Robot machining closed-loop parameter optimization method based on physical information machine learning

Abstract

The invention discloses a robot machining closed-loop parameter optimization method based on physical information machine learning, which comprises the following steps: acquiring key grinding parameters in a robot abrasive belt grinding process, including abrasive particle radius, abrasive belt linear velocity, normal force and robot feeding speed; constructing a physical information neural network prediction model fusing a multi-gate hybrid expert network and a robot grinding mechanism equation; taking minimization of Ra and maximization of MRD as targets, adopting Powerball to improve a multi-target particle swarm algorithm, and remarkably accelerating convergence; inputting the optimized solution set into a processing platform to execute and collect new data, retraining the PINN model by using the new data, repeating the processes of optimization, execution of processing and collection of the new data, model retraining and re-optimization until convergence conditions are met, and outputting a Pareto optimal parameter solution set; according to the method, the MRD and Ra collaborative prediction precision can be remarkably improved, the number of iterations is reduced, and efficient and high-precision grinding machining is achieved.

Description

Technical Field

[0001] This invention belongs to the field of robotic processing technology, and more specifically, relates to a method for optimizing closed-loop parameters in robotic processing based on physical information machine learning. Background Technology

[0002] Robotic belt grinding technology, with its advantages of large workspace, high flexibility, and low cost, has become a key method for precision machining of large structural parts and complex curved surfaces. The core objective of this technology is to optimize key grinding parameters (including abrasive grain radius). sanding belt linear velocity V s Normal force F n Robot feed speed V w Simultaneously improve Material Removal Depth (MRD) and reduce Average Surface Roughness (R). a This allows for efficient and high-precision manufacturing. However, due to the large number of parameter combinations and the complex physical mechanisms involved in the machining process (such as contact mechanics and grinding dynamics), traditional trial-and-error methods are inefficient and costly. There is an urgent need to establish a scientific closed-loop parameter optimization framework to automatically generate the optimal parameter combination.

[0003] Current parameter optimization techniques for robotic belt grinding (RBG) mainly rely on two stages: forward prediction and backward optimization, but these have significant limitations. Regarding forward prediction, existing methods can be divided into three categories: physical mechanism-based models (such as Arcard's wear law and Preston's equation) achieve prediction by establishing tool-workpiece interaction equations; while possessing physical interpretability, they are easily affected by environmental interference, leading to significant prediction errors. Purely data-driven machine learning models can achieve high accuracy with sufficient samples, but their black-box nature results in weak generalization ability and reliance on massive training data, leading to high costs. Hybrid methods, such as Physical Information Neural Networks (PINN), attempt to integrate physical laws and the advantages of data-driven approaches, but existing solutions generally combine MRD and R... a Modeling the prediction model and optimization algorithm separately as independent tasks ignores the physical relationship between them, resulting in limited prediction accuracy. In the reverse optimization stage, multi-objective evolutionary algorithms such as standard multi-objective particle swarm optimization (MOPSO) are widely used to generate Pareto optimal solution sets, but their update process suffers from the gradient vanishing problem, leading to slow convergence speed and requiring hundreds of iterations to obtain a reliable solution set. This results in high computational costs and decision delays, making it difficult to meet the needs of real-time or rapid optimization. More importantly, existing technologies operate the prediction model and optimization algorithm separately: the prediction model does not embed grinding mechanism constraints (such as R...). a(Mathematical correlation with MRD), but the optimization process does not form a closed-loop feedback with the prediction model, which may cause the recommended parameters to violate physical laws or deviate from actual working conditions (such as surface roughness tolerance, workpiece thickness limit).

[0004] In summary, existing technologies face three major bottlenecks in forward prediction, backward optimization, and their synergy: in forward prediction, the mechanistic model exhibits poor adaptability, data-driven models have weak generalization and rely heavily on data, and hybrid models do not fully utilize multi-task correlations and physical constraints; in backward optimization, algorithms such as MOPSO experience slow convergence due to gradient vanishing, limiting real-time performance; and the disconnect between prediction and optimization leads to a lack of physical consistency and system efficiency in parameter recommendations. These shortcomings collectively hinder the application effectiveness of RBG technology in high-precision, high-efficiency manufacturing scenarios, necessitating a closed-loop framework that deeply integrates physical laws, multi-task learning, and accelerated optimization mechanisms to overcome these technological bottlenecks. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a closed-loop parameter optimization method and system for robotic machining based on physical information machine learning. By integrating physical laws with a data-driven multi-task prediction model, an improved multi-objective optimization algorithm, and a closed-loop iteration mechanism, it significantly improves prediction accuracy and generalization ability, accelerates the parameter optimization process, and ensures the physical consistency of recommended parameters and their fit with actual working conditions. It can overcome existing technological bottlenecks, provide a more efficient and accurate parameter optimization solution for robotic belt grinding, and thus promote the development of high-precision manufacturing technology.

