A method and system for constructing an envelope grinding material removal depth prediction model

Abstract

The invention belongs to the technical field of intelligent processing and manufacturing, and discloses a method and system for constructing an envelope grinding material removal depth prediction model. Grinding and polishing experiments are carried out in a source domain and a target domain to obtain material removal depths under different grinding and polishing parameter combinations, and the obtained data set is divided into source domain real data and target domain real data; an envelope grinding and polishing material removal depth mechanism model is established, and the least squares method is used to fit the parameters of the mechanism model; a large amount of virtual simulation data is generated in the source domain through the mechanism model, and a data set is constructed for subsequent model training; a mechanism-data domain adaptive adversarial neural network is established, which includes a data feature extraction layer and a data regression layer; the network loss function includes a source domain real data regression loss function, a source domain virtual simulation data regression loss function, a target domain real data regression loss function and a domain classification loss function; and an Adam optimizer is used to train the model and update the parameters.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of intelligent processing and manufacturing technology, and in particular relates to a method and system for constructing an envelope grinding material removal depth prediction model. Background Art

[0002] In the field of aerospace manufacturing, the manufacturing of complex curved surface parts such as aeroengine blades and blisks is a focus and difficulty that attracts attention both at home and abroad. The contour accuracy of such parts directly affects the performance, service life, complex mechanical properties and fluid mechanics properties of aeroengines, so achieving high contour accuracy in manufacturing is of great significance to ensuring the efficient, safe and reliable operation of aeroengines.

[0003] In the processing of blades and blade disk parts, grinding and polishing as the last process can remove the knife marks left by the milling process, thereby improving the surface quality. At present, the grinding and polishing of blades mainly adopts the method of flexible wheel plus abrasive belt, which can ensure the surface quality and contour accuracy of the blade body after grinding and polishing. However, compared with the blade body surface, the blade edge has the characteristics of a sharp change in curvature and sharp corners (radius of curvature <2mm). Therefore, when the blade edge is ground and polished by flexible wheel grinding and polishing, it is very easy to cause over-grinding and scrap the parts. For this reason, the blade edge grinding and polishing mostly adopts the processing method of envelope grinding and polishing, which can effectively alleviate over-grinding and increase the qualified rate of products. Establishing an accurate material removal depth model can provide theoretical guidance for the process design of envelope grinding and polishing, while the traditional mechanism model has residual errors that lead to inaccurate modeling accuracy. The data-driven model is prone to overfitting due to the small sample characteristics of the data of the material removal calibration experiment. Therefore, establishing a high-precision material removal depth prediction model under small sample experimental data is of great significance for the guidance of the envelope grinding and polishing process of the blade edge.

[0004] In view of the above analysis, the technical problems that need to be solved urgently in the prior art are: when the blade edge is ground and polished by a flexible wheel, it is very easy to cause over-grinding and scrap the parts. The envelope grinding and polishing method lacks an accurate material removal model to guide process planning. Summary of the invention

[0005] In response to the problems existing in the prior art, the present invention provides a method and system for constructing an envelope-type grinding material removal depth prediction model. In order to address the problems that the data in the grinding and polishing material removal depth calibration experiment presents small sample characteristics, which makes the data-driven model prone to overfitting, and the mechanism model has modeling residuals and leads to inaccurate modeling, an envelope-type grinding material removal depth prediction model based on a hybrid drive of mechanism and small data is studied.

[0006] The present invention is implemented as follows: an envelope grinding material removal depth prediction method, comprising:

[0007] S1: First, grinding and polishing experiments are carried out in the source domain and the target domain to obtain the material removal depth under different grinding and polishing parameter combinations, and the obtained data set is divided into the source domain real data and the target domain real data;

[0008] S2: Establish an envelope grinding and polishing material removal depth mechanism model, and use the least squares method to fit the parameters of the mechanism model;

[0009] S3: Generate a large amount of virtual simulation data in the source domain through the mechanism model, and build a data set for subsequent model training;

[0010] S4: Establish a mechanism-data domain adaptive adversarial neural network, which includes a data feature extraction layer and a data regression layer; the network's loss function includes a source domain real data regression loss function, a source domain virtual simulation data regression loss function, a target domain real data regression loss function, and a domain classification loss function;

[0011] S5: Use Adam optimizer to train the model and update its parameters.

[0012] Furthermore, the polishing parameters in step S1 are the contact force F between the tool and the workpiece, the feed speed V of the tool, t , the contact speed between the abrasive belt and the workpiece V c , and the curvature radius R of the workpiece w Through the above different parameter combinations x=[F,V t ,V c ,R w ] The grinding and polishing experiments were carried out at different radii, and the grinding and polishing depth h under the corresponding grinding and polishing parameters was measured. The obtained data were divided into source domain real data and the target domain real data in The grinding and polishing parameters meet the following conditions:

[0013]

[0014] in, and are the minimum and maximum values ​​of F in the source domain, and V in the source domain c The minimum and maximum values ​​of and V in the source domain t The minimum and maximum values ​​of and R in the source domain w The minimum and maximum values ​​of .

[0015] The grinding and polishing parameters meet the following conditions:

[0016]

[0017] in, and is the minimum and maximum value of F in the target domain, and V in the target domain c The minimum and maximum values ​​of and V in the target domain t The minimum and maximum values ​​of and R in the target domain w The minimum and maximum values ​​of .

