward value of the upper-layer process parameter adjustment instruction as the conditional constraint of the lower-layer process parameter adjustment instruction, and determine the parameters of the inter-layer coupling constraint function through the association strength; Construct a global reward function based on the inter-layer coupling constraint function, construct a local reward function according to the standardized state vector, combine the global reward function and the local reward function to generate a level reward function, and weight the level reward function with a level weight coefficient to construct a multi-level policy evaluation function; Record the decision sequence of the upper-layer process parameter adjustment instruction, input the decision sequence into the level decision function to generate a candidate decision set for the lower-layer process parameter adjustment instruction, calculate the change amount of the multi-level policy evaluation function under the candidate decision, determine the feedback adjustment factor according to the change amount and dynamically optimize the output of the level decision function to generate a hierarchical process control strategy.

[0011] In an alternative embodiment, construct an industrial knowledge graph expert system in the form of triples, fuse the initial process optimization strategy and the expert rule base based on the credibility weight, and generate an improved process optimization strategy through a hybrid intelligent method combining symbolic reasoning chain and deep learning feature extraction, including: Construct an industrial knowledge graph expert system, represent the expert experience rules in the form of triples, each triple contains a process parameter node, a process parameter relationship, and a process control rule, and form an expert rule base of the industrial knowledge graph; Calculate the matching degree between the process parameter configuration in the initial process optimization strategy and the process control rules in the expert rule base, assign a credibility weight to each process control rule based on the matching degree, and establish a priority ranking of the process control rules; Construct a symbolic reasoning chain according to the priority order, input the process parameter configuration of the initial process optimization strategy into the symbolic reasoning chain, judge the activation state of the process control rules based on the credibility weight, and output the symbolic reasoning result; Input the process parameter node and the process parameter relationship into the deep learning model for feature extraction to obtain a process optimization feature vector, combine the process optimization feature vector and the symbolic reasoning result to construct a hybrid policy network, and output a candidate process optimization scheme; Construct an evaluation function that includes the rule compliance degree and the optimization effect, calculate the evaluation score of the candidate process optimization scheme under the evaluation function, and adjust the fusion weights of the hybrid strategy network according to the evaluation score to generate an improved process optimization strategy.

[0012] In an alternative embodiment, calculate the matching degree between the process parameter configuration in the initial process optimization strategy and the process control rules in the expert rule base, assign credibility weights to each process control rule based on the matching degree, and establish a priority ranking of the process control rules, including: Construct a semantic feature matrix and a structural feature matrix of the process parameter configuration, input the semantic feature matrix into a bidirectional attention network to calculate the semantic association strength between parameters, and input the structural feature matrix into a graph convolutional network to extract the topological relationship features between parameters; Adopt a hierarchical attention mechanism to calculate the matching scores between the process parameter configuration and the process control rules at different abstraction levels, and combine the semantic association strength and the topological relationship features to construct a rule matching degree evaluation model; Based on a preset rule adaptability evaluation index, input the rule matching degree and the rule adaptability index into a dynamic weighting function to calculate the comprehensive credibility of the process control rules; Adopt a multi-objective optimization algorithm, combine the rule conflict degree and the rule coverage degree, normalize the comprehensive credibility of the process control rules, and establish a priority ranking of the process control rules.

[0013] In the second aspect of the embodiments of the present invention, a robot cable manufacturing process optimization system based on deep reinforcement learning is provided, including: A first unit for collecting cable manufacturing data; constructing a digital twin model, constructing the execution components of the manufacturing equipment as graph network nodes, constructing the process flow between components as graph network edges, and obtaining manufacturing process simulation data through a graph dynamics network and state iteration optimization of thermal coupling and fluid-structure coupling; A second unit for using the cable manufacturing data and the manufacturing process simulation data as states as inputs and the process parameter adjustment instruction as an action output to establish a deep reinforcement learning environment; A third unit for analyzing the causal dependence relationship of the process parameter adjustment instruction in the reinforcement learning training environment, establishing a parameter adjustment dependence graph, and generating an initial process control strategy based on the parameter adjustment dependence graph; A fourth unit for constructing an industrial knowledge graph expert system in the form of a triple, fusing the initial process optimization strategy with the expert rule base based on the credibility weight, and generating an improved process optimization strategy through a hybrid intelligent method combining symbolic reasoning chains and deep learning feature extraction; The fifth unit is used to adjust the process parameters of the robot cable manufacturing equipment according to the improved process optimization strategy, and feed back the cable manufacturing data collected after adjustment to the digital twin model to form a process optimization closed loop.

[0014] In a third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In the embodiments of the present invention, through constructing a digital twin model and a deep reinforcement learning environment, the intelligent optimization of the robot cable manufacturing process is realized, the production cost is effectively reduced, and the manufacturing efficiency and product quality are improved. Compared with traditional methods, the system can adaptively cope with changes in the production environment and fluctuations in process parameters, and maintain stable manufacturing quality; integrating graph network modeling, reinforcement learning and knowledge graph technology, a complete set of hybrid intelligent process optimization methods is formed, reducing the dependence on expert experience, improving the accuracy and response speed of process parameter adjustment, being able to quickly adapt to the manufacturing requirements of different types of cables, and greatly shortening the process debugging time; the established closed-loop optimization system feeds back the cable manufacturing data to the digital twin model in real time, continuously improves the process optimization strategy, realizes the continuous optimization and knowledge accumulation of the manufacturing process, greatly enhances the reliability and stability of the production system, reduces the defective rate, extends the service life of the equipment, and provides an efficient intelligent solution for the field of robot cable manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flowchart of the method for optimizing the robot cable manufacturing process based on deep reinforcement learning according to the embodiments of the present invention; Figure 2 It is a comparative analysis chart of temperature distribution; Figure 3 It is an analysis chart of the stability of process parameter adjustment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below 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 of 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.

[0019] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0020] Figure 1 The following is a schematic flowchart of a method for optimizing a robot cable manufacturing process based on deep reinforcement learning according to an embodiment of the present invention, as Figure 1 shown. The method includes: Collect cable manufacturing data; construct a digital twin model, construct the execution components of the manufacturing equipment as graph network nodes, construct the process flow between components as graph network edges, and obtain manufacturing process simulation data through graph dynamics network and state iterative optimization of thermal coupling and fluid-structure coupling; Use the cable manufacturing data and manufacturing process simulation data as state inputs, and use the process parameter adjustment instruction as the action output to establish a deep reinforcement learning environment; In the reinforcement learning training environment, analyze the causal dependence relationship of the process parameter adjustment instructions, establish a parameter adjustment dependence graph, and generate an initial process control strategy based on the parameter adjustment dependence graph; Construct an industrial knowledge graph expert system in the form of a triple, fuse the initial process optimization strategy with the expert rule base based on the credibility weight, and generate an improved process optimization strategy through a hybrid intelligent method combining symbolic reasoning chain and deep learning feature extraction; Adjust the process parameters of the robot cable manufacturing equipment according to the improved process optimization strategy, and feedback the cable manufacturing data collected after adjustment to the digital twin model to form a process optimization closed loop.

[0021] In an alternative embodiment, constructing a digital twin model, constructing the execution components of the manufacturing equipment as graph network nodes, constructing the process flow between components as graph network edges, and obtaining manufacturing process simulation data through graph dynamics network and state iterative optimization of thermal coupling and fluid-structure coupling includes: Form a state vector from the position vector, velocity vector, force vector, and temperature field distribution of the execution components of the cable manufacturing equipment, and generate a node feature vector through a feature encoder; Collect the heat flux density, mass flow rate, and stress tensor of the process flow between the execution components to form a process flow feature vector; Construct a thermo-mechanical coupling field equation based on the node feature vector and the process flow feature vector, calculate the temperature field distribution and stress tensor, establish a fluid-structure interface coupling model, and generate fluid-structure coupling data; Input the node feature vector and the fluid-structure coupling data into a gated recurrent unit, and use a message passing matrix to fuse adjacent node information to generate a node feature update vector; Calculate the edge attribute change based on the node feature update vector and the process flow feature vector to generate a process flow feature update vector; Input the node feature update vector and the process flow feature update vector into a state transition function to generate a predicted state vector, collect measurement data and perform error correction through a Kalman gain matrix to generate a state correction vector; Use the state correction vector as the new state vector and execute it in a loop until the simulation accuracy meets the preset accuracy threshold.