[0006] To achieve the above objectives, one aspect of the present invention provides a method for optimizing closed-loop parameters of robot processing based on physical information machine learning, comprising the following steps:

[0007] S1: Data collection on the material removal depth (MRD) and surface roughness (R) during robotic belt grinding operations. a The key grinding parameters are used to form the original grinding parameter dataset; a set of parameters that meet the processing standards during the processing are selected as the initial parameters;

[0008] S2: Construct a multi-gate hybrid expert network structure, embed the RBG mechanism equation of robotic grinding, and apply the physical loss function L... u A weighted fusion optimization model based on supervised loss is used. The training data from the original grinding parameter dataset is input into a multi-gate hybrid expert network to train the model, resulting in a trained Physical Information Neural Network (PINN) prediction model. The initial parameters are input into the PINN model, which outputs MRD and R that conform to physical laws. a The collaborative prediction value;

[0009] S3: To minimize R aWith the goal of maximizing MRD, the physically consistent MRD and R... a The collaborative prediction values ​​are generated using a multi-objective particle swarm optimization algorithm based on Powerball to produce an optimized parameter solution set;

[0010] S4: Input the optimized parameter solution set into the robot belt grinding platform for processing, repeat steps S1 to S3, and sequentially obtain the new grinding parameter dataset, the updated PINN model, the new collaborative prediction value, and the new round of optimized parameter solution set. Repeat the process of "optimization → processing and collecting new data → updating dataset → model retraining → re-prediction → re-optimization" until the preset convergence condition is met; and output the final Pareto optimal parameter solution set.

[0011] S5: Input the Pareto optimal parameter solution set into the robot belt grinding platform to perform processing.

[0012] Furthermore, the key grinding parameters mentioned in step S1 include the abrasive grain radius. sanding belt linear velocity V s Normal force F n Robot feed speed V w .

[0013] Further, step S2 includes:

[0014] S21: Construct a multi-gate hybrid expert network architecture, designing N expert networks, each of which is a fully connected feedforward neural network. Extract MRD and R through a shared expert network. a The common characteristics of the tasks; design task-specific gating networks to output the weight distribution of the expert network, which is used to capture differences between tasks and dynamically modulate the combination of expert network outputs;

[0015] S22: Input the training data from the original grinding parameter dataset into both the expert network and the gated network; each expert network outputs its prediction results, and the gated network outputs its weights. The prediction results from each expert network are then weighted and summed to obtain preliminary predictions of the material removal depth and surface roughness.

[0016]

[0017] Where 'a' represents the a-th expert network, and a = 1 corresponds to R. a For prediction, a=2 corresponds to MRD prediction; l represents the l-th expert network; N is the total number of expert networks; The weights represent the outputs of the a-th gating network under the l-th expert network; e l (·) represents the output of the l-th expert network; h a (·) represents the output of the i-th tower network; x represents the set of grinding parameters;

[0018] S23: Based on the RBG processing mechanism and the preliminary predicted values ​​of the material removal depth and surface roughness, establish a physical information loss function between the material removal depth and surface roughness; the physical information loss function is used to measure the deviation between the prediction results and physical laws, ensuring that the predicted values ​​not only fit the data, but also conform to physical constraints;

[0019] S24: Construct supervised loss functions for material removal depth and surface roughness respectively;

[0020] S25: The supervised loss function of material removal depth and surface roughness is weighted and fused with the physical loss to construct the total loss function;

[0021] S26: Use the Adam optimizer to optimize the total loss function, update the network parameters of the multi-gate hybrid expert network, train the physical information neural network prediction model, and obtain the trained physical information neural network PINN prediction model; input the initial parameters into the PINN model, and output the co-predicted values ​​of MRD and that conform to physical laws.

[0022] Furthermore, the physical information loss function is expressed by equation (1):

[0023]

[0024] Among them, L u The physical information loss represents the error between the model's prediction and the physical laws; n is the number of samples. This is the predicted value for surface roughness; is the predicted value of the material removal depth; g(·) is the RBG mechanism equation, representing a mathematical model or function established based on the physical mechanism, used to describe the relationship between surface roughness and material removal depth; ||·|| 2 It represents the square Euclidean norm, which measures the difference between the predicted value and the calculated value based on the physical law. The greater the difference, the greater the loss.

[0025] Furthermore, the total loss function L total This can be expressed by equation (2):

[0026]

[0027] Where λ is a weighting coefficient used to balance the relative importance of physical information loss and other losses; L MRD The supervised loss function for material removal depth; Let be the supervised loss function for surface roughness.

[0028] Further, step S3 includes:

[0029] S31: Define the optimization objective, with the goal of minimizing surface roughness and maximizing material removal depth:

[0030]

[0031] in, and Predicted by physical information machine learning models, These are grinding parameters;

[0032] S32: Randomly generated with a size of N p particle population Where n represents the nth particle; the position of each particle is s. i ∈R 4 Represents a set of grinding parameter combinations; initializes the position and velocity of the particle swarm;

[0033] S33: Use the physical information neural network trained in step S2 to predict the material removal depth and surface roughness corresponding to each set of grinding parameter combinations, and calculate the fitness value of each set of grinding parameter combinations according to the optimization objective.

