A robot process online regulation method adaptive to material uncertainty

Abstract

The application discloses a kind of robot process online regulation and control methods suitable for material uncertainty, and it is related to intelligent control technical field.The application collects the process parameter sequence of the component to be processed, initial topography data and post-processing topography data;Through the differentiable topography evolution agent model, the derivable predicted topography data is output;A two-parameter inversion optimization model is constructed, the actual and predicted topography difference is quantified by mean square error, and the minimum is taken as the goal, the material parameter prior L2 norm constraint and process parameter boundary penalty constraint are embedded;Based on the reverse automatic differentiation algorithm, the gradient vector is solved, and the optimal material parameter is obtained by the random gradient descent algorithm with momentum iterative optimization, and then the future time step process parameter sequence is optimized, and the optimal solution of material parameter and process parameter is output after constraint verification and convergence judgment.The application realizes the accurate online regulation and control of robot process under material uncertainty, and has wide application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and more specifically to an online control method for robotic processes that adapts to material uncertainties. Background Technology

[0002] Developing clean energy and promoting the intelligentization of high-end manufacturing are core pathways for my country to achieve its 2060 carbon neutrality strategic goal and industrial upgrading. Composite materials, due to their advantages such as lightweight, high strength, and corrosion resistance, have been widely used in key areas such as wind power, aerospace, and high-end equipment. In the manufacturing and maintenance of composite material components, surface processing such as grinding and milling is crucial for ensuring assembly accuracy and service reliability. For example, scenarios such as damage repair of wind turbine blades, precision machining of aerospace composite parts, and surface treatment of core components in high-end equipment all require precise control of surface morphology through automated processing. However, the spatial non-uniformity of composite materials (such as differences in fiber layup and resin distribution) and the uncertainty of properties caused by service damage make traditional processing techniques difficult to adapt to dynamically changing material characteristics. This results in poor processing quality stability and low efficiency, becoming a common bottleneck restricting the high-quality development of related industries. To overcome the limitations of manual processing, the industry is gradually promoting robotic automation solutions, and related technologies have made some progress.

[0003] Chinese patent (publication number CN120205872A) discloses a method for optimizing the process parameters of robotic milling of thin-walled composite components for aerospace applications. This method optimizes parameters by constructing a milling force prediction model, thus alleviating the vibration problem in the processing of weakly rigid components. However, this method and most existing methods still have significant common defects. First, the mapping relationship between process parameters and surface morphology lacks high-fidelity and efficient model support, and does not fully consider the nonlinear material removal behavior of multi-mode coupling such as fiber breakage and resin peeling in composite material processing. Traditional finite element simulation calculations are costly and cannot meet the needs of online control. Second, existing robotic milling... Human-based processing relies heavily on fixed parameters or limited teaching. While it can handle simple dynamic changes in a single scenario, it fails to translate physical laws such as material removal mechanisms and mass conservation into explicit constraints, limiting its generalization ability in small-sample, high-value component processing scenarios. Thirdly, it lacks a cross-scenario online control mechanism to adapt to material uncertainties. Existing optimization methods are mostly based on the ideal assumption that material properties are uniform and deterministic, failing to fully consider the commonalities and heterogeneities of composite materials in different industries. When the preset process parameters do not match the actual properties of the materials on site, it can easily lead to underprocessing or overprocessing, seriously affecting product quality and component service safety.

[0004] Therefore, developing a robot adaptive control technology that can scientifically quantify the intrinsic relationship between process parameters, material parameters, and surface morphology, and has strong generalization ability and cross-industry adaptability has become a common need in many fields such as wind power, aerospace, and high-end equipment. Summary of the Invention

[0005] Based on the above-mentioned technical problems, this application discloses a method for online control of robotic processes to adapt to material uncertainties, specifically including:

[0006] Collect the process parameter sequence, initial morphology data, and post-processing morphology data of the component to be processed;

[0007] By using a differentiable morphology evolution proxy model, the input process parameter sequence and initial morphology data are used to output predicted morphology data.

[0008] A two-parameter inversion optimization model is constructed, which combines the regularization terms of prior constraints on material parameters and boundary constraints on process parameters, with the objective of minimizing the morphological difference between the actual morphological 3D scanning data and the predicted morphological data.

[0009] The gradient of the optimization model is solved using an automatic differential algorithm, and the optimal solutions for material parameters and process parameters are obtained through gradient optimization algorithm.