[0022] In a specific embodiment, model the execution components of the cable manufacturing equipment, and construct each execution component as a node in the graph network. Taking an actual application as an example, for a cable extrusion molding equipment, key execution components such as an extruder screw, a heating barrel, a die head, and a cooling tank can be used as nodes. The state vector of each node includes its position vector (x, y, z), velocity vector (vx, vy, vz), force vector (Fx, Fy, Fz), and temperature field distribution T(x, y, z). For example, the state vector of the extruder screw node can be expressed as (150.5mm, 0mm, 0mm, 0mm / s, 0mm / s, 15mm / s, 0N, 0N, 150N, 210°C).

[0023] Encode the state vector of each node through a feature encoder to generate a node feature vector. The feature encoder adopts a three-layer fully connected network structure. The number of neurons in the input layer is the same as the dimension of the state vector, the number of neurons in the hidden layer is 128, and the number of neurons in the output layer is 64, generating a 64-dimensional node feature vector. The ReLU activation function is used to enhance the network expression ability. After encoding, the node feature vector generated by the above extruder screw node represents its key working characteristics.

[0024] At the same time, construct the process flow between the execution components as an edge in the graph network. The system collects process flow feature data, including heat flux density q (unit: W / m²), mass flow rate G (unit: kg / s), and stress tensor σ (unit: Pa), to form a process flow feature vector. Taking the process flow between the extruder screw and the heating barrel as an example, its feature vector can be expressed as (1250W / m², 0.25kg / s, 2.5MPa).

[0025] Based on the node feature vector and the process flow feature vector, a thermo-mechanical coupling field equation is constructed. The system uses the finite difference method to solve the temperature field distribution, with a grid division accuracy of 0.5 mm and a time step of 0.01 s. The heat conduction equation is used to calculate the temperature field in the solid region, and the combined effects of heat convection and heat conduction are considered in the fluid region. For example, when the molten material flows through the die head, the system calculates that the inner wall temperature distribution of the die head is 180 - 190 °C, and the outer wall temperature is 175 - 185 °C, meeting the processing requirements. At the same time, the stress distribution of each executing component is calculated based on the finite element method. The hexahedral element is selected as the element type, with an element size of 1 mm. The maximum stress obtained from the calculation is 2.3 MPa, located at the die head contraction section.

[0026] At the fluid-solid interface, a fluid-solid coupling model is established, and a segregated solution strategy is adopted to alternately solve the fluid domain and the solid domain. In each time step, the solid stress-strain equation is first solved to obtain the deformation field, then the fluid motion equation is solved to obtain the fluid pressure and velocity field, and then information exchange is carried out at the boundary to achieve two-way fluid-solid coupling. For example, the maximum pressure generated by the molten material in the die head on the wall is 1.8 MPa, resulting in an elastic deformation of 0.02 mm of the die head. The system completes the convergence calculation through two iterations.

[0027] The node feature vector and the fluid-solid coupling data are input into a gated recurrent unit (GRU) for temporal feature learning. The GRU unit contains an update gate and a reset gate, with a hidden state dimension of 128 and a time step of 5. The system uses a message passing matrix to fuse the information of adjacent nodes. The message passing radius is set to 1, that is, only directly connected nodes are considered. The size of the message passing matrix is the number of nodes × the number of nodes, and the element value of 1 indicates the existence of a connection between nodes, and 0 indicates no connection. The neighbor information is aggregated through a message aggregation function (using average pooling operation) to generate a node feature update vector.

[0028] Based on the node feature update vector and the process flow feature vector, the system calculates the change in edge attributes. A two-layer fully connected network with a hidden layer dimension of 32 is used as the edge update function to process the feature update vectors of the source node and the target node and the current process flow feature vector to generate a process flow feature update vector. For example, as the rotational speed of the extruder screw increases from 15 rpm to 18 rpm, the flow rate of the molten material is calculated to increase from 0.25 kg / s to 0.28 kg / s, and the heat flux density increases from 1250 W / m² to 1320 W / m² through the edge update function.

[0029] Input the node feature update vector and the process flow feature update vector into the state transition function to generate a predicted state vector. The state transition function adopts a three-layer fully connected network. The dimension of the input layer is equal to the sum of the dimensions of the node feature update vector and the process flow feature update vector. The dimension of the hidden layer is 256, and the dimension of the output layer is equal to the dimension of the state vector. At the same time, the system collects the actual measurement data of the execution component through sensors, including the temperature, pressure, displacement, etc. of key points. The accuracy of the temperature sensor is ±0.5°C, the accuracy of the pressure sensor is ±0.01 MPa, and the accuracy of the displacement sensor is ±0.005 mm.

[0030] Construct a Kalman gain matrix for data fusion. The initial value of the state estimation covariance matrix is set to 0.1×identity matrix, the process noise covariance matrix is set to 0.01×identity matrix, and the measurement noise covariance matrix is set to 0.05×identity matrix. Through the Kalman filter algorithm, the system fuses the predicted state vector with the measurement data, corrects the prediction error, and generates a state correction vector. For example, the predicted value of the die head temperature is 188°C, the measured value is 190°C, and the corrected value obtained after fusion is 189.2°C, correcting the prediction deviation.

[0031] The system takes the state correction vector as the new state vector and loops through the above steps until the preset accuracy threshold is met (temperature error < 1°C, pressure error < 0.05 MPa, displacement error < 0.01 mm).

[0032] In this embodiment, unifying the modeling of structural dynamics, thermal field, and fluid field data to realize the interaction of temperature-stress-fluid can more realistically reflect the complex mechanism in the production process, which plays a key role in improving the simulation accuracy and prediction reliability; by using a feature encoder to encode the states of the execution component (position, speed, force, temperature field) and the process flow (heat flux, mass flow rate, stress tensor) into node and edge features respectively, and performing message passing fusion on the graph structure, the local and global spatio-temporal correlations can be captured, and the model's perception ability of local mutations and overall evolution can be enhanced; based on the thermo-mechanical coupling field equation and the fluid-structure interface model, calculating the temperature distribution and stress state in real time can adjust the simulation parameters for transient working conditions, so as to accurately reproduce the interface interaction and improve the prediction accuracy; using a gated recurrent unit (GRU) for temporal feature update and combining with the Kalman gain matrix to perform online correction on the predicted state can not only eliminate the cumulative error, but also adaptively adjust during the simulation process, improving the convergence speed and the final accuracy; the model is structured and process-oriented, adapting to execution components and complex process networks of different scales; at the same time, based on the combination of graph neural network and Kalman filter, it has good online simulation and fast response capabilities, meeting the real-time monitoring and prediction requirements of industrial sites.

[0033] In an alternative embodiment, a thermo-mechanical coupling field equation is constructed based on the node feature vector and the process flow feature vector, the temperature field distribution and the stress tensor are calculated, a fluid-structure interface coupling model is established, and the fluid-structure coupling data generated includes: The node feature vector is decomposed into a thermal field feature component and a force field feature component, and the process flow feature vector is decomposed into a flow feature component and a heat transfer feature component; Based on the thermal field feature component and the heat transfer feature component, a temperature field control equation is constructed. The temperature field control equation includes terms of material density, specific heat capacity, thermal conductivity, as well as an external heat source term and a deformation heat source term determined by the node feature vector, wherein the deformation heat source term is obtained through the coupling calculation of stress and strain rate; According to the force field feature component and the flow feature component, a thermal stress field equation is established, and the elastic strain term, the thermal expansion strain term and the viscous stress term determined by the process flow feature vector are coupled and calculated; A fluid-structure interface coupling model is established, the temperature gradients of the solid domain and the fluid domain are correlated through the interface heat flux continuity equation, and the stresses of the solid domain and the fluid domain are correlated through the interface stress balance equation; The temperature field control equation and the thermal stress field equation are combined into a coupled matrix equation, and the temperature field distribution and the stress tensor are obtained by using block iterative solution to generate the fluid-structure coupling data.