[0034] S34: Introduce the Powerball nonlinear function into the standard velocity update formula for multi-objective particle swarm optimization to accelerate the update process; update the individual optimal position and the global optimal position in the particle swarm based on the fitness value; update the velocity and position of the particles based on the individual optimal position and the global optimal position;

[0035] S35: Use the hypercube partitioning method to filter out non-dominated solutions from the current particle swarm, and finally output the optimized parameter solution set.

[0036] Furthermore, the velocity update formula for the standard multi-objective particle swarm optimization, after incorporating the Powerball nonlinear function transformation, is expressed by equation (3):

[0037]

[0038] in, It is the velocity of the i-th particle in the d-th dimension during the (k+1)-th iteration; is the velocity of the i-th particle in the d-th dimension during the k-th iteration; c1 and c2 are learning factors; r1 and r2 are random numbers, usually in the range [0,1]; It is the optimal position of the i-th particle in the d-th dimension during the k-th iteration; It is the globally optimal position in the d-th dimension during the k-th iteration; It is the current position of the i-th particle in the d-th dimension during the k-th iteration; σ γ (z)=sign(z)|z| γIt is a Powerball nonlinear function (γ∈[0,1]).

[0039] A second aspect of the present invention provides a robot machining closed-loop parameter optimization system based on physical information machine learning, for implementing the aforementioned robot machining closed-loop parameter optimization method based on physical information machine learning, comprising:

[0040] The data acquisition and processing module is used to collect data on the material removal depth (MRD) and surface roughness (R) during the operation of the robotic belt grinding platform. a The key grinding parameters are used to form the original grinding parameter dataset; the collected data is preprocessed, including cleaning, normalization or standardization; the dataset is divided into training set, validation set and test set; a set of parameters that meet the processing standards during the processing are selected as initial parameters; the key grinding parameters include abrasive grain radius. sanding belt linear velocity V s Normal force F n Robot feed speed V w ;

[0041] The physical information multi-task prediction model construction module is used to build a multi-gate hybrid expert network structure, embedding the RBG mechanism equation of robot grinding, and using the physical loss function L u A weighted fusion optimization model based on supervised loss is used. The training data from the original grinding parameter dataset is input into a multi-gate hybrid expert network to train the model, resulting in a trained Physical Information Neural Network (PINN) prediction model. The initial parameters are input into the PINN model, which outputs MRD and R that conform to physical laws. a The collaborative prediction value;

[0042] A multi-objective optimization module to minimize R a With the goal of maximizing MRD, the physically consistent MRD and R... a The collaborative prediction values ​​are generated using a multi-objective particle swarm optimization algorithm based on Powerball to produce an optimized parameter solution set;

[0043] The closed-loop iterative module is used to input the optimized parameter set into the robotic belt grinding platform for machining, and to collect the actual material removal depth (MRD) and surface roughness (R) in real time during the machining process. aThe data is used to form a new grinding parameter dataset; the physical information neural network (PINN) prediction model is retrained on the new grinding parameter dataset to obtain an updated PINN model; the optimized parameter solution set is input into the updated PINN model for prediction, and a new collaborative prediction value is output; the new collaborative prediction value is input, and the multi-objective optimization module is repeatedly executed to generate a new round of optimized parameter solution set; the process of "optimization → processing and collecting new data → updating dataset → model retraining → re-prediction → re-optimization" is repeated until the preset convergence condition is met.

[0044] A third aspect of the present invention provides an electronic device including a processor and a memory interconnected thereto;

[0045] The memory is used to store computer programs;

[0046] The processor is configured to execute the physical information-based machine learning-based robot machining closed-loop parameter optimization method when the computer program is invoked.

[0047] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that is executed by a processor to implement the robot machining closed-loop parameter optimization method based on physical information machine learning.

[0048] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0049] (1) To address the issues that existing physical mechanism models are susceptible to environmental interference leading to significant prediction errors, and that purely data-driven machine learning models have weak generalization capabilities and rely heavily on massive training data, this invention presents a robot processing closed-loop parameter optimization method and system based on physical information machine learning. This method constructs a multi-gate hybrid expert network structure, weightedly fusing the physical information loss function with the supervised loss. This ensures that the prediction model not only fits the data but also conforms to physical constraints, fully utilizing the advantages of physical laws and data-driven approaches, thereby improving prediction accuracy and generalization ability. Existing hybrid models generally model material removal depth and surface roughness as independent tasks, ignoring the physical correlation between them. This invention, by constructing a physical information loss function, correlates the predictions of material removal depth and surface roughness, fully utilizing the correlation between multiple tasks and further improving prediction accuracy.

[0050] (2) In view of the gradient vanishing problem in existing multi-objective evolutionary algorithms such as standard multi-objective particle swarm optimization (MOPSO), which leads to slow convergence speed and requires hundreds of iterations to obtain a reliable solution set, the robot processing closed-loop parameter optimization method and system based on physical information machine learning of the present invention adopts a multi-objective particle swarm optimization algorithm based on Powerball, and introduces Powerball nonlinear function into the velocity update formula to accelerate the particle swarm update process, thereby improving optimization efficiency, reducing the number of iterations, and reducing computational cost and decision delay.

