Mechanism-data hybrid driven machining deformation online prediction method

Abstract

The invention discloses a mechanism-data hybrid drive machining deformation online prediction method, which comprises the following steps: establishing a residual stress field deformation mechanism model, and calculating the deformation amount of a discrete unit of a part according to an input residual stress value of the discrete unit of the part; a multi-layer perceptron model is established, and the whole deformation of the part is output according to the discrete unit deformation of the part; establishing a deep learning network model, and outputting a hidden layer weight of the multi-layer perceptron model according to the residual stress value of the discrete unit of the part; training the deep learning network model; multi-source data in the machining process are collected online and input into the mechanism model and the deep learning model, a machining deformation prediction result is output, and the influence degree of local stress on overall deformation is quantified. According to the method, the milling deformation of the thin-wall part can be predicted on line, the influence degree of local stress on overall deformation is quantified, and a reference can be provided for a machining deformation control strategy based on residual stress field regulation and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of online prediction of machining deformation, and in particular to a mechanism-data hybrid driven machining deformation online prediction method. BACKGROUND

[0002] Thin-walled parts have the advantages of compact structure and lightweight, and they have played a crucial role in the field of aerospace. With the continuous improvement of the requirements for part precision and performance in modern industry, the machining quality of thin-walled parts has become one of the key factors restricting the development of high-end manufacturing. However, due to the structural characteristics of thin-walled parts, such as thin walls and weak rigidity, they are easily disturbed by various factors during machining, such as cutting force, clamping force and residual stress, etc., which leads to non-negligible machining deformation. This deformation not only reduces the geometric precision and surface quality of the parts, but also causes material waste and a decrease in machining efficiency. Therefore, how to effectively predict and control the machining deformation of thin-walled parts has become a core problem to improve machining quality and efficiency.

[0003] At present, the conventional deformation prediction methods mainly include analytical method and finite element method. The analytical method can describe the influence of residual stress on local deformation during machining by establishing a physical mechanism model; the finite element method realizes the prediction of deformation through numerical simulation. However, these methods have obvious limitations. The residual stress field has global distribution characteristics, and a small change in its local state may lead to significant differences in the overall deformation, and the existing methods are difficult to accurately capture this complex local-global relationship. Secondly, the mechanism model usually relies on a large number of simplifications and assumptions, making it difficult to accurately describe the deformation behavior under the coupling action of multiple factors. In addition, the offline deformation prediction method cannot respond to the dynamic changes in the machining process in real time, resulting in limited prediction accuracy and practicality.

[0004] In recent years, with the rapid development of big data, cloud computing and machine learning technology, intelligent manufacturing systems are gradually transforming towards intelligence and digitization. Under this background, online prediction technology of machining quality has attracted widespread attention, and its core goal is to realize real-time monitoring of physical state and deformation prediction during machining, so as to dynamically adjust the machining parameters to improve the prediction accuracy and control effect. Machine learning technology provides a new solution for online prediction of machining deformation due to its strong non-linear fitting and data processing capabilities. However, although deep learning can theoretically approximate any complex function relationship, pure data-driven prediction models still have significant limitations in practical applications. For example, under unknown working conditions, the prediction accuracy of the model often decreases significantly, which is mainly due to the insufficient diversity and coverage of the training data, making it difficult for the model to generalize to new machining conditions. In addition, pure data-driven models lack a deep understanding of the physical mechanism, making it difficult to explain the complex relationship between local stress distribution and overall deformation, which further limits their application value in actual production. SUMMARY

[0005] The purpose of the present application is to propose a mechanism-data hybrid driving online prediction method of machining deformation, which can online predict the milling machining deformation of thin-walled parts, quantify the influence degree of local stress on the overall deformation, and provide reference for the machining deformation control strategy based on residual stress field regulation.