[0034] In a specific embodiment, during the cable manufacturing process, the node feature vector is decomposed into a thermal field feature component and a force field feature component. The specific implementation is to process the collected node data through a tensor decomposition algorithm. Taking a certain type of robot cable as an example, data points with node temperatures ranging from 120 °C to 180 °C and stress ranges from 15 MPa to 45 MPa are collected. The singular value decomposition method is applied to extract the principal components, and the node feature vector is decomposed into a thermal field feature component (characterizing thermal properties such as temperature and heat flux density) and a force field feature component (characterizing mechanical properties such as stress and strain) through feature mapping. At the same time, the process flow feature vector is decomposed into a flow feature component and a heat transfer feature component, where the flow feature component includes parameters such as flow velocity and pressure, and the heat transfer feature component includes parameters such as convective heat transfer coefficient and fluid temperature. In practical applications, for a copper core cable with a cross-sectional area of 2.5 mm 2 ², the typical values of the node thermal field feature components are [135, 0.45, 0.87], and the typical values of the force field feature components are [28, 0.12, 0.56], where each component represents the normalized representation of the temperature extreme value, the heat flux gradient and the stress extreme value.

[0035] Based on the thermal field feature component and the heat transfer feature component, a temperature field control equation is constructed. This control equation considers physical parameters such as material density, specific heat capacity and thermal conductivity. For example, the density of copper material is 8900 kg / m3 The specific heat capacity is 385 J / (kg·K), and the thermal conductivity is 401 W / (m·K). The external heat source term is determined by the thermal field component of the nodal eigenvector, and the deformation heat source term is obtained through the coupling calculation of stress and strain rate. In actual calculations, for the working condition of an extrusion temperature of 175 °C, the value of the external heat source term is approximately 4.8×10 5 W / m 3 , and the deformation heat source term varies with the strain rate, with a typical value of 3.2×10 4 W / m 3 to 5.7×10 4 W / m³. The control equation of the temperature field is discretized by the finite difference method. The mesh size is set to 0.5 mm, and the time step is 0.01 s to ensure the stability and accuracy of the numerical calculation.

[0036] The thermal stress field equation is established based on the force field characteristic component and the flow characteristic component. This equation couples and calculates the elastic strain term, the thermal expansion strain term, and the viscous stress term determined by the process flow eigenvector. For the cable material, the elastic modulus is 120 GPa, the Poisson's ratio is 0.34, and the thermal expansion coefficient is 17×10 -6 / K. In the thermal stress calculation, the thermal strain caused by the temperature gradient is considered, and the thermal expansion strain term is represented by the product of the temperature change and the thermal expansion coefficient. The viscous stress term is related to the fluid flow characteristics and is determined by the flow characteristic component of the process flow eigenvector. The typical shear stress value varies in the range of 0.8 MPa to 2.1 MPa. The thermal stress field equation is also solved by the finite element method. The element type is selected as a hexahedron eight-node element, and a total of 3200 elements are divided to ensure the calculation accuracy while taking into account the calculation efficiency.

[0037] A fluid-structure interface coupling model is established to effectively connect the solid domain and the fluid domain. The temperature gradients of the solid domain and the fluid domain are correlated through the interface heat flux continuity equation to ensure continuous heat transfer at the interface. This equation requires that the normal heat flux of the solid domain is equal to the normal heat flux of the fluid domain, and the difference between the two does not exceed 0.01 W / m 2 . The stresses in the solid domain and the fluid domain are correlated through the interface stress balance equation to ensure force balance at the interface. In the specific implementation, the alternating finite element method is used to handle the fluid-structure coupling problem. The fluid domain is solved by the computational fluid dynamics method, and the solid domain is solved by the finite element method. The two exchange data at the interface until convergence. The interface mesh division adopts a consistency matching strategy to ensure one-to-one correspondence of the nodes and avoid accuracy loss caused by data interpolation.

[0038] The temperature field control equation and the thermal stress field equation are combined to form a coupled matrix equation, and the temperature field distribution and stress tensor are obtained by using block iteration to solve. During the block iteration process, the temperature field is first solved, then the thermal stress field is calculated based on the temperature field results, and the deformation heat generated by the thermal stress field is fed back into the temperature field calculation, and so on, iterating back and forth until convergence. The convergence criterion is set as the relative error of the temperature field is less than 0.1% and the relative error of the stress field is less than 0.5%. In practical applications, for a cable production line with a manufacturing speed of 18 m / min, the iteration usually needs 5 to 8 times to converge. The finally generated fluid-structure interaction data includes the spatial distribution data of the temperature field and the stress tensor distribution data, and the data accuracy reaches ±1.5 °C and ±0.8 MPa, which can be directly used for process parameter optimization.

[0039] Traditional methods for optimizing cable manufacturing processes mainly rely on empirical models and simplified analyses, often treating thermal analysis and mechanical analysis separately and ignoring the interaction between the two, resulting in insufficient prediction accuracy. Existing technologies usually adopt a one-way thermal-structure coupling method, that is, first calculate the temperature field, and then use the temperature field as a boundary condition to calculate the stress field, which cannot capture the feedback effect of the stress field on the temperature field. The method of this embodiment extracts key physical features through the eigenvector decomposition method, establishes a two-way coupling model of the temperature field and the stress field, introduces a deformation heat source term to consider the influence of the stress field on the temperature field, and at the same time realizes the effective connection between the solid domain and the fluid domain through the fluid-structure interface coupling model. This enables the model to more accurately describe the physical phenomena during the cable manufacturing process, especially to capture the dynamic thermal-mechanical coupling effect under high-speed production conditions.

[0040] As Figure 2 shown, it shows the axial temperature change characteristics during the cable manufacturing process, intuitively presenting the accuracy advantage of this technical solution compared with the traditional one-way coupling model. The temperature curve predicted by this technical solution (blue line) almost completely coincides with the measured data (green dotted line), while there are obvious deviations in the traditional one-way coupling model (red line). Especially in the melt extrusion zone (at 5 - 10 m), this technical solution accurately captures the temperature peak reaching 365 °C, and the error compared with the measured value of 364 °C is only 0.27%, while the predicted value of the traditional model is 348 °C, with an error as high as 4.40%. After entering the cooling zone (at 15 m), the predicted temperature by this technical solution is 292 °C, and the measured value is 294 °C, with an error of only 0.68%, while the traditional model predicts 273 °C, with an error of 7.14%. This significant improvement is due to the fact that this technical solution considers the two-way coupling mechanism between the deformation heat source term and the temperature field, successfully capturing the dynamic thermal-mechanical coupling effect under high-speed production conditions, especially performing well in the area where the temperature gradient changes violently.

[0041] In an alternative embodiment, in the reinforcement learning training environment, analyze the causal dependencies of process parameter adjustment instructions, establish a parameter adjustment dependency graph, and generating an initial process control strategy based on the parameter adjustment dependency graph includes: Collect historical data of process parameter adjustment instructions, establish the joint probability distribution and marginal probability distribution between process parameter adjustment instructions, and calculate the mutual information and partial correlation coefficient between process parameter adjustment instructions using the joint probability distribution and marginal probability distribution; Construct a process parameter adjustment dependency graph based on the mutual information and partial correlation coefficient, and assign weight values to the causal dependencies in the process parameter adjustment dependency graph using the sensitivity analysis results of process parameter adjustment instructions; Form a state vector with the current value and change rate of the process parameter adjustment instruction, and construct a process control strategy according to the hierarchical structure of the process parameter adjustment dependency graph, so that the results of the upper-level process parameter adjustment instructions guide the decision-making of the lower-level process parameter adjustment instructions; Import the process control strategy into the reinforcement learning training environment for testing, record the state transition results and causal effects of process parameter adjustment instructions, and construct a process control strategy evaluation function based on the causal effects and operation costs; Use the process control strategy evaluation function to calculate the stability index of the state vector, and optimize the process control strategy by combining the evaluation results under the perturbation of process parameter adjustment instructions to generate an initial process control strategy.