[0018] Furthermore, the envelope grinding and polishing material removal depth mechanism model in step S2 establishes the grinding and polishing parameters F, V t , V c , R w The relationship between the material removal depth h is as follows:

[0019]

[0020] Among them, u is the material removal coefficient, which is affected by the physical properties of the workpiece and the tool itself and is a constant value during the grinding and polishing process. In the actual experimental process, due to the existence of measurement error ε, the above formula can be expressed as:

[0021]

[0022] The values ​​of u and ε are fitted by combining the real source domain data obtained in step S1 with the least squares method. The specific method is as follows:

[0023] in,

[0024] in, are the contact force between the tool and the workpiece, the contact speed between the abrasive belt and the workpiece, the feed speed of the tool, the workpiece radius and the material removal depth in the i-th group of source domain real data. is the average value of m groups of real material removal depths in the source domain. c i The average value of . m is the number of real data in the source domain.

[0025] Further, in step S3, a large amount of mechanism data is generated by the mechanism model fitted in step S2. First, s groups of grinding and polishing parameter combinations are generated in the source domain. in is the contact force between the tool and the workpiece, the contact speed between the abrasive belt and the workpiece, the feed speed of the tool, the workpiece radius and the material removal depth in the i-th group of source domain virtual simulation data. The calculation method is as follows:

[0026]

[0027] Furthermore, the dataset in S3 contains real data from the source domain. Source domain virtual simulation data and the target domain real data It consists of three parts, the specific components are as follows:

[0028]

[0029] Among them, parameters s, m, and n are the number of source domain virtual data, data source domain real data, and target domain real data, which satisfy s>>m,n, m≈n, where the symbol >> represents much greater than, and the symbol ≈ represents approximately equal to. is the grinding and polishing parameter combination of the real data in the source domain, is the grinding and polishing parameter combination of the source domain virtual simulation data and is the grinding and polishing parameter combination of the real data in the target domain. It can be specifically expressed as:

[0030]

[0031] for The corresponding source domain virtual simulation material removal depth, for The corresponding source domain real material removal depth, for The corresponding target domain real material removal depth. It can be specifically expressed as:

[0032]

[0033] where d i is the data label of the i-th group of data, which can be defined in the form of one-hot encoding, which can be specifically expressed as:

[0034]

[0035] Furthermore, the mechanism-data domain adaptive adversarial neural network in step S4 is composed of a feature extraction layer and a data regression layer. The feature extraction layer has a feature extractor G f (x;θ f ), there are three data regressors and one domain classifier in the data regression layer: source domain virtual simulation data regressor G s,v (z;θs,v ), source domain real data regressor G s,a (z;θ s,a ), target domain real data regressor G t (z;θ t ) and domain classifier G d (z;θ d ). Where x is the input of the network, z is the output of the feature extraction layer and serves as the input of the data regression layer, θ f is the network parameter of the feature extractor, θ s,v is the network parameter of the source domain virtual simulation data regressor, θ s,a is the network parameter of the source domain real data regressor, θ t is the network parameter of the target domain real data regressor, θ d is the network parameter of the domain classifier. The output data of the network There are two parts, is the predicted value of material removal depth, is the probability prediction value of the data label. The specific expression is as follows:

[0036]

[0037] Furthermore, the loss function of the network in step S3 is It consists of four parts: Source domain virtual simulation data loss function Source domain real data loss function Target domain real data simulation function And domain classification loss function and The mean square error loss function is used for calculation, which is defined as follows:

[0038]

[0039]

[0040] Furthermore, the domain classification loss function The cross entropy loss function is used for calculation, which is calculated as follows:

[0041]

[0042] in, is the domain classification loss function of the source domain virtual simulation data, is the domain classification loss function of the real data in the source domain, is the loss function of the real data in the target domain. is the predicted probability of the data label of the i-th data, which is expressed as p1, p2, p3 are the i-th data [x, y]i is the probability of virtual simulation data in the source domain, the probability of real data in the source domain and the probability of real data in the target domain, which satisfies p1+p2+p3=1. The operation L(·) is defined as:

[0043]

[0044] Finally, the loss function of the entire network is It can be calculated as follows:

[0045]

[0046] Among them, λ s,a ,λ t and μ is and The weight parameter of . It is specifically expressed as:

[0047]

[0048] Among them, ρ l is the training process, l is the current training round number, and L is the maximum training round number. s,a is s,a The rate of change coefficient, γ t is t The rate of change coefficient, γ d is the rate of change coefficient of μ.

[0049] Further, the specific implementation method of model training in step S5 is as follows:

[0050] (1) First, set the training parameters of the model, including the maximum number of training rounds L and the learning rate η, and use this to initialize the Adam optimizer. The Adam optimizer adjusts the learning rate of each parameter by calculating the first-order moment estimate and the second-order moment estimate of the gradient, thereby achieving efficient network training. Initialize the network parameters θ f ,θ s,v ,θ s,a ,θ t and θ d .

[0051] (2) Calculate by forward propagation of formula The corresponding material removal prediction depth and the probability prediction value of the data label. Then the loss function of the network is calculated by formulas (13)-(18).

[0052] (3) Back propagation through the Adam optimizer and update the network parameters θ f ,θ s,v ,θ s,a ,θ t and θ d .

[0053] (4) Repeat steps (2)-(3) until the number of training rounds reaches the maximum number of training rounds.