[0010] Preferably, the differentiable morphology evolution proxy model specifically involves: preprocessing the input process parameter sequence and the 3D point cloud of the initial morphology data, wherein the process parameter sequence is normalized and mapped; and the initial morphology data is denoised and sampled using an edge-preserving filtering algorithm to remove outliers.

[0011] The preprocessed initial 3D point cloud of topography is transformed into a graph structure, and the attributes of graph nodes and graph edges are defined; the graph nodes correspond to the sampling points of the topography point cloud, and the weight of the graph edges is determined by the Euclidean distance between adjacent nodes.

[0012] By using a multi-head graph attention mechanism, the process parameter sequence is discretized into process action tuples at time steps, the attention weights of geometric neighbors are calculated, and the node-level process input features are aggregated.

[0013] A spatiotemporally decoupled spatiotemporal graph convolutional network is constructed, in which the spatial graph convolution learns the local interactions of nodes on the surface and captures the spatial evolution law of the morphology; the node-level time series modeling module is responsible for capturing the dynamic changes of the state of each node with processing time; the node-level process input features and graph structure are input to obtain the output of predicted morphology data that can be differentiated with respect to the input.

[0014] Preferably, the dual-parameter inversion optimization model specifically involves: inputting actual morphology data, predicted morphology data, as well as prior material parameters, process parameter boundaries, and initial process parameters; calculating the difference between the two types of morphology data, with minimizing the morphology difference as the primary objective, while embedding two types of regularization constraints—material priors and process boundaries—to solve for the optimal material parameters and process parameters.

[0015] Preferably, the calculation of the difference between the two types of morphological data specifically involves: quantifying the difference between the two types of morphological data using the mean square error (MSE) while ensuring that the difference calculation is differentiable, as shown in the formula:

[0016]

[0017] in, To account for the differences between the two types of morphological data, To obtain the first through 3D scanning Each node thickness For the first The predicted thickness of each node. This represents the total number of nodes.

[0018] Preferably, the two types of regularization constraints, namely material prior and process boundary, are as follows: Since the surrogate model is completely differentiable, a gradient-based optimization algorithm is used for optimization. In order to avoid the inversion results being overly sensitive to measurement noise and to achieve stable inversion, a regularization term is introduced.

[0019] Among them, the prior constraint term for material parameters ensures that the material parameters conform to physical properties through the L2 norm, and constrains the deviation between the material parameters to be inverted and the prior estimates; the boundary constraint term for process parameters constrains the amount of process parameter correction through boundary penalties, ensuring that the corrected parameters are within the physical boundaries.

[0020] Preferably, the primary objective of minimizing morphological differences specifically involves: performing inverse optimization of material and process parameters; first, under preset standard process parameters, optimizing the optimal material parameters; the objective function formula is as follows:

[0021]

[0022] in, For optimal material parameters, To find the minimum state , For actual morphological data, Preset standard process parameters Under the premise that the material parameters are Predicted topographic data at that time;

[0023] Given optimal material parameters, starting from the current moment, we determine the sequence of process parameters for the next Z time steps. The optimization aims to minimize the difference between the predicted and desired morphology while satisfying the limitations of the robot's own motion capabilities. The optimal process parameters are sought, and the objective function formula is as follows:

[0024]

[0025] in, The optimal process parameters are... For the desired morphology, This represents the predicted morphology under optimal material parameters and standard process parameters.

[0026] Preferably, the optimization to obtain the optimal solution for material parameters and the optimal solution for process parameter correction specifically involves: using a back-inverted automatic differentiation algorithm to solve for the gradient values ​​of the total loss function with respect to the material parameters and the process parameter correction values, respectively, to obtain the gradient vector;

[0027] By using the stochastic gradient descent algorithm with momentum, the correction values ​​of material parameters and process parameters are iteratively optimized based on the gradient vector;

[0028] Set convergence criteria, stop the optimization when the iteration meets the convergence criteria, and output the optimal solution for material properties. and optimal solution of process parameters .

[0029] Preferably, obtaining the gradient vector specifically involves: obtaining the total loss function of the two-parameter inversion optimization model; and, based on the total loss function, solving for the gradients of the total loss function with respect to the material parameters and process parameters, respectively, using the following formula:

[0030]

[0031]

[0032] in, Let be the gradient vector of the material parameters. The gradient vector of the process parameters. To predict morphological data, For the total loss function, Mean square error, , For regularization weights, , These are two types of regularization constraints.