[0042] In a specific embodiment, to collect historical data of process parameter adjustment instructions, the process parameter adjustment instruction sequence can be continuously recorded in the production environment. For example, for the semiconductor wafer manufacturing process, the system records the adjustment data of 12 parameters such as temperature adjustment instructions, pressure adjustment instructions, and gas flow adjustment instructions every 5 minutes within 30 days, and a total of 8640 sets of adjustment instruction data are collected. Based on these historical data, the system calculates the joint probability distribution and marginal probability distribution between each process parameter adjustment instruction. Specifically, the system divides the joint occurrence frequency of the temperature adjustment instruction T and the pressure adjustment instruction P by the total number of samples to obtain the joint probability P(T, P), and calculates their respective marginal probabilities P(T) and P(P). Subsequently, the system calculates the mutual information through the logarithmic ratio of the joint probability and the marginal probability. For example, the mutual information between temperature T and pressure P can be obtained through a specific calculation of their joint and marginal probabilities, and the value is 0.76, indicating a strong information correlation between these two parameter adjustment instructions. At the same time, the system calculates the partial correlation coefficient. After excluding the influence of other parameters, the partial correlation coefficient between temperature and pressure is 0.65.

[0043] Based on the calculated mutual information and partial correlation coefficient, the system constructs a process parameter adjustment dependency graph. In the dependency graph, nodes represent process parameter adjustment instructions, and edges represent the dependency relationships between parameters. The system sets the mutual information threshold to 0.5. When the mutual information between two parameters exceeds this threshold, a connection is established in the dependency graph. For the semiconductor wafer manufacturing environment, the dependency graph constructed by the system shows that the temperature adjustment instruction has dependency relationships with five other parameters, and the gas flow rate instruction has dependency relationships with four parameters. Further, the system determines the weights of the dependency relationships through sensitivity analysis. The sensitivity analysis includes perturbing each parameter by ±5% and recording the change amplitudes of other parameters. The results show that the influence weight of temperature adjustment on the pressure parameter is 0.82, and the influence weight on the gas flow rate parameter is 0.67. These weight values are assigned to the corresponding edges of the dependency graph to form a weighted process parameter adjustment dependency graph.

[0044] Using the constructed process parameter adjustment dependency graph, the system forms a state vector from the current values and their change rates of each process parameter adjustment instruction. For example, the state vector of the temperature parameter is (T, ΔT / Δt), where T is the current temperature value and ΔT / Δt is the temperature change rate. According to the hierarchical structure of the dependency graph, the system identifies that the temperature adjustment instruction is at the top layer, the pressure and gas flow rate adjustment instructions are at the second layer, and the remaining parameters are at lower layers. Based on this, a hierarchical process control strategy is constructed: when the temperature parameter needs to be adjusted, the system first calculates the target value and rate of temperature adjustment, and then determines the adjustment directions and amplitudes of these parameters according to the dependency relationships between temperature, pressure, and gas flow rate. In practical applications, when the temperature needs to be increased from 250°C to 270°C, the system determines based on the dependency relationship that the pressure needs to be adjusted from 5.2 bar to 5.6 bar, and the gas flow rate needs to be adjusted from 120 sccm to 135 sccm.

[0045] The constructed process control strategy is imported into the reinforcement learning training environment for testing. The training environment simulates the actual production scenario. The system records 1000 parameter adjustment operations, the state transition results and causal effects of the parameters after each operation. An evaluation function is constructed based on the causal effect and operation cost. The evaluation function considers the accuracy, stabilization speed, and energy consumption of parameter adjustment. Specifically, when the temperature adjustment accuracy is within the range of ±1°C, the stabilization time is less than 60 seconds, and the increase in energy consumption does not exceed 3%, the evaluation score is 92 points (out of 100).

[0046] The stability index of the state vector is calculated using an evaluation function. For the state vector composed of 12 parameters, the calculated stability index is 0.85. Random perturbations are introduced to the process parameters to observe the robustness of the control strategy. After introducing a random perturbation of ±3°C to the temperature parameter, the temperature can still be adjusted to near the target value, but the stabilization time increases to 65 seconds and the evaluation score drops to 85 points. Based on the results of the perturbation test, the parameter adjustment rate and adjustment order in the control strategy are optimized. The adjustment rate of the temperature parameter is adjusted from the original 2°C / minute to 1.8°C / minute, and the cooperative control logic between parameters is adjusted to better handle perturbations. Under the same perturbation conditions, the stabilization time of the optimized control strategy is reduced to 52 seconds and the evaluation score is increased to 90 points, successfully generating an initial process control strategy with strong robustness.

[0047] In an alternative embodiment, the current value and change rate of the process parameter adjustment instruction are combined to form a state vector, and a process control strategy is constructed according to the hierarchical structure of the process parameter adjustment dependency graph. The result of the upper-layer process parameter adjustment instruction guiding the decision-making of the lower-layer process parameter adjustment instruction includes: The current value and change rate of the process parameter adjustment instruction are combined to form an initial state vector, the mean and standard deviation of the process parameter adjustment instruction are calculated, and the initial state vector is dynamically normalized to generate a normalized state vector; Based on the process parameter adjustment dependency graph, the path weights between adjacent nodes are calculated, the path weights are substituted into the node level determination function, the level value of each node corresponding to the process parameter adjustment instruction is calculated, and the association strength between the nodes within the layer is calculated according to the level value and the path weights; Taking the normalized state vector as the input, a multi-layer conditional strategy function is constructed according to the level value, the level reward value of the upper-layer process parameter adjustment instruction is used as the conditional constraint for the lower-layer process parameter adjustment instruction, and the parameters of the inter-layer coupling constraint function are determined through the association strength; Based on the inter-layer coupling constraint function, a global reward function is constructed, a local reward function is constructed according to the normalized state vector, the global reward function and the local reward function are combined to generate a level reward function, and the level reward function is weighted using a level weight coefficient to construct a multi-level strategy evaluation function; The decision sequence of the upper-layer process parameter adjustment instruction is recorded, the decision sequence is input into the level decision function to generate a candidate decision set for the lower-layer process parameter adjustment instruction, the change amount of the multi-level strategy evaluation function under the candidate decision is calculated, the feedback adjustment factor is determined according to the change amount, and the output of the level decision function is dynamically optimized to generate a hierarchical process control strategy.

[0048] In a specific implementation, the current values and change rates of process parameter adjustment instructions such as extrusion temperature, cooling rate, and tension control are collected during the manufacturing process of the cable. The mean and standard deviation of the instruction sequence of each process parameter are calculated. Taking the extrusion temperature as an example, the current temperature value is 185°C, the temperature value at the previous moment is 183°C, and the change rate is 2°C per time unit. The current value and the change rate are combined to form an initial state vector [185, 2]. The mean of the historical data of this parameter is 180°C, and the standard deviation is 5°C. The standardized state vector is [(185 - 180) / 5, 2 / 1.5] = [1, 1.33]. Similar processing is performed on all process parameters to generate a complete standardized state vector.

[0049] A directed dependency graph is established based on the physical correlations between parameters during the cable manufacturing process. For example, the extrusion temperature affects the melt fluidity, which in turn affects the wire diameter stability. By analyzing the process data, the path weights between parameter nodes are calculated. The influence weight of the extrusion temperature on the wire diameter is 0.7, and its influence on the insulation strength is 0.5. These weights are substituted into the hierarchical determination function. The hierarchical determination function uses the weighted ratio of the in-degree and out-degree of the node to determine the hierarchy. For example, the in-degree of the extrusion temperature is 1 (only controlled by the equipment), and the weighted out-degree is 2.3 (affecting multiple downstream parameters). The calculated hierarchy value is 2.3, belonging to the upper-level control parameter. The hierarchy value of the cooling rate is 1.8, and the hierarchy value of the tension control is 1.2. Among the parameters in the same layer, such as the extrusion temperature and the extrusion pressure, both are upper-level parameters. The correlation strength is determined by calculating their common influencing targets. Their common influence on the wire diameter stability is 0.65, and the set correlation strength is 0.65.