[0051] (3) In view of the problem that the existing technology system operates the prediction model and the optimization algorithm separately, the prediction model does not embed the grinding mechanism constraint, and the optimization process does not form a closed loop feedback with the prediction model, which may lead to the recommended parameters violating physical laws or deviating from the actual working conditions, the robot machining closed loop parameter optimization method and system based on physical information machine learning of the present invention inputs the optimal parameter solution set into the robot belt grinding platform through the closed loop iteration module, collects the actual data in real time during the processing, and feeds the measured data back to the physical information neural network prediction model to evaluate the measured data and update the prediction model, forming a closed loop iterative optimization, ensuring the physical consistency of the parameter recommendation, and improving the processing quality and model accuracy.

[0052] (4) The robot machining closed-loop parameter optimization method and system based on physical information machine learning of the present invention breaks through the existing technical bottleneck by integrating physical laws, multi-task learning and accelerated optimization closed-loop framework. It can automatically generate the optimal parameter combination and achieve the goal of reducing surface roughness while increasing the material removal depth. This significantly improves the efficiency and accuracy of robot belt grinding and provides a more efficient and reliable solution for high-precision manufacturing scenarios. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating a robot machining closed-loop parameter optimization method based on physical information machine learning, according to an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the internal logic of a robot processing closed-loop parameter optimization method based on physical information machine learning according to an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of a robot machining closed-loop parameter optimization system based on physical information machine learning according to an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0058] like Figure 1 and Figure 2 As shown, one aspect of the present invention provides a method for optimizing closed-loop parameters of robot processing based on physical information machine learning, comprising the following steps:

[0059] S1: Data collection on the material removal depth (MRD) and surface roughness (R) during robotic belt grinding operations. a The key grinding parameters are used to form the original grinding parameter dataset; a set of parameters that meet the processing standards during the processing are selected as the initial parameters;

[0060] S2: Construct a multi-gate hybrid expert network structure, embed the RBG mechanism equation of robotic grinding, and apply the physical loss function L... u A weighted fusion optimization model based on supervised loss is used. The training data from the original grinding parameter dataset is input into a multi-gate hybrid expert network to train the model, resulting in a trained Physical Information Neural Network (PINN) prediction model. The initial parameters are input into the PINN model, which outputs MRD and R that conform to physical laws. a The collaborative prediction value;

[0061] S3: To minimize R a With the goal of maximizing MRD, the physically consistent MRD and R... a The collaborative prediction values ​​are generated using a multi-objective particle swarm optimization algorithm based on Powerball to produce an optimized parameter solution set;

[0062] S4: Input the optimized parameter solution set into the robot belt grinding platform for processing, repeat steps S1 to S3, and sequentially obtain the new grinding parameter dataset, the updated PINN model, the new collaborative prediction value, and the new round of optimized parameter solution set. Repeat the process of "optimization → processing and collecting new data → updating dataset → model retraining → re-prediction → re-optimization" until the preset convergence condition is met; and output the final Pareto optimal parameter solution set.

[0063] S5: Input the Pareto optimal parameter solution set into the robot belt grinding platform to perform processing.

[0064] Further, step S1 includes:

[0065] Determine the factors affecting material removal depth (MRD) and surface roughness (R). a Key grinding parameters; the key grinding parameters include abrasive grain radius. sanding belt linear velocity V s Normal force F n Robot feed speed V w ;

[0066] Based on actual processing capabilities and process requirements, the value range and sampling step size of each key grinding parameter were set; a robotic grinding experimental platform was independently built, integrating ABB industrial robots, belt sanders, and high-precision measuring equipment, generating 80 sets of orthogonal experimental data covering the entire parameter range. MRD was obtained through a self-developed measurement system, while Ra was accurately measured by a digital surface roughness tester to ensure comprehensive and reliable data.

[0067] The input parameters and their corresponding output results are organized into data samples to form the original grinding parameter dataset.

[0068] The original grinding parameter data is preprocessed, including cleaning (removing outliers and filling missing values), normalization or standardization, and eliminating the influence of dimensions.

[0069] The preprocessed grinding parameter dataset was divided into a training set and a test set.

[0070] Further, step S2 includes:

[0071] S21: Construct a multi-gate hybrid expert network architecture, designing N expert networks, each of which is a fully connected feedforward neural network. Extract MRD and R through a shared expert network. a The common characteristics of the tasks; design task-specific gating networks to output the weight distribution of the expert network, which is used to capture differences between tasks and dynamically modulate the combination of expert network outputs;

[0072] S22: Input the training data from the original grinding parameter dataset into both the expert network and the gated network; each expert network outputs its prediction results, and the gated network outputs its weights. The prediction results from each expert network are then weighted and summed to obtain preliminary predictions of the material removal depth and surface roughness.