[0054] Another object of the present invention is to provide an envelope grinding material removal depth prediction system for realizing the envelope grinding material removal depth prediction method, comprising:

[0055] Dataset division module: Grinding and polishing experiments are conducted in the source domain and the target domain to obtain the material removal depth under different grinding and polishing parameter combinations, and the obtained data set is divided into the source domain real data and the target domain real data;

[0056] Mechanism model building module: used to build the envelope grinding and polishing material removal depth mechanism model, and use the least squares method to fit the parameters of the mechanism model;

[0057] Virtual simulation data generation module: used to generate a large amount of virtual simulation data in the source domain through the mechanism model and construct a data set for subsequent model training;

[0058] Adversarial neural network establishment module: Establish a mechanism-data domain adaptive adversarial neural network, which includes a data feature extraction layer and a data regression layer; the network's loss function includes a source domain real data regression loss function, a source domain virtual simulation data regression loss function, a target domain real data regression loss function, and a domain classification loss function;

[0059] Model training and parameter updating module: Adam optimizer is used to train the model and update its parameters.

[0060] Another object of the present invention is to provide a computer device, the computer device includes a memory and a processor, the memory stores a computer program, when the computer program is executed by the processor, the processor executes the steps of the envelope grinding material removal depth prediction method.

[0061] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the envelope grinding material removal depth prediction method.

[0062] Another object of the present invention is to provide an information data processing terminal, which includes the envelope grinding material removal depth prediction system.

[0063] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0064] First, the present invention can achieve accurate prediction of the removal depth of envelope grinding and polishing materials under small samples through a large amount of virtual simulation data and a small amount of experimental data.

[0065] The present invention generates an envelope material removal mechanism model, which establishes a direct relationship between material removal depth and contact force, contact speed between workpiece and tool, feed speed of tool and radius of workpiece, and provides theoretical support for the generation of a large amount of virtual simulation data.

[0066] The present invention is implemented through a mechanism-data domain adaptive adversarial neural network. During the training process of the network, the weight parameters of the loss function can be adaptively updated according to the training progress, so that the general structure of the network can be quickly updated by a large amount of virtual data in the early stage of training, and then the weight of the loss function of the real data is increased during the training process, thereby improving the prediction accuracy of the real data.

[0067] The data set of the mechanism-data domain adaptive adversarial neural network consists of source domain virtual simulation data, source domain real data and target domain real data, which can simultaneously ensure the prediction accuracy of the established prediction model in both the source domain and the target domain.

[0068] Second, the expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0069] The present invention mainly focuses on the modeling problem of envelope material removal of blade edge parts of complex curved surface parts in aerospace. The envelope grinding and polishing method of the blade edge can alleviate the stress concentration of the contact part between the tool and the blade edge, increase the contact area, thereby reducing over-grinding and reducing the scrap rate. At the same time, the proposed envelope material removal depth prediction method can accurately predict the removal depth with small parameters to guide the planning of the grinding and polishing process. The technology of the present invention can be applied to the planning guidance of blade edge grinding process of complex curved surface parts in aerospace and other manufacturing enterprises, and can improve the yield and production efficiency of related products. The technical solution of the present invention solves the technical problems that people have always been eager to solve but have never been successful:

[0070] At present, the blade edge of blades and blade disk parts is usually polished by manual grinding or grinding and polishing. The manual grinding method often depends on the experience of the workers, and it is difficult to ensure the consistency of the quality after grinding and polishing. The contact wheel grinding method is very easy to over-grind due to the high stress in the contact part between the tool and the workpiece, which makes process planning difficult. Therefore, an envelope grinding method can be used for the blade edge of such parts. However, there are relatively few theoretical studies on the prediction model of envelope grinding material removal depth. The mechanism model has poor accuracy due to the existence of assumptions, and the data-driven model often requires a large amount of training data to fit the model parameters. The data of the grinding and polishing calibration experiment often presents the characteristics of small samples, which makes the data model easy to overfit. Therefore, the current material removal modeling methods are difficult to accurately realize the prediction of the depth of material removal by envelope grinding. Therefore, the present invention provides a mechanism-based and small data hybrid driven prediction model for the material removal depth in the envelope grinding process, and adopts a mechanism-data adaptive domain adversarial neural network to predict the material removal depth under different grinding and polishing process parameters, wherein the data set is composed of source domain real data, source domain virtual simulation data and target domain real data. The source domain virtual simulation data is generated by the mechanism model. The loss function of the mechanism-data adaptive domain adversarial neural network is composed of the source domain real data loss function, the source domain virtual simulation data loss function, the target domain real data simulation function and the domain classification loss function. At the same time, the weights of the loss functions of each part can be adaptively updated during the training process. In the early stage of training, the general structure of the network is obtained through a large amount of virtual simulation data, and then in the later stage of training, the weight of the real data is increased to achieve accurate prediction of the real data under a small amount of data. The invention solves the problem of inaccurate prediction of material removal in envelope grinding and polishing.

[0071] Third, the present invention achieves significant technical progress.

[0072] 1. Accurately predict material removal depth:

[0073] By establishing a mechanism model and an adaptive adversarial neural network, the present invention can accurately predict the material removal depth under different grinding and polishing parameter combinations, thereby improving the control accuracy of the grinding and polishing process.

[0074] 2. Data domain adaptive capability:

[0075] By generating a large amount of virtual data in the source domain and combining it with an adaptive adversarial neural network, the present invention solves the problem of different data distributions in the source domain and the target domain, and improves the generalization ability and prediction accuracy of the model.

[0076] 3. Efficient network training:

[0077] The Adam optimizer is used to adjust the learning rate by calculating the first-order moment estimate and second-order moment estimate of the gradient, which achieves efficient network training, shortens the model training time, and improves the overall efficiency.