[0050] Using the standardized state vector as the input, a hierarchical strategy is constructed based on the calculated hierarchy values. For parameters with a hierarchy value greater than 2.0 (such as the extrusion temperature), an upper-level strategy function is constructed. This function calculates the parameter adjustment direction and amplitude based on the current state vector. When the standardized extrusion temperature state is [1, 1.33], the upper-level strategy function may output an adjustment instruction to increase by 0.2°C. The parameters of the inter-layer coupling constraint function are determined by the correlation strength. The correlation strength between the extrusion temperature and the cooling rate is 0.6. The conditional strategy function of the cooling rate will use 0.6 as the coupling coefficient to calculate the co-variation of the cooling rate based on the adjustment of the extrusion temperature.

[0051] The global reward function is based on the final product quality metrics (such as cable insulation strength, conductor resistivity, etc.), while the local reward function is based on the stability of process parameters and the deviation from the target. Taking wire diameter stability as an example, the target wire diameter is 2.5mm, and the wire diameter value in the standardized state vector is [0.8, 0.1] (indicating that the current wire diameter is 0.8 times the standard deviation and the change rate is 0.1 times the standard deviation), and the calculated local reward value is 0.85. The global reward and the local reward are combined to form a hierarchical reward function. The hierarchical reward weight of the upper-layer parameter (extrusion temperature) is 0.7, the weight of the middle-layer parameter (cooling rate) is 0.5, and the weight of the lower-layer parameter (tension control) is 0.3. The hierarchical rewards are weighted to constitute a multi-level policy evaluation function.

[0052] Record the historical decision sequence of the upper-layer process parameters (such as extrusion temperature), such as [increase by 0.5°C, maintain, decrease by 0.3°C]. Input this decision sequence into the hierarchical decision function to generate a candidate decision set for the lower-layer parameters (such as cooling rate), including multiple candidate adjustment options such as [increase by 5%, increase by 2%, maintain, decrease by 3%]. For each candidate decision, calculate the change amount of the multi-level policy evaluation function. Increasing the cooling rate by 5% results in an increase of 0.12 in the evaluation function, and increasing it by 2% results in an increase of 0.15 in the evaluation function. Select the latter as the optimization decision. Calculate the feedback adjustment factor based on the actual execution effect. For example, if it is expected to increase the temperature by 0.5°C to improve wire diameter stability, but the actual effect is lower than expected, calculate the adjustment factor of 0.8 to adjust the amplitude of the next decision. Continuously improve the effectiveness of the hierarchical process control strategy by dynamically optimizing the hierarchical decision function.

[0053] In robotic automated cable manufacturing, the current value of the extrusion temperature is collected as 192°C, with a change rate of 1°C per time unit, the cooling rate is 60%, and the change rate is -2% per time unit. The average extrusion temperature is 190°C, the standard deviation is 4°C, the average cooling rate is 65%, the standard deviation is 7%, and the standardized state vector is [0.5, 0.25, -0.71, -0.29]. According to the dependency graph, the hierarchical value of the extrusion temperature is calculated as 2.4, and the hierarchical value of the cooling rate is 1.8, determining that the extrusion temperature is the upper-layer control parameter. When uneven cable insulation thickness is detected, the calculation result of the multi-level policy evaluation function shows that the combined decision of adjusting the extrusion temperature to 194°C and reducing the cooling rate to 58% can obtain the highest reward value of 0.87. Generate the upper-layer parameter decision "increase the temperature by 2°C", generate a candidate decision set for the lower-layer parameters through the hierarchical decision function, and calculate that the evaluation function increases by 0.15 when the cooling rate is reduced by 2%. Finally, generate a coordinated adjustment strategy to achieve a 15% improvement in insulation thickness uniformity.

[0054] Existing methods for optimizing cable manufacturing processes mainly rely on the independent control of single parameters and cannot effectively handle the complex interactions between parameters. Traditional methods usually adopt rule-based control strategies. For example, when the cable diameter is too large, the extrusion speed is reduced, but it is difficult to cope with multi-parameter coupling scenarios. The method of this embodiment solves the parameter coupling problem in traditional methods by constructing a dependency graph between process parameters, processing the parameters in layers, and using a deep reinforcement learning framework to calculate the optimal adjustment strategy. In traditional methods, adjusting one parameter may cause other parameters to get out of control, while this method realizes the collaborative optimization between parameters through a hierarchical reward function and a coupling constraint function.

[0055] As Figure 3 shown, it demonstrates the significant differences between this technical solution and multiple traditional methods in the control of key cable manufacturing parameters. The chart records the real-time fluctuations of three key parameters (extrusion temperature, traction speed, and cooling water flow rate) during a 500-minute production process. The extrusion temperature curve under this technical solution shows that starting from the initial 192.3°C, it rises steadily to 205.5°C after about 120 minutes and remains stable, with the fluctuation range controlled within ±0.1°C. In contrast, the temperature under single-parameter independent control fluctuates periodically, reaching a maximum of 213.2°C and a minimum of 190.5°C, with an oscillation amplitude exceeding 22°C. Similarly, the traction speed of this technical solution is adjusted smoothly from 2.12 m / s to 2.97 m / s, while the rule-based control strategy fluctuates repeatedly between 2.05 - 3.25 m / s; the cooling water flow rate parameter also shows the same trend, with this technical solution stabilizing at 5.75 L / s, and the PID control method oscillating between 4.10 - 6.42 L / s. This significant improvement in stability fully proves that this technical solution can effectively suppress the negative interactions between parameters through hierarchical decision-making and coupling constraint mechanisms, achieving global stable control.

[0056] In an alternative embodiment, an industrial knowledge graph expert system in the form of triples is constructed, and the initial process optimization strategy is fused with the expert rule base based on credibility weights. The improved process optimization strategy is generated through a hybrid intelligent method that combines symbolic reasoning chains and deep learning feature extraction, including: Construct an industrial knowledge graph expert system, represent expert experience rules in the form of triples, and each triple contains process parameter nodes, process parameter relationships, and process control rules to form an expert rule base of the industrial knowledge graph; Calculate the matching degree between the process parameter configuration in the initial process optimization strategy and the process control rules in the expert rule base, assign credibility weights to each process control rule based on the matching degree, and establish a priority ranking for the process control rules; Construct a symbolic inference chain according to the priority order, input the process parameter configuration of the initial process optimization strategy into the symbolic inference chain, judge the activation state of the process control rules based on the credibility weight, and output the symbolic inference result; Input the process parameter nodes and process parameter relationships into a deep learning model for feature extraction to obtain process optimization feature vectors, combine the process optimization feature vectors with the symbolic inference results to construct a hybrid strategy network, and output candidate process optimization solutions; Construct an evaluation function that includes rule compliance and optimization effect, calculate the evaluation score of the candidate process optimization solution under the evaluation function, and adjust the fusion weight of the hybrid strategy network according to the evaluation score to generate an improved process optimization strategy.

[0057] In a specific implementation, represent the expert experience rules in the form of triples, where each triple includes a process parameter node, a process parameter relationship, and a process control rule. In the specific implementation process, for a certain iron and steel smelting process, the expert rule "when the temperature exceeds 1200 °C and the carbon content is lower than 0.5%, the carbon powder addition amount needs to be increased by 0.2 kg" can be represented as multiple triples: (temperature, numerical relationship, 1200 °C), (carbon content, numerical relationship, 0.5%), (condition combination, trigger rule, increase carbon powder), (carbon powder addition, adjustment amount, 0.2 kg). These triples form a connected graph structure through relationships to form an expert rule base of the industrial knowledge graph. Use a graph database to store these triples for efficient query and reasoning.