[0033] Preferably, the stochastic gradient descent algorithm for the driving force specifically comprises: obtaining the gradient vector obtained by automatic differentiation, and setting a learning rate. Momentum coefficient ;

[0034] When updating the momentum term, the iteration count is incremented by 1. Using the stochastic gradient descent algorithm with momentum, the momentum terms corresponding to the material parameters and process parameters are updated by combining the current gradient vector with the momentum cache value of the previous iteration. Momentum is used to alleviate gradient oscillations and accelerate convergence.

[0035] Based on the updated momentum term and learning rate, the material and process parameters are iteratively updated. The learning rate controls the step size, and the momentum term guides the update direction. The formula is:

[0036]

[0037]

[0038] in, , This is the momentum term for the current iteration. This represents the current iteration number. , This is the momentum term from the previous iteration. , Let be the gradient vector of the current iteration. The momentum coefficient; , The parameter values ​​are updated in the current iteration. , These are the parameter values ​​from the previous iteration. This is the learning rate.

[0039] Preferably, stopping the optimization when the iteration meets the convergence condition specifically involves: verifying the updated material parameters and process parameters to confirm that they respectively meet the prior constraints of the material parameters and the boundary constraints of the process parameters; if they do not meet the constraints, parameter pruning is performed.

[0040] Based on the norm of the joint gradient vector of the iteration, it is determined whether the current iteration meets the convergence condition. If it does, the iteration stops and the optimal solutions for material properties and process parameters are output.

[0041] Compared with the prior art, the technical solution of this application has the following technical effects:

[0042] This invention utilizes a dual-parameter inversion optimization model to first dynamically invert the optimal solution of material parameters based on actual morphology data, and then optimize the process parameter sequence in combination with the optimal material parameters. This addresses the property uncertainties caused by the spatial non-uniformity of composite materials and service damage, achieving dynamic matching of material properties and process parameters. This effectively avoids problems such as over-processing and under-processing caused by traditional fixed-parameter processing, and significantly improves the consistency and reliability of processing quality across different scenarios.

[0043] This invention transforms physical laws such as material removal mechanisms and mass conservation into explicit regularization terms by using prior material constraints and boundary penalty constraints for process parameters. At the same time, it adopts an implicit training logic of pre-training with mechanism data and fine-tuning with high-fidelity data, which greatly improves the model's generalization ability in small-sample, high-value component processing scenarios.

[0044] This invention focuses on the common correlation between composite material processing technology, materials, and morphology. Through modular design, it adapts to the processing needs of components in different fields. Its closed-loop control mechanism of perception-inversion-optimization-execution can be directly embedded into existing robotic processing systems without large-scale hardware modifications, thus lowering the threshold for technology implementation. This promotes the transformation of multiple industries from experience-driven traditional processing modes to precise control intelligent processing modes, and helps the intelligent upgrading of the manufacturing industry.

[0045] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more easily understood, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0046] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0048] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0049] Figure 1 This is a flowchart of an online control method for robotic processes that adapts to material uncertainties.

[0050] Figure 2 This is an architecture diagram of an online control method for robotic processes that adapts to material uncertainties.

[0051] Figure 3 This is a diagram of the architecture of the two-parameter inversion optimization model in this application;

[0052] Figure 4 A schematic diagram illustrating the operation of the two-parameter inversion optimization model in this application for two-parameter optimization;

[0053] Figure 5 This is a schematic diagram of the wind turbine blade block and the integrated data acquisition and polishing robot in the embodiments of this application;

[0054] Figure 6This is a data diagram illustrating the process of optimizing gradients through inverse automatic differentiation in this embodiment of the application.

[0055] Figure 7 This is a probability density comparison chart of the grinding results and target morphology deviations of each method in the embodiments of this application;

[0056] Figure 8 This is a graph showing the comprehensive performance indicators of each method in the embodiments of this application. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0058] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0059] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0060] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0061] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0062] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0063] Example 1 mainly describes an online control method for robotic processes that adapts to material uncertainties, such as... Figure 1 , Figure 2 As shown, it specifically includes:

[0064] Collect the process parameter sequence, initial morphology data, and post-processing morphology data of the component to be processed;

[0065] By using a differentiable morphology evolution proxy model, the input process parameter sequence and initial morphology data are used to output predicted morphology data.

[0066] A two-parameter inversion optimization model is constructed, which combines the regularization terms of prior constraints on material parameters and boundary constraints on process parameters, with the objective of minimizing the morphological difference between the actual morphological 3D scanning data and the predicted morphological data.

[0067] The gradient of the optimization model is solved using an automatic differential algorithm, and the optimal solutions for material parameters and process parameters are obtained through gradient optimization algorithm.