[0073]

[0074] Where 'a' represents the a-th expert network, and a = 1 corresponds to R. a For prediction, a=2 corresponds to MRD prediction; l represents the l-th expert network; N is the total number of expert networks; The weights represent the outputs of the a-th gating network under the l-th expert network; e l (·) represents the output of the l-th expert network; ha (·) represents the output of the i-th tower network; x represents the set of grinding parameters;

[0075] S23: Based on the RBG processing mechanism and the preliminary predicted values ​​of the material removal depth and surface roughness, establish a physical information loss function between the material removal depth and surface roughness; the physical information loss function is used to measure the deviation between the prediction results and physical laws, ensuring that the predicted values ​​not only fit the data, but also conform to physical constraints;

[0076] S24: Construct supervised loss functions for material removal depth and surface roughness respectively;

[0077] S25: The supervised loss function of material removal depth and surface roughness is weighted and fused with the physical loss to construct the total loss function;

[0078] S26: Optimize the total loss function using the Adam optimizer, update the network parameters of the multi-gate hybrid expert network, train the physical information neural network prediction model, and obtain the trained physical information neural network (PINN) prediction model; input the initial parameters into the PINN model, and output the MRD and R values ​​that conform to physical laws. a The collaborative prediction value.

[0079] Furthermore, the physical information loss function is expressed by equation (1):

[0080]

[0081] Among them, L u The physical information loss represents the error between the model's prediction and the physical laws; n is the number of samples. This is the predicted value for surface roughness; is the predicted value of the material removal depth; g(·) is the RBG mechanism equation, representing a mathematical model or function established based on the physical mechanism, used to describe the relationship between surface roughness and material removal depth; ||·|| 2 It represents the square Euclidean norm, which measures the difference between the predicted value and the calculated value based on the physical law. The greater the difference, the greater the loss.

[0082] Furthermore, the total loss function L total This can be expressed by equation (2):

[0083]

[0084] Where λ is a weighting coefficient used to balance the relative importance of physical information loss and other losses; L MRD The supervised loss function for material removal depth; Let be the supervised loss function for surface roughness.

[0085] Further, step S3 includes:

[0086] S31: Define the optimization objective, with the goal of minimizing surface roughness and maximizing material removal depth:

[0087]

[0088] in, and Predicted by physical information machine learning models, These are grinding parameters;

[0089] S32: Randomly generated with a size of N p particle population Where n represents the nth particle; the position of each particle is s. i ∈R 4 Represents a set of grinding parameter combinations; initializes the position and velocity of the particle swarm;

[0090] S33: Use the physical information neural network trained in step S2 to predict the material removal depth and surface roughness corresponding to each set of grinding parameter combinations, and calculate the fitness value of each set of grinding parameter combinations according to the optimization objective.

[0091] S34: Introduce the Powerball nonlinear function into the standard velocity update formula for multi-objective particle swarm optimization to accelerate the update process; update the individual optimal position and the global optimal position in the particle swarm based on the fitness value; update the velocity and position of the particles based on the individual optimal position and the global optimal position;

[0092] S35: Use the hypercube partitioning method to filter out non-dominated solutions from the current particle swarm, and finally output the optimized parameter solution set;

[0093] Furthermore, the velocity update formula for the standard multi-objective particle swarm optimization, after incorporating the Powerball nonlinear function transformation, is expressed by equation (3):

[0094]

[0095] in, It is the velocity of the i-th particle in the d-th dimension during the (k+1)-th iteration; is the velocity of the i-th particle in the d-th dimension during the k-th iteration; c1 and c2 are learning factors; r1 and r2 are random numbers, usually in the range [0,1]; It is the optimal position of the i-th particle in the d-th dimension during the k-th iteration; It is the globally optimal position in the d-th dimension during the k-th iteration; It is the current position of the i-th particle in the d-th dimension during the k-th iteration; σ γ (z)=sign(z)|z| γ It is a Powerball nonlinear function (γ∈[0,1]).

[0096] Further, step S4 includes

[0097] S41: Input the optimized parameter set into the robotic belt grinding platform for machining, and collect the actual material removal depth (MRD) and surface roughness (R) in real time during the machining process. a data,

[0098] Create a new grinding parameter dataset;

[0099] S42: For the new grinding parameter dataset, retrain the physical information neural network PINN prediction model described in step S2 to obtain the updated PINN model;

[0100] S43: Input the optimized parameter solution set into the updated PINN model described in step S42 to make predictions and output new collaborative prediction values;

[0101] S44: Taking the new collaborative prediction value as input, repeat step S3 (optimization) to generate a new set of optimized parameters;

[0102] S45: Repeat steps S41 to S44 of “optimization → processing and collecting new data → updating dataset → model retraining → re-predicting → re-optimization” until the preset convergence condition is met; after the convergence condition is met, output the final Pareto optimal parameter solution set.

[0103] This invention improves the velocity update formula of the particle swarm optimization algorithm by introducing the Powerball nonlinear function, thereby accelerating the convergence process and ultimately obtaining a uniformly distributed Pareto optimal parameter set. It also alleviates the gradient vanishing problem through an exponential decay mechanism, significantly improving convergence efficiency. This method adjusts the particle update behavior, enabling the algorithm to explore and develop the search space more effectively, thus increasing the probability of finding the global optimum.

[0104] The closed-loop parameter optimization method proposed in this invention can significantly improve the overall performance of robotic belt grinding (RBG) manufacturing.