[0006] To achieve the above technical purpose, the technical scheme adopted by the present application is:

[0007] A mechanism-data hybrid driving online prediction method of machining deformation, the method comprising the following steps:

[0008] S1, carrying out thin-walled part milling machining test, layer cutting for several thin-walled parts, collecting cutting force, clamping force, acceleration, overall deformation and maximum deformation position of thin-walled parts during the machining process of each thin-walled part;

[0009] S2, sequentially carrying out finite element mesh division for each thin-walled part to obtain several part discrete units; obtaining the internal residual stress state of the thin-walled part according to the collected cutting force, clamping force and acceleration, obtaining the residual stress value of each part discrete unit; establishing a residual stress field deformation mechanism model, calculating the deformation of each part discrete unit according to the input residual stress value of each part discrete unit;

[0010] S3, integrating the deformation of all part discrete units corresponding to each thin-walled part, the distance between the center point of the part discrete unit and the maximum deformation position, and the overall deformation to generate a first training data set and a first test set; taking the deformation of all part discrete units corresponding to any one thin-walled part as the input data of the multilayer perception machine model, taking the distance vector composed of the distance between the center point of all part discrete units and the maximum deformation position as the initial hidden layer weight vector of the multilayer perception machine model, and taking the overall deformation of the thin-walled part as the output data, the first training data set and the first test set are used to train and verify the multilayer perception machine model; output and save the hidden layer weight vector of the trained multilayer perception machine model;

[0011] S4, introducing a residual module into a U-net model to construct a deep learning network model; integrating the deformation of all part discrete units of each thin-walled part and the hidden layer weight vector of the corresponding multilayer perception machine model to generate a second training data set and a second test set; taking the deformation of all part discrete units of each thin-walled part as the input data of the deep learning network model, and taking the hidden layer weight vector of the corresponding multilayer perception machine model as the output data of the deep learning network model, the second training data set and the second test set are used to train and verify the deep learning network model;

[0012] S5, for the thin-walled part being milled, the cutting force, clamping force and acceleration of the thin-walled part in the machining process are collected on-line to obtain the residual stress values of each part discrete unit corresponding to the thin-walled part, and then the deformation of all part discrete units corresponding to the thin-walled part is calculated by using the residual stress field deformation mechanism model; the deformation of all part discrete units corresponding to the thin-walled part calculated is input into the deep learning network model trained to output the hidden layer weight vector of the multi-layer perception model; the hidden layer weight vector of the multi-layer perception model output and the deformation of all part discrete units corresponding to the thin-walled part are input into the multi-layer perception model together to output the overall deformation of the thin-walled part.

[0013] Step S1 further comprises:

[0014] A thin-walled part milling experiment is carried out to remove material in a layer removal manner, and the cutting force and acceleration in the thin-walled part milling process are recorded; during the thin-walled part milling process, a thin film pressure sensor is used to collect clamping force data in the thin-walled part milling process.

[0015] A laser displacement sensor is fixed at the bottom of the thin-walled part, a line laser beam is aligned with the weak rigidity area of the part, and deformation data in the thin-walled part milling process is collected, including overall deformation and maximum deformation position.

[0016] Further, in step S2, the thin-walled part is divided into m layers, the deflection of the neutral layer of the thin-walled part is set to zero, a mechanism model of processing deformation caused by residual stress after material removal is established by using thin plate bending theory, and a residual stress field deformation mechanism model is formed in combination with cantilever beam theory:

[0017] ;

[0018] In the formula, represents the deflection change value caused by the change of the residual stress field after the i-th layer of material removal in the part discrete unit; represents the deflection change value caused by the concentrated load F applied by the tool to the part during the machining process of the thin-walled part according to the cantilever beam theory; represents the deformation of the part discrete unit.

[0019] Further, the deflection change value of the center point position caused by the change of the residual stress field after the i-th layer of material removal in the part discrete unit is calculated by using the following formula :

[0020] ;

[0021] In the formula, t is the thickness of each layer of the thin-walled part, i is the current number of layers removed, h i and h i+1respectively, is the Poisson's ratio of the thin-walled part plate, E is the elastic modulus, σ x,i-1,i is the residual stress in the X direction of the i-th layer of the thin-walled part after removing the i-1-th layer, σ y,i-1,i is the residual stress in the Y direction of the i-th layer of the thin-walled part after removing the i-1-th layer, σ y,i-1,1 is the residual stress in the Y direction of the i-th layer of the thin-walled part after removing the i-1-th layer, σ x,i-1,1 is the residual stress in the X direction of the i-th layer of the thin-walled part after removing the i-1-th layer.

[0022] the deflection change value of the center point position is the deflection change value caused by the change of the residual stress field of the material after removing the i-th layer in the part discrete unit .