[0068] Furthermore, the differentiable morphology evolution proxy model is specifically as follows: the input process parameter sequence and the 3D point cloud of the initial morphology data are preprocessed, wherein the process parameter sequence is normalized and mapped; the initial morphology data is denoised and sampled using an edge-preserving filtering algorithm to remove outliers.

[0069] The preprocessed initial 3D point cloud of topography is transformed into a graph structure, and the attributes of graph nodes and graph edges are defined; the graph nodes correspond to the sampling points of the topography point cloud, and the weight of the graph edges is determined by the Euclidean distance between adjacent nodes.

[0070] By using a multi-head graph attention mechanism, the process parameter sequence is discretized into process action tuples at time steps, the attention weights of geometric neighbors are calculated, and the node-level process input features are aggregated.

[0071] A spatiotemporally decoupled spatiotemporal graph convolutional network is constructed, in which the spatial graph convolution learns the local interactions of nodes on the surface and captures the spatial evolution law of the morphology; the node-level time series modeling module is responsible for capturing the dynamic changes of the state of each node with processing time; the node-level process input features and graph structure are input to obtain the output of predicted morphology data that can be differentiated with respect to the input.

[0072] Furthermore, the differentiable morphology evolution proxy model specifically involves: preprocessing the input process parameter sequence and the 3D point cloud of the initial morphology data, wherein the process parameter sequence is normalized and mapped; and the initial morphology data is denoised and sampled using an edge-preserving filtering algorithm to remove outliers.

[0073] The preprocessed initial 3D point cloud of topography is transformed into a graph structure, and the attributes of graph nodes and graph edges are defined; the graph nodes correspond to the sampling points of the topography point cloud, and the weight of the graph edges is determined by the Euclidean distance between adjacent nodes.

[0074] By using a multi-head graph attention mechanism, the process parameter sequence is discretized into process action tuples at time steps, the attention weights of geometric neighbors are calculated, and the node-level process input features are aggregated.

[0075] A spatiotemporally decoupled spatiotemporal graph convolutional network is constructed, in which the spatial graph convolution learns the local interactions of nodes on the surface and captures the spatial evolution law of the morphology; the node-level time series modeling module is responsible for capturing the dynamic changes of the state of each node with processing time; the node-level process input features and graph structure are input to obtain the output of predicted morphology data that can be differentiated with respect to the input.

[0076] Furthermore, such as Figure 3 The diagram shows the architecture of the two-parameter inversion optimization model. Specifically, the two-parameter inversion optimization model is as follows: input actual morphology data, predicted morphology data, as well as prior material parameters, boundary process parameters, and initial process parameters; calculate the difference between the two types of morphology data, with minimizing the morphology difference as the main objective, and simultaneously embed two types of regularization constraints: prior material parameters and boundary process parameters, to solve for the optimal material parameters and process parameters.

[0077] Furthermore, the difference between the two types of morphological data is calculated as follows: Under the premise of ensuring the difference calculation is differentiable, the difference between the two types of morphological data is quantified by the mean squared error (MSE), as shown in the formula:

[0078]

[0079] in, To account for the differences between the two types of morphological data, To obtain the first through 3D scanning Each node thickness For the first The predicted thickness of each node. This represents the total number of nodes.

[0080] Furthermore, such as Figure 4The diagram shows the operation of the two-parameter inversion optimization model. There are two types of regularization constraints: material prior and process boundary. Specifically, since the surrogate model is completely differentiable, a gradient-based optimization algorithm is used for optimization. In order to avoid the inversion results being overly sensitive to measurement noise and to achieve stable inversion, a regularization term is introduced.

[0081] Among them, the prior constraint term for material parameters ensures that the material parameters conform to physical properties through the L2 norm, and constrains the deviation between the material parameters to be inverted and the prior estimates. The expression is as follows:

[0082]

[0083] in, For the prior regularization term of the material parameters, These are prior estimates of material parameters. It is an L2 norm;

[0084] Process parameter boundary constraints passed The function generates boundary penalties to constrain the correction amount of process parameters, ensuring that the corrected parameters are within the physical boundaries. The expression is:

[0085]

[0086] in, For process parameter boundary regularization terms, This is the amount of process parameter correction. To preset standard process parameters, This represents the maximum value of the process parameter. This represents the minimum value of the process parameters. To obtain the maximum value.