[0105] In terms of processing quality, the MMoE multi-task prediction model (PINN) with physical information enhancement significantly reduces the prediction error of material removal depth and surface roughness: the mean absolute percentage error (MAPE) of material removal depth prediction is reduced from 42.4% in the Shared-Bottom method to 13.8%, and the mean absolute percentage error of surface roughness prediction is reduced from 11.7% in the Shared-Bottom method to 7.0%, which can fundamentally ensure the accuracy of parameter optimization input.

[0106] In terms of production efficiency, the innovative Powerball-MOPSO algorithm reduces the number of optimization iterations by an average of 87% (the number of convergence iterations is reduced from 197 steps to 25 steps), while the generated Pareto solution set is more evenly distributed, significantly shortening the decision cycle.

[0107] In terms of resource consumption, the closed-loop mechanism forces the parameter set to strictly meet the physical constraints of grinding, avoids invalid parameter experiments, and reduces the raw material loss due to trial and error by more than 80%. The recommended parameters have been verified to improve the material removal depth by 18.5% while controlling the surface roughness within the precision range of 2.1–3.6μm, and reduce the energy consumption per workpiece by 27.3%.

[0108] In terms of intelligent operation, the prediction model is automatically updated through real-time feedback of processing data, which enables parameters to adapt to changes in working conditions and reduces the frequency of manual intervention by 90%. The recommended parameters for two sets of standard abrasive sizes can be directly applied to industrial production lines, which significantly improves the surface quality compliance rate of complex curved surface grinding.

[0109] like Figure 3 As shown, a second aspect of the present invention provides a robot machining closed-loop parameter optimization system based on physical information machine learning, for implementing the above-mentioned design method, comprising:

[0110] The data acquisition and processing module is used to collect key grinding parameters that affect the material removal depth and surface roughness during the operation of the robotic belt grinding platform, forming a raw grinding parameter dataset. in For grinding parameters, y j =[MRD j ,R a,j ] T The measured values ​​are used as the initial parameters. The collected data undergoes preprocessing, including cleaning, normalization, or standardization. The dataset is divided into training, validation, and test sets. A set of parameters conforming to processing standards is selected as the initial parameters. The key grinding parameters include the abrasive grain radius. sanding belt linear velocity V s Normal force F n Robot feed speed V w;

[0111] The physical information multi-task prediction model construction module is used to build a multi-gate hybrid expert network structure, embedding the RBG mechanism equation of robot grinding, and using the physical loss function L u A weighted fusion optimization model based on supervised loss is used. The training data from the original grinding parameter dataset is input into a multi-gate hybrid expert network to train the model, resulting in a trained Physical Information Neural Network (PINN) prediction model. The initial parameters are input into the PINN model, which outputs MRD and R that conform to physical laws. a The collaborative prediction value;

[0112] The multi-objective optimization module is used to define multi-objective optimization problems to minimize R. a With the goal of maximizing MRD, the physically consistent MRD and R... a The collaborative prediction values ​​are generated using a multi-objective particle swarm optimization algorithm based on Powerball to produce an optimized parameter solution set;

[0113] The closed-loop iterative module is used to input the optimized parameter set into the robotic belt grinding platform for machining, and to collect the actual material removal depth (MRD) and surface roughness (R) in real time during the machining process. a The data is used to form a new grinding parameter dataset; the physical information neural network (PINN) prediction model is retrained on the new grinding parameter dataset to obtain an updated PINN model; the optimized parameter solution set is input into the updated PINN model for prediction, and a new collaborative prediction value is output; the new collaborative prediction value is input, and the multi-objective optimization module is repeatedly executed to generate a new round of optimized parameter solution set; the process of "optimization → processing and collecting new data → updating dataset → model retraining → re-prediction → re-optimization" is repeated until the preset convergence condition is met.

[0114] This invention, through the aforementioned modules, constitutes a complete closed-loop parameter optimization system, which can automatically generate the optimal parameter combination and achieve dynamic calibration of parameter recommendations through real-time feedback and periodic model retraining. Through iterative optimization, it continuously improves processing quality and model accuracy.

[0115] It should be noted that the robot machining closed-loop parameter optimization system based on physical information machine learning provided in this embodiment can be a computer program (including program code) running on a computer device. For example, the robot machining closed-loop parameter optimization system based on physical information machine learning is an application software. The robot machining closed-loop parameter optimization system based on physical information machine learning can be used to execute the corresponding steps in the above-described method provided in the embodiments of this application.

[0116] In some feasible implementations, the robot machining closed-loop parameter optimization system based on physical information machine learning provided in this embodiment can be implemented in a combination of hardware and software. As an example, the robot machining closed-loop parameter optimization system based on physical information machine learning provided in this application embodiment can be a processor in the form of a hardware decoding processor, which is programmed to execute the robot machining closed-loop parameter optimization method based on physical information machine learning provided in this application embodiment. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0117] In some feasible implementations, the robot machining closed-loop parameter optimization system based on physical information machine learning provided in this embodiment can be implemented in software. It can be software in the form of programs and plug-ins, and includes a series of modules to implement the robot machining closed-loop parameter optimization method based on physical information machine learning provided in this embodiment of the invention.