[0058] Calculate the matching degree between the process parameter configuration in the initial process optimization strategy and the process control rules in the expert rule base. In practical applications, taking the heat treatment process as an example, assume that the initial process parameter configuration is: holding temperature 950 °C, holding time 2 hours, and cooling rate 15 °C / minute. There is a rule in the expert rule base: "For alloy steel heat treatment, the holding temperature should be between 900 - 1000 °C, the holding time should be between 1.5 - 2.5 hours, and the cooling rate should not exceed 20 °C / minute". By calculating the matching degree, the matching degree of the holding temperature is 100% (950 °C is within the range of 900 - 1000 °C), the matching degree of the holding time is 100% (2 hours is within the range of 1.5 - 2.5 hours), and the matching degree of the cooling rate is 100% (15 °C / minute is less than 20 °C / minute). The overall matching degree of this rule is 100%, so a higher credibility weight of 0.95 is assigned. For a rule with a lower matching degree, such as a rule suggesting "the holding temperature should be 1050 °C", which does not match the initial configuration, a lower credibility weight of 0.3 is assigned. By calculating the matching degree of all rules and assigning weights, establish a priority ranking of the process control rules.

[0059] Construct a symbolic inference chain according to the priority order, and input the process parameter configuration of the initial process optimization strategy into the symbolic inference chain. In the case of the aluminum alloy extrusion process, the initial process parameters are: extrusion temperature 480 °C, extrusion speed 10 mm / s, extrusion ratio 20:1. The inference chain checks the rules one by one according to the priority from high to low. The high-priority rule "the extrusion temperature should be maintained between 450 - 500 °C" is satisfied, so the activation status is "satisfied"; the medium-priority rule "the product of the extrusion speed and the extrusion ratio should not exceed 300" calculates the result as 10×20 = 200, which is less than 300, so the activation status is "satisfied"; the low-priority rule "the extrusion temperature should be lower than 470 °C" is not satisfied, but the confidence weight of this rule is only 0.4, which is lower than the threshold of 0.6, so it is not adopted. The symbolic inference result is: the current process parameter configuration is basically reasonable, but the extrusion temperature can be considered to be appropriately reduced to 470 °C to improve the product quality.

[0060] Input the process parameter nodes and process parameter relationships into the deep learning model for feature extraction. In the implementation process, a graph neural network is used to process the process parameter relationship graph. For the injection molding process, the process parameter nodes include: mold temperature 80 °C, injection pressure 120 MPa, holding pressure time 6 s, cooling time 20 s, etc. The graph neural network extracts a 28-dimensional process optimization feature vector [0.72, 0.45, 0.89,..., 0.56] by aggregating neighbor node information. Combine these feature vectors with the symbolic inference results to construct a hybrid strategy network. The input layer of the network receives the encoding of the feature vectors and the symbolic inference results, processes them through the fully connected layer and the activation function, and outputs the candidate process optimization plan: adjust the mold temperature to 85 °C, maintain the injection pressure at 120 MPa, extend the holding pressure time to 8 s, and shorten the cooling time to 18 s.

[0061] Construct an evaluation function, which includes two parts of indicators: rule compliance and optimization effect. The rule compliance calculates the degree of compliance of the candidate solution with the expert rules, and the optimization effect evaluates the improvement of product quality after adjusting process parameters. In the case of the ceramic sintering process, the candidate optimization solution is: sintering temperature 1150°C, holding time 3 hours, heating rate 2°C / minute, and cooling rate 1.5°C / minute. The calculated result of rule compliance is 0.92 (highly compliant with expert rules), and the predicted value of the optimization effect is 0.85 (significantly improving product density and strength). The comprehensive evaluation score is 0.892, which is higher than the preset threshold of 0.8. Therefore, this candidate solution is accepted. Based on the evaluation score, adjust the fusion weights of the hybrid strategy network, increase the weight of the symbolic reasoning result on the temperature parameter to 0.7, and reduce the weight of deep learning on the cooling parameter to 0.3, generating an improved process optimization strategy: sintering temperature 1140°C, holding time 3 hours, heating rate 2°C / minute, and cooling rate 1.7°C / minute. Practical applications have proved that the improved strategy has increased the product qualification rate from 92% to 97.5% and reduced energy consumption by 8.3%, fully demonstrating the advantages of the hybrid intelligent method in industrial process optimization.

[0062] In an alternative embodiment, calculate the matching degree between the process parameter configuration in the initial process optimization strategy and the process control rules in the expert rule base, assign a credibility weight to each process control rule based on the matching degree, and establish a priority ranking of process control rules, including: Construct a semantic feature matrix and a structural feature matrix of the process parameter configuration, input the semantic feature matrix into a bidirectional attention network to calculate the semantic association strength between parameters, and input the structural feature matrix into a graph convolutional network to extract the topological relationship features between parameters; Adopt a hierarchical attention mechanism to calculate the matching scores between the process parameter configuration and the process control rules at different abstraction levels, and combine the semantic association strength and topological relationship features to construct a rule matching degree evaluation model; Based on the preset rule adaptability evaluation index, input the rule matching degree and the rule adaptability index into a dynamic weighting function to calculate the comprehensive credibility of the process control rules; Adopt a multi-objective optimization algorithm, combine the rule conflict degree and the rule coverage degree, normalize the comprehensive credibility of the process control rules, and establish a priority ranking of the process control rules.

[0063] In a specific implementation, constructing the semantic feature matrix and the structural feature matrix of process parameter configuration is the basis of the optimization process. For the construction of the semantic feature matrix, the system first collects process parameters related to cable manufacturing, including insulation material type, wire diameter specification, insulation layer thickness, processing temperature, cooling time, etc. For each parameter, the system uses a pre-trained word vector model in the domain knowledge base to convert the parameter description into a 300-dimensional vector representation. For example, when processing the parameter "hot pressing forming of XLPE insulated cable", the system decomposes it into keywords such as "XLPE", "insulation", "cable", "hot pressing", "forming", etc., obtains their word vectors respectively, and then obtains the semantic vector of this parameter through weighted average. For all n process parameters, an n×300-dimensional semantic feature matrix is formed. In practical applications, when processing 50 process parameters, the generated semantic feature matrix has a dimension of 50×300.

[0064] During the construction of the structural feature matrix, the system analyzes the dependency relationships between process parameters and establishes a parameter association graph. For example, there is a direct association between wire diameter specification and insulation layer thickness, and there may be an indirect association with cooling time. The system analyzes the process flow to establish an association strength value for each pair of parameters and constructs an n×n adjacency matrix. When n is 50, a 50×50 structural feature matrix is generated. In the matrix, a value of 1 indicates a direct association between parameters, and a value of 0 indicates no direct association. For example, the association value between "heating temperature" and "holding time" is 1, while the association value with "packaging material" may be 0.

[0065] The semantic association strength is calculated using a bidirectional attention network, and the semantic feature matrix is input into this network. The network contains a multi-head self-attention mechanism with 8 attention heads, and the hidden layer dimension of each head is 64. By calculating the attention weights between parameters, an n×n semantic association strength matrix is obtained, where the element values range from 0 to 1, indicating the semantic similarity between parameters. For example, "hot pressing forming temperature" and "material melting point" may obtain a high semantic association strength of 0.85, while the association strength with "packaging size" may only be 0.12.

[0066] The topological relationship feature extraction uses a graph convolutional network with the structural feature matrix as the input. The network contains 3 graph convolutional layers with hidden layer dimensions of 128, 64, and 32 respectively. Through the message passing mechanism, each parameter node aggregates information from adjacent nodes to capture the high-order topological relationships between parameters. The output is a 32-dimensional topological feature vector for each parameter, forming an n×32 topological feature matrix. In practical applications, these features can represent the structural importance of parameters in the entire process flow chart. For example, some parameters may be key nodes connecting multiple process steps.