[0023] Further, the deflection change value w B caused by the concentrated load F applied by the tool to the part during the processing of the thin-walled part includes the following steps:

[0024] The part discrete unit is regarded as a cantilever beam, and the contact process of the tool and the cantilever beam is regarded as the application of a concentrated load F to the end of the cantilever beam, so that the cantilever beam as a whole is bent and deformed, and the bending moment M on any cross section is:

[0025] ;

[0026] In the formula, represents the concentrated load applied by the tool to the part, represents the distance between the bending moment section and the free end, and l is the length of the cantilever beam; the abscissa of the bending moment section B is substituted to obtain the deflection change value :

[0027] ;

[0028] In the formula, I is the section moment of inertia, represents the elastic modulus.

[0029] Further, in step S3, the multi-layer perception machine model includes an input layer, a hidden layer, a full connection layer and an output layer connected in turn;

[0030] Wherein, the input layer of the multi-layer perception machine model is a one-dimensional vector data, and the length is equal to the number n of part discrete units; the number of neurons of the hidden layer is set to the number n of part discrete units, and the initial hidden layer weight vector is a distance vector composed of the distance between each part discrete unit and the maximum deformation position; the neuron activation function adopted by the full connection layer is LeakyReLU function.

[0031] Further, in step S4, a residual module is introduced into the U-net model to construct a deep learning network model, and the input and output of the deep learning network model are one-dimensional vectors, wherein the input of the deep learning network model is the deformation of all part discrete units of the thin-walled part, and the output is the hidden layer weight vector w of the trained multi-layer perceptron model n ;

[0032] The deep learning network model comprises an encoding module and a decoding module; the encoding module comprises a first residual module, a first pooling layer, a second residual module, a second pooling layer, a third residual module and a third pooling layer, and the encoding module extracts features from the input deformation of the part discrete units; the decoding module comprises a convolution layer, a fourth residual module, a first transposed pooling layer, a fifth residual module, a second transposed pooling layer, a sixth residual module, a third transposed pooling layer, a seventh residual module, a fourth transposed pooling layer, an eighth residual module and a ninth residual module, and the decoding module decodes the features extracted by the encoding module and outputs the predicted hidden layer weight vector w n ;

[0033] The residual module comprises a 1×1 convolution kernel, a 3×1 convolution kernel and a LeakyReLu activation function connected in sequence.

[0034] Further, in step S5, mean square error (MSE) is used as the loss function during training of the deep learning network model:

[0035] ;

[0036] wherein B s is the batch size, v=1,2,…, B s， is the hidden layer weight vector of the multi-layer perceptron model in the second training set, is the hidden layer weight vector predicted for the data in the second training set; the deep learning network model is evaluated by the following evaluation function:

[0037] ;

[0038] wherein N is the number of training periods, is the hidden layer weight vector of the multi-layer perceptron model in the second test set, is the hidden layer weight vector of the multi-layer perceptron model predicted for the data in the second test set.

[0039] Further, the method further comprises:

[0040] inputting the deformation of the part discrete units into the deep learning network model to output the hidden layer weight vector w of the multi-layer perceptron model n; the hidden layer weight vector w n As a quantitative result of the influence degree of local stress on overall deformation.

[0041] Further, the extracted hidden layer weight vector w n And the center point coordinates of the part discrete unit correspond, according to the HSV color value, the hidden layer weight vector w n Assign color, draw a three-dimensional cloud chart combined with the corresponding coordinates, and visualize the quantitative results of the influence degree of local stress on overall deformation.

[0042] Compared with the prior art, the beneficial effects of the present application are as follows:

[0043] The quantitative method of the influence degree of local stress on overall deformation of the present application combines the advantages of mechanism model and data-driven model, uses the learning ability of data-driven model to make up for the shortcomings of mechanism model, and provides theoretical support for data-driven model by using the prediction data set to provide workpiece machining process deformation related features, and realizes online prediction of machining deformation. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The flow chart of the mechanism-data hybrid driven machining deformation online prediction method of the present application; Figure 2 The structural diagram of the residual stress field deformation mechanism model; Figure 3 The structure diagram of the MLP model; Figure 4 The structure diagram of the URes-net model; Figure 5 The schematic diagram of the machining system and sensor arrangement; Figure 6a The training loss diagram of the MLP model; Figure 6b The training loss diagram of the URes-net model; Figure 7 The machining deformation prediction result schematic diagram; wherein, (a) corresponds to the first group of cutting machining data, (b) corresponds to the second group of cutting machining data; Figure 8 The cloud chart of the quantitative results of the influence degree of local stress on overall deformation; Mark: 1, wireless rotary dynamometer, 2, thin film pressure sensor, 3, acceleration sensor, 4, laser displacement sensor, 5, deformation measurement area. DETAILED DESCRIPTION

[0045] The embodiments of the present application are further described in detail below in combination with the drawings.

[0046] AsFigure 1 As shown, the application discloses a mechanism-data hybrid driving online prediction method for machining deformation, which comprises the following steps:

[0047] S1, a milling test of thin-walled parts is carried out, and the cutting force, clamping force, acceleration, overall deformation and maximum deformation position of each thin-walled part in the machining process are collected by layer-by-layer cutting of a plurality of thin-walled parts;

[0048] S2, finite element mesh division is carried out for each thin-walled part in turn to obtain a plurality of part discrete units; the internal residual stress state of the thin-walled part is predicted according to the collected cutting force, clamping force and acceleration to obtain the residual stress value of each part discrete unit; a residual stress field deformation mechanism model is established, and the deformation of each part discrete unit is calculated according to the input residual stress value of each part discrete unit;

[0049] S3, the deformation of all part discrete units corresponding to each thin-walled part, the distance between the center point of the part discrete unit and the maximum deformation position, and the overall deformation are integrated to generate a first training data set and a first test set; the deformation of all part discrete units corresponding to any thin-walled part is taken as the input data of the multilayer perceptron model, the distance vector composed of the distance between the center point of all part discrete units and the maximum deformation position is taken as the initial hidden layer weight vector of the multilayer perceptron model, and the overall deformation of the thin-walled part is taken as the output data; the first training data set and the first test set are used to train and verify the multilayer perceptron model; the hidden layer weight vector of the trained multilayer perceptron model is output and saved;

[0050] S4, a residual module is introduced into the U-net model to construct a deep learning network model; the deformation of all part discrete units of each thin-walled part and the hidden layer weight vector of the corresponding multilayer perceptron model are integrated to generate a second training data set and a second test set; the deformation of all part discrete units of each thin-walled part is taken as the input data of the deep learning network model, and the hidden layer weight vector of the corresponding multilayer perceptron model is taken as the output data of the deep learning network model; the second training data set and the second test set are used to train and verify the deep learning network model;

[0051] S5, for the thin-walled part being milled, the cutting force, clamping force and acceleration of the thin-walled part in the machining process are collected on-line to obtain the residual stress values of each part discrete unit corresponding to the thin-walled part, and then the deformation of all part discrete units corresponding to the thin-walled part is calculated by using the residual stress field deformation mechanism model; the calculated deformation of all part discrete units corresponding to the thin-walled part is input into the trained deep learning network model, and the hidden layer weight vector of the multi-layer perception model is output; the output hidden layer weight vector of the multi-layer perception model and the deformation of all part discrete units corresponding to the thin-walled part are input into the multi-layer perception model together, and the overall deformation of the thin-walled part is output.

[0052] In step S1, the material is removed in layers during the milling of the thin-walled part, and the data collection process includes:

[0053] A milling test of the thin-walled part is carried out to remove the material in layers, and the cutting force and acceleration during the milling of the thin-walled part are recorded; during the milling of the thin-walled part, a film pressure sensor is used to collect the clamping force data during the milling of the thin-walled part; a laser displacement sensor is fixed at the bottom of the thin-walled part, and a line laser beam is aligned with the weak rigidity area of the part to collect deformation data during the milling of the thin-walled part, including overall deformation and maximum deformation position.