[0078] 4. Comprehensive consideration of various grinding and polishing parameters:

[0079] The present invention systematically handles the influence of various grinding and polishing parameters such as contact force, feed speed, contact speed and workpiece curvature radius through a mechanism model, thereby realizing precise control of complex grinding and polishing processes.

[0080] 5. Innovative data construction and model design:

[0081] By generating virtual data through a mechanism model and combining it with an adversarial neural network, the present invention proposes an innovative data construction and model design method, which solves the problems of insufficient data and unstable models in traditional methods.

[0082] Specific application scenarios

[0083] 1. Aerospace Industry:

[0084] The processing of high-precision parts requires extremely high surface quality and material removal accuracy. The method of the present invention can significantly improve the processing quality and consistency.

[0085] 2. Automobile Manufacturing:

[0086] For machining of engine parts and other parts with complex geometric shapes, the method of the present invention can optimize the grinding and polishing force, ensure the uniformity of material removal, and improve the reliability and service life of the parts.

[0087] 3. Shipbuilding industry:

[0088] By precisely controlling the grinding and polishing force and material removal depth, the surface profile manufacturing accuracy of the blade edge of marine turbine blade parts can be significantly improved, reducing subsequent processing steps and reducing costs.

[0089] The present invention systematically solves many technical problems in grinding and polishing force planning by introducing accurate curvature radius calculation, material removal mechanism model, probability density function, Monte Carlo simulation and Bayesian theorem, greatly improves the accuracy, efficiency and consistency of industrial grinding and polishing, and achieves significant technological progress. These advances have broad application prospects and huge economic benefits in industries with high precision requirements such as aerospace, automobile manufacturing, and shipbuilding.

[0090] Fourth, the parameters, algorithms, mathematical models and significant improvements of the technical solution of the present invention.

[0091] 1. Technical issues

[0092] (1) Insufficient accuracy in predicting material removal depth:

[0093] Traditional methods make it difficult to accurately predict the material removal depth, resulting in unstable workpiece processing quality.

[0094] (2) Data imbalance and inter-domain differences:

[0095] The different data distributions in the source domain and the target domain result in insufficient generalization ability of the model.

[0096] (3) It is difficult to deal with the comprehensive impact of multiple parameters:

[0097] The grinding and polishing process involves multiple parameters, and traditional methods are difficult to systematically deal with their comprehensive effects.

[0098] (4) Low network training efficiency:

[0099] High-dimensional data and complex models lead to low training efficiency and high time cost.

[0100] 2. Significant progress

[0101] (1) High-precision prediction:

[0102] By establishing an accurate relationship model between grinding and polishing parameters and material removal depth and using the least squares method to fit the parameters, high-precision material removal depth prediction is achieved.

[0103] (2) Adaptive Adversarial Neural Network:

[0104] Through the mechanism-data domain adaptive adversarial neural network, the problem of different data distribution in the source domain and the target domain is solved, and the generalization ability of the model is improved.

[0105] (3) Multi-parameter comprehensive processing:

[0106] By establishing an envelope grinding and polishing material removal depth mechanism model, the comprehensive influence of various grinding and polishing parameters is systematically processed to achieve accurate grinding and polishing force planning.

[0107] (4) Efficient network training:

[0108] The Adam optimizer is used to improve the network training efficiency, shorten the training time, and achieve efficient model training and parameter optimization.

[0109] The present invention systematically solves many technical problems in grinding and polishing force planning by accurately calculating grinding and polishing parameters, establishing a material removal mechanism model, generating virtual data and combining adaptive adversarial neural networks, significantly improving the accuracy, efficiency and consistency of industrial grinding and polishing. These technological advances have broad application prospects and huge economic benefits in industries with high precision requirements such as aerospace, automobile manufacturing, and shipbuilding. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] Figure 1 is a flow chart of an envelope grinding material removal depth prediction method provided by an embodiment of the present invention;

[0111] Figure 2 It is a structural diagram of an envelope grinding material removal depth prediction system provided in an embodiment of the present invention.

[0112] Figure 3 1. The following are the actual blade part envelope grinding material removal depth prediction result diagrams provided by an embodiment of the present invention: (a) blade edge profile before and after grinding and polishing; (b) material removal depth prediction result; (c) material removal depth prediction error.

[0113] Figure 4 This is a comparison of the prediction results of the envelope grinding material removal depth prediction model provided by the embodiment of the present invention with the mechanism model and the data-driven model: (a) comparison of material removal prediction results in the source domain; (b) comparison of material removal errors in the source domain; (c) comparison of material removal prediction results in the target domain; (d) comparison of material removal errors in the target domain. DETAILED DESCRIPTION

[0114] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0115] Example 1: High-precision grinding of aircraft engine blades

[0116] Aircraft engine blades are key components that require extremely high surface quality and material removal depth accuracy. Traditional grinding methods are difficult to meet these requirements.

[0117] 1) Data collection and grouping (S1)

[0118] Grinding and polishing experiments were carried out to collect material removal depth data under different combinations of grinding and polishing parameters.

[0119] Grinding and polishing parameters include contact force (F), feed speed (V t ), contact speed (V c ) and the blade curvature radius (R w ).

[0120] The data is divided into real data in the source domain (laboratory environment) and the target domain (actual production environment).

[0121] 2) Establish and fit the mechanism model (S2)

[0122] Establish a model of the material removal depth mechanism of envelope grinding and polishing;

[0123] The model parameters were fitted using the least squares method.

[0124] 3) Virtual data generation (S3)

[0125] A large amount of virtual simulation data is generated in the source domain using the mechanism model, and a dataset is constructed for model training.