[0087] Furthermore, with minimizing morphological differences as the primary objective, specifically: optimization of material and process parameters is performed through inversion. First, under preset standard process parameters, the optimal material parameters are sought, and the objective function formula is:

[0088]

[0089] in, For optimal material parameters, To find the minimum state , For actual morphological data, Preset standard process parameters Under the premise that the material parameters are Predicted topographic data at that time;

[0090] Given optimal material parameters, starting from the current moment, we determine the sequence of process parameters for the next Z time steps. The optimization aims to minimize the difference between the predicted and desired morphology while satisfying the limitations of the robot's own motion capabilities. The optimal process parameters are sought, and the objective function formula is as follows:

[0091]

[0092] in, The optimal process parameters are... For the desired morphology, This represents the predicted morphology under optimal material parameters and standard process parameters.

[0093] Furthermore, the optimal solutions for material parameters and process parameter corrections are obtained through optimization. Specifically, the inverse automatic differentiation algorithm is used to solve the gradient values ​​of the total loss function with respect to the material parameter and process parameter correction values, respectively, to obtain the gradient vector.

[0094] By using the stochastic gradient descent algorithm with momentum, the correction values ​​of material parameters and process parameters are iteratively optimized based on the gradient vector;

[0095] Set convergence criteria, stop the optimization when the iteration meets the convergence criteria, and output the optimal solution for material properties. and optimal solution of process parameters .

[0096] Furthermore, the gradient vector is obtained, specifically by obtaining the total loss function of the two-parameter inversion optimization model, as shown in the formula:

[0097]

[0098] in, For the total loss function, Mean square error, , For regularization weights, , These are two types of regularization constraints;

[0099] Based on the total loss function, the gradients of the total loss function with respect to material parameters and process parameters are calculated separately, as shown in the following formulas:

[0100]

[0101]

[0102] in, Let be the gradient vector of the material parameters. The gradient vector of the process parameters. To predict morphological data.

[0103] Furthermore, the stochastic gradient descent algorithm with momentum is specifically as follows: the gradient vector obtained through automatic differentiation is acquired, and a preset learning rate is used. Momentum coefficient ;

[0104] When updating the momentum term, the iteration count is incremented by 1. Using the stochastic gradient descent algorithm with momentum, the momentum terms corresponding to the material parameters and process parameters are updated by combining the current gradient vector with the momentum cache value of the previous iteration. Momentum is used to alleviate gradient oscillations and accelerate convergence.

[0105] Based on the updated momentum term and learning rate, the material and process parameters are iteratively updated. The learning rate controls the step size, and the momentum term guides the update direction. The formula is:

[0106]

[0107]

[0108] in, , This is the momentum term for the current iteration. This represents the current iteration number. , This is the momentum term from the previous iteration. , Let be the gradient vector of the current iteration. The momentum coefficient; , The parameter values ​​are updated in the current iteration. , These are the parameter values ​​from the previous iteration. This is the learning rate.

[0109] Furthermore, when the iteration meets the convergence condition, the optimization stops. Specifically, the updated material parameters and process parameters are checked to confirm that they meet the prior constraints of the material parameters and the boundary constraints of the process parameters, respectively. If they do not meet the constraints, parameter pruning is performed.

[0110] Based on the norm of the joint gradient vector of the iteration, it is determined whether the current iteration satisfies the convergence condition. If it does, the iteration stops, and the optimal solutions for material properties and process parameters are output, as shown in the formula:

[0111]

[0112] in, For the first The norm of the joint gradient vector in the nth iteration. This is the preset norm threshold for the joint gradient vector.

[0113] This embodiment details a method for online control of robot processes to adapt to material uncertainties. The method collects the process parameter sequence, initial morphology data, and post-processing morphology data of the component to be processed. It outputs differentiable predicted morphology data through a differentiable morphology evolution proxy model. A two-parameter inversion optimization model is constructed, aiming to minimize the difference between the actual and predicted morphology by quantifying the mean square error. Prior L2 norm constraints for material parameters and boundary penalty constraints for process parameters are embedded. The gradient vector is solved based on an inverse automatic differentiation algorithm, and iterative optimization is achieved using a stochastic gradient descent algorithm with kinetic energy. The optimal material parameters are first obtained, and then the process parameter sequence for future time steps is optimized. After constraint verification and convergence judgment, the optimal solutions for material and process parameters are output.

[0114] Example 2, based on Example 1, details the process of optimizing material and process parameters for repairing long-term 1.5MW wind turbine blade blocks in wind farms using this method, as follows:

[0115] like Figure 5 As shown on the right, there is an erosion damage area of ​​about 1.2m × 0.8m at the leading edge of the blade block, accompanied by coating peeling and fiber exposure. The material is glass fiber reinforced epoxy resin composite material. Due to long-term service, there are spatial non-uniformity problems such as resin aging and uneven fiber layup density. Traditional fixed parameter grinding is prone to over-grinding or under-grinding.