[0067] The hierarchical attention mechanism is used to calculate the matching score between the process parameter configuration and the process control rules. This mechanism consists of three levels: parameter-level attention, rule-level attention, and global attention. The parameter-level attention calculates the contribution of each parameter to the rule matching, and normalizes the attention weights using the softmax function; the rule-level attention calculates the importance of different control rules; the global attention integrates the results of the first two levels to obtain the final matching score. For the rule of "when the wire diameter is greater than 2.5 mm and the insulating material is PVC, the hot pressing temperature should be controlled at 165 ± 5 °C" in wire and cable manufacturing, the system will calculate the attention weights of parameters such as "wire diameter", "insulating material", and "hot pressing temperature" respectively, for example, they may be 0.3, 0.35, and 0.35 respectively.

[0068] The rule matching degree evaluation model combines the semantic association strength and topological relationship features. The model adopts a multi-layer perceptron structure, the input layer dimension is n×n + n×32, it contains two hidden layers with dimensions of 256 and 128 respectively, and the output layer is the rule matching degree score. During the model training process, the successful cases in the historical process data are used as positive samples, and the failed cases are used as negative samples to optimize the model parameters through the cross-entropy loss function. In practical applications, when inputting a specific process parameter configuration such as "wire diameter 2.8 mm, PVC insulation, hot pressing temperature 170 °C", for the above rule, the model may output a high matching degree of 0.92.

[0069] The rule adaptability evaluation indicators include three dimensions: rule precision, rule coverage, and rule stability. The rule precision verifies the applicability of the rule through historical data; the rule coverage measures the coverage range of the rule for the process parameter space; the rule stability evaluates the robustness of the rule under small parameter changes. For example, for the temperature control rule, the system analyzes the proportion of successful applications of this rule in historical data to calculate the precision, and may obtain 0.95; analyzes the proportion of process parameter combinations covered by this rule in the total parameter space to obtain the coverage rate, which may be 0.75; obtains the stability score of 0.88 through parameter perturbation testing.

[0070] The comprehensive credibility calculation adopts a dynamic weighting function, which adaptively adjusts the weights of each indicator according to the current process state. The function uses the softmax form, the input is the rule matching degree and the rule adaptability indicators, and the output is the comprehensive credibility score. At the starting stage of the wire and cable manufacturing process, the system may attach more importance to the rule precision and give a weight of 0.5; during the stable production stage, it attaches more importance to the rule stability and adjusts its weight to 0.6. In an application example, the matching degree of a certain temperature control rule is 0.92, the precision is 0.95, the coverage rate is 0.75, and the stability is 0.88. After dynamic weighting calculation, the comprehensive credibility is 0.89.

[0071] The priority sorting of process control rules adopts a multi-objective optimization algorithm, combining rule conflict degree and rule coverage. The rule conflict degree is obtained through the analysis of semantic similarity and target consistency between rules; the rule coverage represents the coverage range of rules for the process parameter space. The system uses the NSGA-II multi-objective optimization algorithm, sets the population size to 100, and the number of evolutionary generations to 50, and obtains the normalized result of the comprehensive credibility. When two rules conflict, such as "the hot pressing temperature is controlled at 165±5°C" and "the hot pressing temperature should be greater than 170°C", the system calculates their conflict degree to be 0.9. Considering their respective credibilities comprehensively, the former is 0.89 and the latter is 0.72, and finally determines that the priority of the former is higher than that of the latter.

[0072] Compared with the prior art, traditional optimization methods for robot cable manufacturing processes mainly rely on expert experience and fixed rule bases, lacking in-depth analysis of the complex correlations between parameters. Common methods such as fuzzy logic control and case-based reasoning methods cannot effectively handle optimization problems in high-dimensional parameter spaces. In the prior art, rule priorities are usually determined by expert scoring or simple weighting methods, making it difficult to adapt to dynamic production environments. The method of this embodiment introduces a dual-channel feature extraction mechanism that combines semantic features and structural features, captures the complex correlations between parameters through deep learning technology; adopts a hierarchical attention mechanism to achieve multi-level rule matching and evaluation; introduces a comprehensive credibility calculation method with dynamic weighting to improve the adaptability of rule application; and processes rule conflicts through multi-objective optimization.

[0073] The robot cable manufacturing process optimization system based on deep reinforcement learning in the embodiments of the present invention includes: The first unit is used to collect cable manufacturing data; construct a digital twin model, construct the execution components of manufacturing equipment as graph network nodes, construct the process flow between components as graph network edges, and obtain manufacturing process simulation data through the graph dynamics network and state iterative optimization of thermal coupling and fluid-structure coupling; The second unit is used to establish a deep reinforcement learning environment by taking the cable manufacturing data and manufacturing process simulation data as state inputs and the process parameter adjustment instruction as the action output; The third unit is used to analyze the causal dependency relationship of the process parameter adjustment instruction in the reinforcement learning training environment, establish a parameter adjustment dependency graph, and generate an initial process control strategy based on the parameter adjustment dependency graph; The fourth unit is used to construct an industrial knowledge graph expert system in the form of triples, fuse the initial process optimization strategy with the expert rule base based on the credibility weight, and generate an improved process optimization strategy through a hybrid intelligent method that combines symbolic reasoning chains and deep learning feature extraction; The fifth unit is used to adjust the process parameters of the robot cable manufacturing equipment according to the improved process optimization strategy, and feed back the cable manufacturing data collected after adjustment to the digital twin model to form a closed loop for process optimization.

[0074] In a third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0075] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0076] The present invention may be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions for performing various aspects of the present invention loaded thereon.

[0077] 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 foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the robot cable manufacturing process based on deep reinforcement learning, characterized in that, Including: Collecting cable manufacturing data; Constructing a digital twin model, constructing the execution components of manufacturing equipment as graph network nodes, constructing the process flows between components as graph network edges, and obtaining manufacturing process simulation data through graph dynamics network and state iterative optimization of thermal coupling and fluid-structure coupling; Taking the cable manufacturing data and manufacturing process simulation data as state inputs and the process parameter adjustment instructions as action outputs to establish a deep reinforcement learning environment; In the reinforcement learning training environment, analyzing the causal dependence relationship of the process parameter adjustment instructions, establishing a parameter adjustment dependence graph, and generating an initial process control strategy based on the parameter adjustment dependence graph; Constructing an industrial knowledge graph expert system in the form of triples, fusing the initial process optimization strategy with the expert rule base based on credibility weights, and generating an improved process optimization strategy through a hybrid intelligent method combining symbolic reasoning chain and deep learning feature extraction; Adjusting the process parameters of the robotic cable manufacturing equipment according to the improved process optimization strategy, and feeding back the cable manufacturing data collected after adjustment to the digital twin model to form a process optimization closed loop.

2. The method according to claim 1, characterized in that Constructing a digital twin model, constructing the execution components of manufacturing equipment as graph network nodes, constructing the process flows between components as graph network edges, and obtaining manufacturing process simulation data through graph dynamics network and state iterative optimization of thermal coupling and fluid-structure coupling, including: Composing the position vector, velocity vector, force vector, and temperature field distribution of the execution components of the cable manufacturing equipment into a state vector, and generating a node feature vector through a feature encoder; Collecting the heat flux density, mass flow rate, and stress tensor of the process flow between the execution components, and composing a process flow feature vector; Constructing a thermal coupling field equation based on the node feature vector and the process flow feature vector, calculating the temperature field distribution and stress tensor, establishing a fluid-structure interface coupling model, and generating fluid-structure coupling data; Inputting the node feature vector and the fluid-structure coupling data into a gated recurrent unit, and fusing adjacent node information using a message passing matrix to generate a node feature update vector; Calculating the change of edge attributes according to the node feature update vector and the process flow feature vector to generate a process flow feature update vector; Inputting the node feature update vector and the process flow feature update vector into a state transfer function to generate a predicted state vector, collecting measurement data and performing error correction through a Kalman gain matrix to generate a state correction vector; Taking the state correction vector as a new state vector and executing it in a loop until the simulation accuracy meets the preset accuracy threshold.