[0126] 4) Training Adaptive Adversarial Neural Network (S4)

[0127] Construction mechanism - data domain adaptive adversarial neural network, including data feature extraction layer and data regression layer.

[0128] The network loss function includes the source domain virtual simulation data loss function, the source domain real data loss function, the target domain real data loss function and the domain classification loss function.

[0129] 5) Model training and optimization (S5)

[0130] Adam optimizer is used for model training and parameter updating.

[0131] Set the maximum number of training rounds, learning rate and other parameters, and perform back propagation and parameter updates.

[0132] Through this method, the material removal depth in the high-precision grinding process of aircraft engine blades was accurately predicted and controlled, which improved the surface quality of the blades and reduced the rejection rate.

[0133] Example 2: Grinding of blade edges of marine steam turbines

[0134] The blade edge of a marine steam turbine requires high-precision grinding to ensure the performance and service life of the turbine. Traditional grinding methods are difficult to meet the requirements of high precision and high efficiency.

[0135] 1) Data collection and grouping (S1)

[0136] Grinding and polishing experiments were carried out in the laboratory and actual production environment to collect material removal depth data under different combinations of grinding and polishing parameters.

[0137] The grinding and polishing parameters include contact force, feed speed, contact speed and the curvature radius of the blade edge.

[0138] The data is divided into real data in the source domain and real data in the target domain.

[0139] 2) Establish and fit the mechanism model (S2)

[0140] A material removal depth mechanism model of envelope grinding and polishing was established.

[0141] 3) Virtual data generation (S3)

[0142] A large amount of virtual simulation data is generated in the source domain using the mechanism model, and a dataset is constructed for model training.

[0143] 4) Training Adaptive Adversarial Neural Network (S4)

[0144] Construction mechanism - data domain adaptive adversarial neural network, including data feature extraction layer and data regression layer.

[0145] The network loss function includes the source domain virtual simulation data loss function, the source domain real data loss function, the target domain real data loss function and the domain classification loss function.

[0146] 5) Model training and optimization (S5)

[0147] Adam optimizer is used for model training and parameter updating.

[0148] Set the maximum number of training rounds, learning rate and other parameters, and perform back propagation and parameter updates.

[0149] Through this method, the material removal depth in the high-precision grinding process of the blade edge of a marine steam turbine blade was accurately predicted and controlled, which improved the machining contour accuracy of the blade edge, shortened the machining time, and reduced the production cost.

[0150] Through the above two specific application examples, the present invention demonstrates its significant technical progress and practical application value in high-precision grinding and material removal depth prediction. These advances have broad application prospects and huge economic benefits in industries with high precision requirements such as aerospace and shipbuilding.

[0151] like Figure 1 As shown, the present invention provides a mechanism and data hybrid driven envelope material removal prediction model. Specifically, the method for establishing the model is as follows: Figure 1 As shown, the following steps are included:

[0152] The specific implementation steps are as follows:

[0153] S1: First, grinding and polishing experiments are carried out in the source domain and the target domain to obtain the material removal depth under different grinding and polishing parameter combinations, and the obtained data set is divided into the real data of the source domain and the real data of the target domain.

[0154] S2: Establish an envelope grinding and polishing material removal depth mechanism model, and use the least squares method to fit the parameters of the mechanism model.

[0155] S3: Generate a large amount of virtual simulation data in the source domain through the mechanism model, and construct a data set for subsequent model training.

[0156] S4: Establish a mechanism-data domain adaptive adversarial neural network, which includes a data feature extraction layer and a data regression layer. The network's loss function includes a source domain real data regression loss function, a source domain virtual simulation data regression loss function, a target domain real data regression loss function, and a domain classification loss function.

[0157] S5: Use Adam optimizer to train the model and update its parameters.

[0158] The polishing parameters in step S1 are the contact force F between the tool and the workpiece, the feed speed V of the tool, t , the contact speed between the abrasive belt and the workpiece V c , and the curvature radius R of the workpiece w Through the above different parameter combinations x=[F,V t ,V c ,R w ] The grinding and polishing experiments were carried out at different radii, and the grinding and polishing depth h under the corresponding grinding and polishing parameters was measured. The obtained data were divided into source domain real data and the target domain real data in The grinding and polishing parameters meet the following conditions:

[0159]

[0160] in, and are the minimum and maximum values ​​of F in the source domain, and V in the source domain c The minimum and maximum values ​​of and V in the source domain t The minimum and maximum values ​​of and R in the source domain w The minimum and maximum values ​​of .

[0161] The grinding and polishing parameters meet the following conditions:

[0162]

[0163] in, and is the minimum and maximum value of F in the target domain, and V in the target domain c The minimum and maximum values ​​of and V in the target domain t The minimum and maximum values ​​of and R in the target domain w The minimum and maximum values ​​of .

[0164] The envelope grinding and polishing material removal depth mechanism model in step S2 establishes the grinding and polishing parameters F, V t , V c , R w The relationship between the material removal depth h is as follows:

[0165]

[0166] Among them, u is the material removal coefficient, which is affected by the physical properties of the tool and the workpiece and is a constant value during the grinding and polishing process. In the actual experimental process, due to the existence of measurement error ε, the above formula can be expressed as:

[0167]

[0168] The values ​​of u and ε are fitted by combining the real source domain data obtained in step S1 with the least squares method. The specific method is as follows:

[0169] in,

[0170] in, are the contact force between the tool and the workpiece, the contact speed between the abrasive belt and the workpiece, the feed speed of the tool, the workpiece radius and the material removal depth in the i-th group of source domain real data. is the average value of m groups of real material removal depths in the source domain. c i The average value of . m is the number of real data in the source domain.