[0116] like Figure 5 As shown on the left, this method utilizes an integrated data acquisition and polishing robot for data acquisition and polishing repair. The robot's core modules include an adaptive suction cup, a six-degree-of-freedom robotic arm, a force-controlled polisher, a 3D structured light scanner, an integrated vacuum cleaner, and a computing terminal. The adaptive suction cup enables stable adsorption and positioning of the robot on the curved surface of the blade. The six-degree-of-freedom robotic arm, equipped with the force-controlled polisher and the 3D structured light scanner, completes polishing and morphology scanning. The contact force control accuracy of the force-controlled polisher is ±0.1N, ensuring stable force control during the polishing process. The 3D structured light scanner has a scanning accuracy of 0.02mm and is used to acquire initial / actual morphology data. The integrated vacuum cleaner simultaneously collects polishing dust, improving the working environment. The computing terminal is equipped with a GPU server based on the PyTorch framework for proxy model training, gradient solving, and parameter optimization.

[0117] The initial morphological data of the damaged area of ​​the blade block was collected by the three-dimensional structured light scanner of the integrated robot, and 1.2 million three-dimensional point clouds were obtained. After edge-preserving filtering and noise reduction and removal of abnormal points, 300,000 key feature points were sampled and retained.

[0118] Set the preset standard process parameter sequence: grinding head speed 3000 r / min, feed speed 5 mm / s, normal contact force 8 N; obtain the prior parameters of this type of blade material: hardness 35-45 HRC, elastic modulus 28-32 GPa; process parameter boundary: speed 2000-4000 r / min, feed speed 3-8 mm / s, contact force 5-12 N.

[0119] The preprocessed initial topography point cloud is transformed into a graph structure. The graph nodes correspond to the sampling points (including 3D coordinates, normal vectors, and initial thickness attributes). The graph edge weights are determined by the Euclidean distance between adjacent nodes (<5mm is considered an effective connection).

[0120] By employing a multi-head graph attention mechanism, the preset process parameter sequence is discretized into process action tuples with 0.1s time steps. , The geometric neighbor attention weights are calculated for pose, contact force, feed rate, and rotational speed, respectively, and then aggregated to obtain node-level process input features.

[0121] A spatial-temporal decoupled network architecture was constructed. Spatial graph convolution learned the local interactions of surface nodes, and the temporal modeling module used gated recurrent units to capture dynamic changes. The model was trained using 80 sets of historical grinding data of similar blades and embedded with mass conservation physical constraints. The model prediction error converged to ±0.03mm.

[0122] The objective function of the two-parameter inversion is to minimize the mean square error between the actual and predicted morphologies. Prior constraints on material parameters and boundary constraints on process parameters are set. The gradient is solved using inverse automatic differentiation, and iterative optimization is performed using the SGD algorithm (η=0.001, γ=0.9) to obtain the following result: Figure 6 The diagram shows the data of the gradient optimization process using inverse automatic differentiation, based on... Figure 6 The optimal material parameters can be determined. The hardness of the damage core area is 38.2 HRC and the elastic modulus is 29.7 GPa, while the hardness of the edge area is 42.5 HRC and the elastic modulus is 31.3 GPa.

[0123] based on The optimal process parameters were obtained by rolling optimization of the process parameter sequence for the next 20 time steps: core damage zone (resin aging) rotation speed 2800 r / min, feed 4 mm / s, contact force 6.5 N; edge zone (fiber densification) rotation speed 3500 r / min, feed 6 mm / s, contact force 9.2 N.

[0124] The robot performs grinding according to the optimized parameters, triggering a local topography scan every 5mm² of area being ground, updating the data in real time and repeatedly performing inversion optimization.

[0125] After implementing this method, parameter optimization experiments were conducted using existing methods: Aerospace Composite Milling Optimization Method (ACPO) and Dual-Source Curing Repair Method (DSCR). The Aerospace Composite Milling Optimization Method (ACPO) is based on a preset milling force prediction model. It inputs the initial blade morphology and preset process parameters, and obtains fixed optimal process parameters (rotation speed 3200 r / min, feed 5 mm / s, contact force 8 N) through database matching, which are directly used for grinding operations. The Dual-Source Curing Repair Method (DSCR) only optimizes the repair material ratio and the UV-microwave curing process. The grinding process uses traditional fixed parameters (rotation speed 3000 r / min, feed 5 mm / s, contact force 8 N), without involving morphology prediction and process parameter optimization.