3. The method according to claim 2, wherein Constructing a thermal coupling field equation based on the node feature vector and the process flow feature vector, calculating the temperature field distribution and stress tensor, establishing a fluid-structure interface coupling model, and generating fluid-structure coupling data, including: Decomposing the node feature vector into a thermal field feature component and a force field feature component, and decomposing the process flow feature vector into a flow feature component and a heat transfer feature component; Construct a temperature field control equation based on the thermal field characteristic components and the heat transfer characteristic components. The temperature field control equation includes terms of material density, specific heat capacity, thermal conductivity, as well as an external heat source term and a deformation heat source term determined by the node characteristic vector, where the deformation heat source term is obtained through the coupling calculation of stress and strain rate; Establish a thermal stress field equation according to the force field characteristic components and the flow characteristic components, and perform a coupling calculation on the elastic strain term, the thermal expansion strain term, and the viscous stress term determined by the process flow characteristic vector; Establish a fluid-structure interface coupling model, associate the temperature gradients of the solid domain and the fluid domain through the interface heat flux continuity equation, and associate the stresses of the solid domain and the fluid domain through the interface stress balance equation; Form a coupled matrix equation from the temperature field control equation and the thermal stress field equation, and use block iterative solution to obtain the temperature field distribution and stress tensor, generating fluid-structure coupling data.

4. The method according to claim 1, wherein In the reinforcement learning training environment, analyze the causal dependence relationship of process parameter adjustment instructions, establish a parameter adjustment dependence graph, and generate an initial process control strategy based on the parameter adjustment dependence graph, including: Collect historical data of process parameter adjustment instructions, establish the joint probability distribution and marginal probability distribution between process parameter adjustment instructions, and calculate the mutual information quantity and partial correlation coefficient between process parameter adjustment instructions using the joint probability distribution and marginal probability distribution; Construct a process parameter adjustment dependence graph based on the mutual information quantity and partial correlation coefficient, and assign weight values to the causal dependence relationships in the process parameter adjustment dependence graph using the sensitivity analysis results of process parameter adjustment instructions; Form a state vector from the current value and change rate of the process parameter adjustment instruction, and construct a process control strategy according to the hierarchical structure of the process parameter adjustment dependence graph, so that the results of the upper-level process parameter adjustment instructions guide the decision-making of the lower-level process parameter adjustment instructions; Import the process control strategy into the reinforcement learning training environment for testing, record the state transition results and causal effects of the process parameter adjustment instructions, and construct a process control strategy evaluation function based on the causal effects and operation costs; Calculate the stability index of the state vector using the process control strategy evaluation function, and optimize the process control strategy in combination with the evaluation results under the perturbation of the process parameter adjustment instructions to generate an initial process control strategy.

5. The method according to claim 4, wherein Form a state vector from the current value and change rate of the process parameter adjustment instruction, and construct a process control strategy according to the hierarchical structure of the process parameter adjustment dependence graph, so that the results of the upper-level process parameter adjustment instructions guide the decision-making of the lower-level process parameter adjustment instructions, including: Form an initial state vector from the current value and change rate of the process parameter adjustment instruction, calculate the mean and standard deviation of the process parameter adjustment instruction, and perform dynamic normalization processing on the initial state vector to generate a normalized state vector; Calculate the path weights of adjacent nodes based on the process parameter adjustment dependence graph, substitute the path weights into the node level determination function, calculate the level values of the nodes corresponding to each process parameter adjustment instruction, and calculate the association strength between the nodes within the layer according to the level values and path weights; Taking the standardized state vector as the input, construct a multi-layer conditional policy function according to the hierarchical value, use the hierarchical reward value of the upper-layer process parameter adjustment instruction as the conditional constraint of the lower-layer process parameter adjustment instruction, and determine the parameters of the inter-layer coupling constraint function through the association strength; Construct a global reward function based on the inter-layer coupling constraint function, construct a local reward function according to the standardized state vector, combine the global reward function and the local reward function to generate a hierarchical reward function, and weight the hierarchical reward function with the hierarchical weight coefficient to construct a multi-level policy evaluation function; Record the decision sequence of the upper-layer process parameter adjustment instruction, input the decision sequence into the hierarchical decision function, generate a candidate decision set of the lower-layer process parameter adjustment instruction, calculate the change amount of the multi-level policy evaluation function under the candidate decision, determine the feedback adjustment factor according to the change amount and dynamically optimize the output of the hierarchical decision function to generate a hierarchical process control strategy.

6. The method according to claim 1, characterized in that, Construct an industrial knowledge graph expert system in the form of triples, fuse the initial process optimization strategy with the expert rule base based on the credibility weight, and generate an improved process optimization strategy through a hybrid intelligent method combining symbolic reasoning chain and deep learning feature extraction, including: Construct an industrial knowledge graph expert system, represent the expert experience rules in the form of triples, and each triple contains a process parameter node, a process parameter relationship, and a process control rule to form an expert rule base of the industrial knowledge graph; Calculate the matching degree between the process parameter configuration in the initial process optimization strategy and the process control rules in the expert rule base, assign a credibility weight to each process control rule based on the matching degree, and establish a priority ranking of the process control rules; Construct a symbolic reasoning chain according to the priority order, input the process parameter configuration of the initial process optimization strategy into the symbolic reasoning chain, and judge the activation state of the process control rules based on the credibility weight to output the symbolic reasoning result; Input the process parameter node and the process parameter relationship into the deep learning model for feature extraction to obtain a process optimization feature vector, combine the process optimization feature vector and the symbolic reasoning result to construct a hybrid strategy network, and output a candidate process optimization plan; Construct an evaluation function including rule compliance and optimization effect, calculate the evaluation score of the candidate process optimization plan under the evaluation function, and adjust the fusion weight of the hybrid strategy network according to the evaluation score to generate an improved process optimization strategy.

7. The method according to claim 6, wherein Calculate the matching degree between the process parameter configuration in the initial process optimization strategy and the process control rules in the expert rule base, assign a credibility weight to each process control rule based on the matching degree, and establish a priority ranking of the process control rules, including: Construct a semantic feature matrix and a structural feature matrix of the process parameter configuration, input the semantic feature matrix into the bidirectional attention network to calculate the semantic association strength between parameters, and input the structural feature matrix into the graph convolutional network to extract the topological relationship features between parameters; Adopt a hierarchical attention mechanism to calculate the matching scores between the process parameter configuration and the process control rules at different abstraction levels, and combine the semantic association strength and the topological relationship features to construct a rule matching degree evaluation model; Based on a preset rule adaptability evaluation index, the rule matching degree and the rule adaptability index are input into a dynamic weighting function to calculate the comprehensive credibility of the process control rules; Using a multi-objective optimization algorithm, combining the rule conflict degree and the rule coverage degree, the comprehensive credibility of the process control rules is normalized to establish a priority ranking of the process control rules.

8. A robot cable manufacturing process optimization system based on deep reinforcement learning, for implementing the method described in any one of the preceding claims 1-7, characterized in that, Including: The first unit is used to collect cable manufacturing data; A digital twin model is constructed. The execution components of the manufacturing equipment are constructed as graph network nodes, and the process flows between the components are constructed as graph network edges. Through the graph dynamics network and state iterative optimization of thermal coupling and fluid-structure coupling, the manufacturing process simulation data is obtained; The second unit is used to establish a deep reinforcement learning environment by taking the cable manufacturing data and the manufacturing process simulation data as state inputs and the process parameter adjustment instruction as the action output; The third unit is used to analyze the causal dependence relationship of the process parameter adjustment instruction in the reinforcement learning training environment, establish a parameter adjustment dependence graph, and generate an initial process control strategy based on the parameter adjustment dependence graph; The fourth unit is used to construct an industrial knowledge graph expert system in the form of triples, fuse the initial process optimization strategy with the expert rule base based on the credibility weight, and generate an improved process optimization strategy through a hybrid intelligent method combining symbolic reasoning chain and deep learning feature extraction; The fifth unit is used to adjust the process parameters of the robot cable manufacturing equipment according to the improved process optimization strategy, and feedback the cable manufacturing data collected after adjustment to the digital twin model to form a process optimization closed loop.

9. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

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

Images