[0171] The large amount of mechanism data described in step S3 is generated by the mechanism model fitted in step S2. First, s groups of grinding and polishing parameter combinations are generated in the source domain. in is the contact force between the tool and the workpiece, the contact speed between the abrasive belt and the workpiece, the feed speed of the tool, the workpiece radius and the material removal depth in the i-th group of source domain virtual simulation data. The calculation method is as follows:

[0172]

[0173] The datasets described in S3 contain real data from the source domain Source domain virtual simulation data and the target domain real data It consists of three parts, the specific components are as follows:

[0174]

[0175] Among them, parameters s, m, and n are the number of source domain virtual data, data source domain real data, and target domain real data, which satisfy s>>m,n, m≈n, where the symbol >> represents much greater than, and the symbol ≈ represents approximately equal to. is the grinding and polishing parameter combination of the real data in the source domain, is the grinding and polishing parameter combination of the source domain virtual simulation data and is the grinding and polishing parameter combination of the real data in the target domain. It can be specifically expressed as:

[0176]

[0177] for The corresponding source domain virtual simulation material removal depth, for The corresponding source domain real material removal depth, for The corresponding target domain real material removal depth. It can be specifically expressed as:

[0178]

[0179] where d i is the data label of the i-th group of data, which can be defined in the form of one-hot encoding, which can be specifically expressed as:

[0180]

[0181] The mechanism in step S4 - the data domain adaptive adversarial neural network consists of a feature extraction layer and a data regression layer. The feature extraction layer has a feature extractor G f (x;θ f ), there are three data regressors and one domain classifier in the data regression layer: source domain virtual simulation data regressor G s,v (z;θ s,v ), source domain real data regressor G s,a (z;θ s,a ), target domain real data regressor G t (z;θ t ) and domain classifier G d (z;θ d ). Where x is the input of the network, z is the output of the feature extraction layer and serves as the input of the data regression layer, θ f is the network parameter of the feature extractor, θ s,v is the network parameter of the source domain virtual simulation data regressor, θ s,a is the network parameter of the source domain real data regressor, θt is the network parameter of the target domain real data regressor, θ d is the network parameter of the domain classifier. The output data of the network There are two parts, is the predicted value of material removal depth, is the probability prediction value of the data label. The specific expression is as follows:

[0182]

[0183] The loss function of the network in step S3 It consists of four parts: Source domain virtual simulation data loss function Source domain real data loss function Target domain real data loss function And domain classification loss function and The mean square error loss function is used for calculation, which is defined as follows:

[0184]

[0185] The domain classification loss function described The cross entropy loss function is used for calculation, which is calculated as follows:

[0186]

[0187] in, is the domain classification loss function of the source domain virtual simulation data, is the domain classification loss function of the real data in the source domain, is the loss function of the real data in the target domain. is the predicted probability of the data label of the i-th data, which is expressed as p1, p2, p3 are the i-th data [x, y] i is the probability of virtual simulation data in the source domain, the probability of real data in the source domain and the probability of real data in the target domain, which satisfies p1+p2+p3=1. The operation L(·) is defined as:

[0188]

[0189] Finally, the loss function of the entire network is It can be calculated as follows:

[0190]

[0191] Among them, λ s,a ,λ t and μ is and The weight parameter of . It is specifically expressed as:

[0192]

[0193] Among them, ρ l is the training process, l is the current training round number, and L is the maximum training round number. s,a is s,a The rate of change coefficient, γ t is t The rate of change coefficient, γ d is the rate of change coefficient of μ.

[0194] The specific implementation method of the training of the model described in step S5 is as follows:

[0195] (1) First, set the training parameters of the model, including the maximum number of training rounds L and the learning rate η, and use this to initialize the Adam optimizer. The Adam optimizer adjusts the learning rate of each parameter by calculating the first-order moment estimate and the second-order moment estimate of the gradient, thereby achieving efficient network training. Initialize the network parameters θ f ,θ s,v ,θ s,a ,θ t and θ d .

[0196] (2) Calculate by forward propagation of formula The corresponding material removal prediction depth and the probability prediction value of the data label. Then the loss function of the network is calculated by formulas (13)-(18).

[0197] (3) Back propagation through the Adam optimizer and update the network parameters θ f ,θ s,v ,θ s,a ,θ t and θ d .

[0198] (4) Repeat steps (2)-(3) until the number of training rounds reaches the maximum number of training rounds.

[0199] like Figure 2 As shown, the envelope grinding material removal depth prediction system provided by the embodiment of the present invention includes:

[0200] Dataset division module: Grinding and polishing experiments are conducted in the source domain and the target domain to obtain the material removal depth under different grinding and polishing parameter combinations, and the obtained data set is divided into the source domain real data and the target domain real data;

[0201] Mechanism model building module: used to build the envelope grinding and polishing material removal depth mechanism model, and use the least squares method to fit the parameters of the mechanism model;

[0202] Virtual simulation data generation module: used to generate a large amount of virtual simulation data in the source domain through the mechanism model and construct a data set for subsequent model training;

[0203] Adversarial neural network establishment module: Establish a mechanism-data domain adaptive adversarial neural network, which includes a data feature extraction layer and a data regression layer; the network's loss function includes a source domain real data regression loss function, a source domain virtual simulation data regression loss function, a target domain real data regression loss function, and a domain classification loss function;

[0204] Model training and parameter updating module: Adam optimizer is used to train the model and update its parameters.