[0126] After polishing, the morphology data of the wind turbine blade blocks of each method were collected by scanning. The optimization effects of the three methods were statistically analyzed, and the effect comparison and verification data are shown in Table 1 below.

[0127] Table 1. Comparison of Results and Validation Data

[0128] Evaluation indicators Method of the present invention ACPO (Anti-milling Optimization Method) for Aerospace Composites Dual-source curing repair method DSCR Average deviation between final morphology and target morphology 0.02mm 0.08mm 0.12mm morphological mean square error 0.01mm 0.3mm 0.2mm Over-grinding / under-grinding incidence 2.3% 8.7% 5.1% Single-damage area repair time 3.4h 4.2h 4.5h Post-repair bonding surface flatness pass rate 98.7% 92.3% 88.5% Dust emissions per unit area during grinding 12.5g / m² 18.3g / m² 20.1g / m² Cost of materials per unit area 128.5 yuan / m² 187.2 yuan / m² 153.8 yuan / m²

[0129] According to Table 1 and Figure 7 The probability density comparison data of the grinding results and target morphology deviations shown in the figure demonstrates that the morphology accuracy of the method of this invention is significantly higher than that of the two comparative methods, and the probability density plot is more concentrated, proving that this method more stably addresses the spatial non-uniformity problem of wind turbine blade materials; according to Table 1 and Figure 8 As shown in the comprehensive performance index chart of each method, this method, due to the dual optimization of material parameters and process parameters, outperforms the other two methods in terms of over-grinding / under-grinding rate, single-damage area repair time, and post-repair bonding surface flatness qualification rate. It also achieves lower dust emission per unit area and lower material cost per unit area, verifying the processing accuracy, quality stability, operation efficiency, environmental friendliness, and economy of this method.

[0130] This embodiment describes in detail the process of optimizing material and process parameters for repairing 1.5MW wind turbine blade blocks that have been in long-term service in wind farms using this method. The process is compared with existing methods, and the results verify that after optimizing the material and process parameters using the method of this invention, the economic efficiency of the repair operation is improved while ensuring processing accuracy and quality stability, making it more suitable for large-scale industrial application in wind power operation and maintenance scenarios.

[0131] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A robotic process online regulation method suitable for material uncertainty, characterized in that, include: Collect the process parameter sequence, initial morphology data, and post-processing morphology data of the component to be processed; By using a differentiable morphology evolution proxy model, the input process parameter sequence and initial morphology data are used to output predicted morphology data. A two-parameter inversion optimization model is constructed, which combines the regularization terms of prior constraints on material parameters and boundary constraints on process parameters, with the objective of minimizing the morphological difference between the actual morphological 3D scanning data and the predicted morphological data. The gradient of the optimization model is solved using an automatic differential algorithm, and the optimal solutions for material parameters and process parameters are obtained through gradient optimization algorithm.

2. The online robot process regulation method for material uncertainty adaptation according to claim 1, wherein, The differentiable morphology evolution proxy model specifically involves: preprocessing the input process parameter sequence and the 3D point cloud of the initial morphology data, wherein the process parameter sequence is normalized and mapped; and the initial morphology data is denoised and sampled using an edge-preserving filtering algorithm to remove outliers. The preprocessed initial 3D point cloud of topography is transformed into a graph structure, and the attributes of graph nodes and graph edges are defined; the graph nodes correspond to the sampling points of the topography point cloud, and the weight of the graph edges is determined by the Euclidean distance between adjacent nodes. By using a multi-head graph attention mechanism, the process parameter sequence is discretized into process action tuples at time steps, the attention weights of geometric neighbors are calculated, and the node-level process input features are aggregated. A spatiotemporally decoupled spatiotemporal graph convolutional network is constructed, in which the spatial graph convolution learns the local interactions of nodes on the surface and captures the spatial evolution law of the morphology; the node-level time series modeling module is responsible for capturing the dynamic changes of the state of each node with processing time; the node-level process input features and graph structure are input to obtain the output of predicted morphology data that can be differentiated with respect to the input.

3. The method of claim 1, wherein the material uncertainty is a material property uncertainty. The dual-parameter inversion optimization model specifically involves: inputting actual morphology data, predicted morphology data, as well as prior material parameters, boundary process parameters, and initial process parameters; calculating the difference between the two types of morphology data, with minimizing the morphology difference as the primary objective, while embedding two types of regularization constraints—prior material parameters and boundary process parameters—to solve for the optimal material parameters and process parameters.