[0118] A third aspect of the present invention also provides an electronic device, Figure 4 This is a schematic diagram of the electronic device in this embodiment, as shown below. Figure 4 As shown, the electronic device 1000 in this embodiment may include: a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the electronic device 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The user interface 1003 may include a display screen and a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 4As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0119] like Figure 4 In the electronic device 1000 shown, the network interface 1004 provides network communication functions; the user interface 1003 is mainly used to provide an input interface for users; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the above-mentioned optimization method.

[0120] It should be understood that in some feasible implementations, the processor 1001 described above may be a central processing unit (CPU), which may also be other general-purpose processors, DSPs, ASICs, FPGAs, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0121] In specific implementation, the aforementioned electronic device 1000 can perform the above-described actions through its built-in functional modules. Figure 1 The implementation methods provided for each step are detailed in the above-mentioned implementation methods, and will not be repeated here.

[0122] This application also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement... Figure 1 The methods provided in each step are detailed in the implementation methods provided in the above steps, and will not be repeated here.

[0123] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0124] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing closed-loop parameters in robot machining based on physical information machine learning, characterized in that, Includes the following steps: S1: Data collection on the material removal depth (MRD) and surface roughness (R) during robotic belt grinding operations. a The key grinding parameters are used to form the original grinding parameter dataset; a set of parameters that meet the processing standards during the processing are selected as the initial parameters; S2: Construct a multi-gate hybrid expert network structure, embed the RBG mechanism equation of robotic grinding, and apply the physical loss function L... u A weighted fusion optimization model based on supervised loss is used. The training data from the original grinding parameter dataset is input into a multi-gate hybrid expert network to train the model, resulting in a trained Physical Information Neural Network (PINN) prediction model. The initial parameters are input into the PINN model, which outputs MRD and R that conform to physical laws. a The collaborative prediction value; S3: To minimize R a With the goal of maximizing MRD, the physically consistent MRD and R... a The collaborative prediction values ​​are generated using a multi-objective particle swarm optimization algorithm based on Powerball to produce an optimized parameter solution set; S4: Input the optimized parameter solution set into the robot belt grinding platform for processing, repeat steps S1 to S3, and sequentially obtain the new grinding parameter dataset, the updated PINN model, the new collaborative prediction value, and the new round of optimized parameter solution set. Repeat the process of "optimization → processing and collecting new data → updating dataset → model retraining → re-prediction → re-optimization" until the preset convergence condition is met; and output the final Pareto optimal parameter solution set. S5: Input the Pareto optimal parameter solution set into the robot belt grinding platform to perform processing.

2. The method for optimizing closed-loop parameters of robot processing based on physical information machine learning according to claim 1, characterized in that: The key grinding parameters mentioned in step S1 include abrasive grain radius. sanding belt linear velocity V s Normal force F n Robot feed speed V w .

3. The method for optimizing closed-loop parameters of robot processing based on physical information machine learning according to claim 2, characterized in that, Step S2 includes: S21: Construct a multi-gate hybrid expert network architecture, designing N expert networks, each of which is a fully connected feedforward neural network. Extract MRD and R through a shared expert network. a The common characteristics of the tasks; design task-specific gating networks to output the weight distribution of the expert network, which is used to capture differences between tasks and dynamically modulate the combination of expert network outputs; S22: Input the training data from the original grinding parameter dataset into both the expert network and the gated network; each expert network outputs its prediction results, and the gated network outputs its weights. The prediction results from each expert network are then weighted and summed to obtain preliminary predictions of the material removal depth and surface roughness. Where 'a' represents the a-th expert network, and a = 1 corresponds to R. a For prediction, a=2 corresponds to MRD prediction; l represents the l-th expert network; N is the total number of expert networks; The weights represent the outputs of the a-th gating network under the l-th expert network; e l (·) represents the output of the l-th expert network; h a (·) represents the output of the l-th expert network; x represents the set of grinding parameters; S23: Based on the RBG processing mechanism and the preliminary predicted values ​​of the material removal depth and surface roughness, establish a physical information loss function between the material removal depth and surface roughness; the physical information loss function is used to measure the deviation between the prediction results and physical laws, ensuring that the predicted values ​​not only fit the data, but also conform to physical constraints; S24: Construct supervised loss functions for material removal depth and surface roughness respectively; S25: The supervised loss function of material removal depth and surface roughness is weighted and fused with the physical loss to construct the total loss function; S26: Use the Adam optimizer to optimize the total loss function, update the network parameters of the multi-gate hybrid expert network, train the physical information neural network prediction model, and obtain the trained physical information neural network PINN prediction model; input the initial parameters into the PINN model, and output the co-predicted values ​​of MRD and that conform to physical laws.

4. A method for optimizing closed-loop parameters of robot processing based on physical information machine learning according to any one of claims 1-3, characterized in that: The physical information loss function is expressed by equation (1): Among them, L u The physical information loss represents the error between the model's prediction and the physical laws; n is the number of samples. This is the predicted value for surface roughness; is the predicted value of the material removal depth; g(·) is the RBG mechanism equation, representing a mathematical model or function established based on the physical mechanism, used to describe the relationship between surface roughness and material removal depth; ||·|| 2 It represents the square Euclidean norm, which measures the difference between the predicted value and the calculated value based on the physical law. The greater the difference, the greater the loss.