[0205] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the envelope grinding material removal depth prediction method.

[0206] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the envelope grinding material removal depth prediction method.

[0207] An application embodiment of the present invention provides an information data processing terminal, which includes an envelope grinding material removal depth prediction system.

[0208] The envelope grinding and polishing proposed in the present invention is an envelope grinding material removal depth prediction model based on mechanism and small data hybrid drive, which can be used for predicting the material removal of the blade edge of aerospace blades and blade disk parts. Figure 3 As a specific embodiment of the method, Figure 3 (a) shows the contour of the blade edge of a certain type of aircraft engine disk before and after envelope grinding and polishing. Figure 3 (b) and Figure 3 (c) are respectively the prediction results and prediction errors of the material removal depth using the method proposed in the present invention. It can be seen from the figure that the method proposed in the present invention can well predict the material removal depth in the envelope grinding process, with a maximum prediction error of 10 microns and a mean prediction error of 5.08 microns. In order to illustrate the effectiveness of the method proposed in the present invention, this section verifies the relevant effects of the method through an envelope grinding and polishing calibration experiment. The grinding and polishing parameter combination table 1 designed in this experiment is shown below:

[0209] Table 1 Combination of grinding and polishing parameters for calibration experiment

[0210]

[0211] Calculate V using the data in the table t and V c The formula is as follows:

[0212]

[0213] Where R a is the radius of the belt sander shaft. This experiment uses a total of 108 groups of experimental data, 3*3*3*4, to verify the proposed material removal prediction model. At the same time, the mechanism model and the traditional data-driven model are used as comparative experiments. The experimental data are first divided into real data in the source domain and real data in the target domain. The data in the source domain are first fitted by the mechanism model using the least squares method, and then a large amount of virtual data is generated in the source domain through the mechanism model. Then the generated virtual data and real data are used as training data for the proposed envelope grinding material removal depth prediction model based on a hybrid drive of mechanism and small data. The prediction results of the proposed model, mechanism model and traditional data-driven model are shown in the figure. Figure 4 As shown in Table 2. From the results, it can be seen that the average error and maximum error of the proposed method in the source domain are reduced by 47.2% and 60.0% respectively compared with the traditional data model prediction, and the average error and maximum error are reduced by 5.0% and 29.5% respectively compared with the mechanism model prediction; the average error and maximum error of the proposed method in the target domain are reduced by 11.2% and 15.8% respectively compared with the traditional data model prediction, and the average error and maximum error are reduced by 22.8% and 6.7% respectively compared with the mechanism model prediction. The results show that the proposed model has a higher material removal depth prediction accuracy than the existing mechanism model and data-driven model, which further illustrates the effectiveness of the material removal prediction model proposed in the present invention.

[0214] Table 2 Comparison of model prediction errors

[0215]

[0216] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0217] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for predicting the depth of material removal by enveloping grinding, characterized in that: include: S1: Grinding and polishing experiments are carried out in the source domain and the target domain to obtain the material removal depth under different grinding and polishing parameter combinations, and the obtained data set is divided into the source domain real data and the target domain real data; S2: Establish an envelope grinding and polishing material removal depth mechanism model, and use the least squares method to fit the parameters of the mechanism model; S3: Generate a large amount of virtual simulation data in the source domain through the mechanism model, and build a data set for subsequent model training; S4: Establish a mechanism-data domain adaptive adversarial neural network, which includes a data feature extraction layer and a data regression layer; the network's loss function includes a source domain real data regression loss function, a source domain virtual simulation data regression loss function, a target domain real data regression loss function, and a domain classification loss function; S5: Use Adam optimizer to train the model and update its parameters; The envelope grinding and polishing material removal depth mechanism model in step S2 establishes the grinding and polishing parameters F, V t , V c , R w The relationship between the material removal depth and the specific expression is as follows: Among them, u is the material removal coefficient, which is affected by the physical properties of the tool and the workpiece. It is a constant value in the grinding and polishing process. In the actual experimental process, due to the existence of measurement error ε, the above formula can be expressed as: The values ​​of u and ε are fitted by combining the real source domain data obtained in step S1 with the least squares method. The specific method is as follows: in, Among them, F i s,a , are the contact force between the tool and the workpiece, the contact speed between the abrasive belt and the workpiece, the feed speed of the tool, the workpiece radius and the material removal depth in the i-th group of source domain real data, respectively. is the average value of the material removal depth in the source domain, c i The average value of , m is the number of real data in the source domain; The mechanism in step S4 - the data domain adaptive adversarial neural network consists of a feature extraction layer and a data regression layer. The feature extraction layer has a feature extractor G f (x;θ f ), there are three data regressors and one domain classifier in the data regression layer: source domain virtual simulation data regressor G s,v (z;θ s,v ), source domain real data regressor G s,a (z;θ s,a ), target domain real data regressor G t (z;θ t ) and domain classifier G d (z;θ d ), where x is the input of the network, z is the output of the feature extraction layer and serves as the input of the data regression layer, θ f is the network parameter of the feature extractor, θ s,v is the network parameter of the source domain virtual simulation data regressor, θ s,a is the network parameter of the source domain real data regressor, θ t is the network parameter of the target domain real data regressor, θ d is the network parameter of the domain classifier; the output data of the network There are two parts, is the predicted value of material removal depth, is the probability prediction value of the data label; the specific expression is as follows:

2. The envelope grinding material removal depth prediction method according to claim 1, characterized in that: The polishing parameters in step S1 are the contact force F between the tool and the workpiece, the feed speed V of the tool, t , the contact speed between the abrasive belt and the workpiece V c , and the curvature radius R of the workpiece w ; Through the above different parameter combinations x=[F,V t ,V c ,R w ] The grinding and polishing experiments were carried out at different radii, and the grinding and polishing depth h under the corresponding grinding and polishing parameters was measured; the obtained data was divided into source domain real data and the target domain real data in The grinding and polishing parameters meet the following conditions: in, and are the minimum and maximum values ​​of F in the source domain, and V in the source domain c The minimum and maximum values ​​of and V in the source domain t The minimum and maximum values ​​of and R in the source domain w The minimum and maximum values ​​of ; The grinding and polishing parameters meet the following conditions: in, and is the minimum and maximum value of F in the target domain, and V in the target domain c The minimum and maximum values ​​of and V in the target domain t The minimum and maximum values ​​of and R in the target domain w The minimum and maximum values ​​of .

3. The envelope grinding material removal depth prediction method according to claim 1, characterized in that: In step S3, a large amount of mechanism data is generated by the mechanism model fitted in step S2. First, s groups of grinding and polishing parameter combinations are generated in the source domain. Among them, F i s,v , is the contact force between the tool and the workpiece, the contact speed between the abrasive belt and the workpiece, the feed speed of the tool, the workpiece radius and the material removal depth in the i-th group of source domain virtual simulation data; The calculation method is as follows:

4. The envelope grinding material removal depth prediction method according to claim 1, characterized in that: The dataset in S3 contains the real data in the source domain Source domain virtual simulation data and the target domain real data It consists of three parts, the specific components are as follows: Among them, the parameters s, m, and n are the numbers of virtual data in the source domain, real data in the data source domain, and real data in the target domain, which satisfy s>>m,n, and m≈n, where the symbol >> represents much greater than, and the symbol ≈ represents approximately equal to; is the grinding and polishing parameter combination of the real data in the source domain, is the grinding and polishing parameter combination of the source domain virtual simulation data and is the grinding and polishing parameter combination of the real data in the target domain; it can be specifically expressed as: for The corresponding source domain virtual simulation material removal depth, for The corresponding source domain real material removal depth, for The corresponding target domain real material removal depth can be specifically expressed as: where d i is the data label of the i-th group of data, which can be defined in the form of one-hot encoding, which can be specifically expressed as:

5. The envelope grinding material removal depth prediction method according to claim 1, characterized in that: The loss function of the network in step S4 It consists of four parts: Source domain virtual simulation data loss function Source domain real data loss function Target domain real data simulation function And domain classification loss function and The mean square error loss function is used for calculation, which is defined as follows:

6. The envelope grinding material removal depth prediction method according to claim 1, characterized in that: Domain Classification Loss Function The cross entropy loss function is used for calculation, which is calculated as follows: in, is the domain classification loss function of the source domain virtual simulation data, is the domain classification loss function of the real data in the source domain, is the loss function of the real data in the target domain; is the predicted probability of the data label of the i-th data, which is expressed as p1, p2, p3 are the i-th data [x, y] i is the probability of virtual simulation data in the source domain, the probability of real data in the source domain and the probability of real data in the target domain, which satisfies p1+p2+p3=1; the operation L(·) is defined as: Finally, the loss function of the entire network is It can be calculated as follows: Among them, λ s,a ,λ t and μ is and The weight parameter is: Among them, ρ l is the training process, l is the number of current training rounds, and L is the maximum number of training rounds; γ s,a is s,a The rate of change coefficient, γ t is t The rate of change coefficient, γ d is the rate of change coefficient of μ.

7. The envelope grinding material removal depth prediction method according to claim 1, characterized in that: The specific implementation method of model training in step S5 is as follows: (1) First, set the training parameters of the model, including the maximum number of training rounds L and the learning rate η, and use this to initialize the Adam optimizer. The Adam optimizer adjusts the learning rate of each parameter by calculating the first-order moment estimate and the second-order moment estimate of the gradient, thereby achieving efficient network training and initializing the network parameters θ f ,θ s,v ,θ s,a ,θ t and θ d ; (2) Calculate by forward propagation of formula The corresponding material removal prediction depth and the probability prediction value of the data label are then used to calculate the network loss function using formulas (52)-(57); (3) Back propagation through the Adam optimizer and update the network parameters θ f ,θ s,v ,θ s,a ,θ t and θ d ; (4) Repeat steps (2)-(3) until the number of training rounds reaches the maximum number of training rounds.

8. An envelope grinding material removal depth prediction system for implementing the envelope grinding material removal depth prediction method according to any one of claims 1 to 7, characterized in that: include: Dataset division module: Grinding and polishing experiments are conducted in the source domain and the target domain to obtain the material removal depth under different grinding and polishing parameter combinations, and the obtained data set is divided into the source domain real data and the target domain real data; Mechanism model building module: used to build the envelope grinding and polishing material removal depth mechanism model, and use the least squares method to fit the parameters of the mechanism model; Virtual simulation data generation module: used to generate a large amount of virtual simulation data in the source domain through the mechanism model and construct a data set for subsequent model training; Adversarial neural network establishment module: Establish a mechanism-data domain adaptive adversarial neural network, which includes a data feature extraction layer and a data regression layer; the network's loss function includes a source domain real data regression loss function, a source domain virtual simulation data regression loss function, a target domain real data regression loss function, and a domain classification loss function; Model training and parameter updating module: Adam optimizer is used to train the model and update its parameters.

Images