4. The method of claim 3, wherein the material uncertainty is determined by a material model. The calculation of the difference between the two types of morphological data specifically involves: ensuring that the difference calculation is differentiable, quantifying the difference between the two types of morphological data using the mean square error (MSE), as shown in the formula: in, To account for the differences between the two types of morphological data, To obtain the first through 3D scanning Each node thickness For the first The predicted thickness of each node. This represents the total number of nodes.

5. The method for online control of robotic processes to adapt to material uncertainties according to claim 4, characterized in that, The two types of regularization constraints, namely material prior and process boundary, are as follows: Since the surrogate model is completely differentiable, a gradient-based optimization algorithm is used for optimization. In order to avoid the inversion results being overly sensitive to measurement noise and to achieve stable inversion, a regularization term is introduced. Among them, the prior constraint term for material parameters ensures that the material parameters conform to physical properties through the L2 norm, and constrains the deviation between the material parameters to be inverted and the prior estimates; the boundary constraint term for process parameters constrains the amount of process parameter correction through boundary penalties, ensuring that the corrected parameters are within the physical boundaries.

6. The method for online control of robotic processes to adapt to material uncertainties according to claim 5, characterized in that, The primary objective of minimizing morphological differences specifically involves: performing inverse optimization of material and process parameters; firstly, under preset standard process parameters, finding the optimal material parameters; the objective function formula is as follows: in, For optimal material parameters, To find the minimum state , For actual morphological data, Preset standard process parameters Under the premise that the material parameters are Predicted topographic data at that time; Given optimal material parameters, starting from the current moment, we determine the sequence of process parameters for the next Z time steps. The optimization aims to minimize the difference between the predicted and desired morphology while satisfying the limitations of the robot's own motion capabilities. The optimal process parameters are sought, and the objective function formula is as follows: in, The optimal process parameters are... For the desired morphology, This represents the predicted morphology under optimal material parameters and standard process parameters.

7. The method for online control of robotic processes to adapt to material uncertainties according to claim 1, characterized in that, The optimization process to obtain the optimal solutions for material parameters and process parameter corrections involves: using an inverse automatic differentiation algorithm to solve for the gradient values ​​of the total loss function with respect to the material parameters and process parameter corrections, respectively, and obtaining the gradient vectors. By using the stochastic gradient descent algorithm with momentum, the correction values ​​of material parameters and process parameters are iteratively optimized based on the gradient vector; Set convergence criteria, stop the optimization when the iteration meets the convergence criteria, and output the optimal solution for material properties. and optimal solution of process parameters .

8. The method for online control of robotic processes to adapt to material uncertainties according to claim 7, characterized in that, The gradient vector is obtained by: acquiring the total loss function of the two-parameter inversion optimization model; and, based on the total loss function, solving for the gradients of the total loss function with respect to the material parameters and process parameters, respectively, using the following formula: in, Let be the gradient vector of the material parameters. The gradient vector of the process parameters. To predict morphological data, For the total loss function, Mean square error, , For regularization weights, , These are two types of regularization constraints.

9. The method for online control of robotic processes to adapt to material uncertainties according to claim 8, characterized in that, The stochastic gradient descent algorithm for the driving force specifically involves: obtaining the gradient vector obtained by automatic differentiation, and setting a learning rate. Momentum coefficient ; When updating the momentum term, the iteration count is incremented by 1. Using the stochastic gradient descent algorithm with momentum, the momentum terms corresponding to the material parameters and process parameters are updated by combining the current gradient vector with the momentum cache value of the previous iteration. Momentum is used to alleviate gradient oscillations and accelerate convergence. Based on the updated momentum term and learning rate, the material and process parameters are iteratively updated. The learning rate controls the step size, and the momentum term guides the update direction. The formula is: in, , This is the momentum term for the current iteration. This represents the current iteration number. , This is the momentum term from the previous iteration. , Let be the gradient vector of the current iteration. The momentum coefficient; , The parameter values ​​are updated in the current iteration. , These are the parameter values ​​from the previous iteration. This is the learning rate.

10. The method for online control of robotic processes to adapt to material uncertainties according to claim 9, characterized in that, The step of stopping optimization when the iteration meets the convergence condition is specifically: verifying the updated material parameters and process parameters to confirm that they respectively meet the prior constraints of material parameters and the boundary constraints of process parameters; if they do not meet the constraints, parameter pruning is performed. Based on the norm of the joint gradient vector of the iteration, it is determined whether the current iteration meets the convergence condition. If it does, the iteration stops and the optimal solutions for material properties and process parameters are output.

Images