5. The method for optimizing closed-loop parameters of robot processing based on physical information machine learning according to claim 4, characterized in that: The total loss function L total This can be expressed by equation (2): Where λ is a weighting coefficient used to balance the relative importance of physical information loss and other losses; L MRD The supervised loss function for material removal depth; Let be the supervised loss function for surface roughness.

6. A method for optimizing closed-loop parameters of robot machining based on physical information machine learning according to any one of claims 1-3 or 5, characterized in that: Step S3 includes: S31: Define the optimization objective, with the goal of minimizing surface roughness and maximizing material removal depth: in, and R predicted by the physical information machine learning model a With the value of MRD, It is a set of grinding parameters; f 1 (·) represents the first optimization objective; f 2 (·) represents the second optimization objective; S32: Randomly generated with a size of N p particle population Where n represents the nth particle; the position of each particle is s. i ∈R 4 Represents a set of grinding parameter combinations; initializes the position and velocity of the particle swarm; S33: Use the physical information neural network trained in step S2 to predict the material removal depth and surface roughness corresponding to each set of grinding parameter combinations, and calculate the fitness value of each set of grinding parameter combinations according to the optimization objective. S34: Introduce the Powerball nonlinear function into the standard velocity update formula for multi-objective particle swarm optimization to accelerate the update process; update the individual optimal position and the global optimal position in the particle swarm based on the fitness value; update the velocity and position of the particles based on the individual optimal position and the global optimal position; S35: Use the hypercube partitioning method to filter out non-dominated solutions from the current particle swarm, and finally output the optimized parameter solution set.

7. The method for optimizing closed-loop parameters of robot processing based on physical information machine learning according to claim 6, characterized in that: The velocity update formula for the standard multi-objective particle swarm optimization, after introducing the Powerball nonlinear function transformation, is expressed by equation (3): in, It is the velocity of the i-th particle in the d-th dimension during the (k+1)-th iteration; is the velocity of the i-th particle in the d-th dimension during the k-th iteration; c1 and c2 are learning factors; r1 and r2 are random numbers, usually in the range [0,1]; It is the optimal position of the i-th particle in the d-th dimension during the k-th iteration; It is the globally optimal position in the d-th dimension during the k-th iteration; It is the current position of the i-th particle in the d-th dimension during the k-th iteration; σ γ (z)=sign(z)|z| γ It is a Powerball nonlinear function, γ∈[0,1].

8. A closed-loop parameter optimization system for robot processing based on physical information machine learning, characterized in that, A method for optimizing closed-loop parameters of robot machining based on physical information machine learning as described in any one of claims 1-7, comprising: The data acquisition and processing module is used to collect key grinding parameters affecting material removal depth and surface roughness during the operation of the robotic belt grinding platform, forming a raw grinding parameter dataset; preprocessing the collected data, including cleaning, normalization, or standardization; and dividing the dataset into training, validation, and test sets; the key grinding parameters include abrasive grain radius. sanding belt linear velocity V s Normal force F n Robot feed speed V w ; The physical information multi-task prediction model construction module is used to build a multi-gate hybrid expert network structure, embedding the RBG mechanism equation of robot grinding, and using the physical loss function L u A weighted fusion optimization model based on supervised loss is used. The training data from the original grinding parameter dataset is input into a multi-gate hybrid expert network to train the model, resulting in a trained Physical Information Neural Network (PINN) prediction model. The initial parameters are input into the PINN model, which outputs MRD and R that conform to physical laws. a The collaborative prediction value; The multi-objective optimization module is used to define multi-objective optimization problems to minimize R. a With the goal of maximizing MRD, the physically consistent MRD and R... a The collaborative prediction values ​​are generated using a multi-objective particle swarm optimization algorithm based on Powerball to produce an optimized parameter solution set; The closed-loop iterative module is used to input the optimized parameter set into the robotic belt grinding platform for machining, and to collect the actual material removal depth (MRD) and surface roughness (R) in real time during the machining process. a The data is used to form a new grinding parameter dataset; the physical information neural network (PINN) prediction model is retrained on the new grinding parameter dataset to obtain an updated PINN model; the optimized parameter solution set is input into the updated PINN model for prediction, and a new collaborative prediction value is output; the new collaborative prediction value is input, and the multi-objective optimization module is repeatedly executed to generate a new round of optimized parameter solution set; the process of "optimization → execution of processing to collect new data → update dataset → model retraining → re-prediction → re-optimization" is repeated until the preset convergence condition is met.

9. An electronic device, characterized in that, It includes a processor and a memory, which are interconnected; The memory is used to store computer programs; The processor is configured to execute, when the computer program is invoked, the robot machining closed-loop parameter optimization method based on physical information machine learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement the robot machining closed-loop parameter optimization method based on physical information machine learning as described in any one of claims 1 to 7.

Images