[0054] In step S2, the internal residual stress state of the thin-walled part is predicted by using an IncepU-net network model. Specifically, the convolution layer in the U-net network model is replaced by an Inception module with deeper feature extraction to construct an IncepU-net network model, the input data of the IncepU-net network model is the cutting force, clamping force and acceleration, and the output data is the internal residual stress state of the thin-walled part, and then the residual stress value of each part discrete unit is obtained. The processing process of the IncepU-net network model on the input data set is as follows: input data-two inception modules-pooling layer-two inception modules-pooling layer-two inception modules-pooling layer-two inception modules-pooling layer-two inception modules-pooling layer, which completes the coding stage of the data; and then through the convolution layer-two transpose inception modules-transposed pooling layer-two transpose inception modules-transposed pooling layer-two transpose inception modules-transposed pooling layer-two transpose inception modules-transposed pooling layer-two transpose inception modules-full connection layer, which completes the decoding stage of the data.

[0055] In step S2, the workpiece is divided into m layers, a mechanism model of the machining deformation caused by the residual stress after material removal is established by using the thin plate bending theory, and a discrete unit deformation mechanism model is formed in combination with the cantilever beam theory.Figure 2 is the structural diagram of the deformation mechanism model of the residual stress field. Wherein, l , b , h 1 is the length, width and height of the cantilever beam respectively. At the same time, h 1 is also the remaining thickness of the material before removing the first layer, h 2 is the remaining thickness of the material before removing the second layer. L , W , H are the length, width and height of the discrete unit respectively. t is the thickness of the layer removed material. u represents the distance between the bending moment section and the free end. σ x is the stress in the x direction, σ y is the stress in the y direction, F is the concentrated load applied to the part. Specifically, in a part discrete unit, the deflection change caused by the change of the residual stress field after removing a layer of material (the i-th layer) is RSi ; according to the cantilever beam theory, the deflection change caused by the concentrated load F applied to the part by the tool during the machining of the thin-walled part is B ; combining the two theories, the deformation mechanism model of the cantilever beam discrete unit is shown in the figure, and the calculation formula is as follows:

[0056] ;

[0057] In order to facilitate calculation, in this embodiment, the deflection change value of the center point (x, y) is taken as the deflection change value of the part discrete unit, and the deflection change of the center point (x, y) caused by the change of the residual stress field after removing a layer of material (the i-th layer) is , which is expressed as the following formula:

[0058] ;

[0059] In the formula, t is the thickness of each layer of the workpiece material, i is the number of layers currently removed, h i is the thickness of the workpiece before removing the i-th layer, μ is the Poisson's ratio of the material, E is the elastic modulus, σ x,i-1,i is the residual stress in the X direction of the i-th layer after removing the i-1-th layer, σ y,i-1,i is the residual stress in the Y direction of the i-th layer after removing the i-1-th layer, σ y,i-1,1 is the residual stress in the Y direction of the i-th layer after removing the i-1-th layer, σ x,i-1,1 is the residual stress in the X direction of the i-th layer after removing the i-1-th layer.

[0060] In step S2, the part discrete unit is regarded as a cantilever beam, and the contact process of the tool with the cantilever beam is regarded as a concentrated load F applied to the end of the cantilever beam, so that the cantilever beam as a whole is bent and deformed, and the bending moment M in any cross section is:

[0061] ;

[0062] In the formula, u represents the distance between the bending moment section and the free end, and l is the length of the cantilever beam; the abscissa u=l of the bending moment section B is brought in to obtain the deflection calculation formula, and the following is obtained:

[0063] .

[0064] In step S3, the structure of the MLP model (multilayer perceptron model) is a sequentially connected input layer, a hidden layer, a fully connected layer and an output layer, and the MLP model structure is as shown in Figure 3 In this embodiment, the length of the one-dimensional vector data of the input layer of the multilayer perceptron model is equal to the number n of part discrete units, the number of neurons of the hidden layer is equal to the number n of part discrete units, the initial hidden layer weight of the multilayer perceptron model is a distance vector composed of the distance of each part discrete unit to the maximum deformation position, the neuron activation function of the fully connected layer is a LeakyReLU function, and the output of the function is S(). The deformation amount of each part discrete unit corresponding to the thin-walled part, the distance between the center point of the part discrete unit and the maximum deformation position, and the overall deformation amount are integrated to generate a first training data set. The deformation amount of the n part discrete units corresponding to any one thin-walled part in the first training data set constitutes a one-dimensional vector x n As the input data of the multilayer perceptron model, the distance vector composed of the distance between the center point of the part discrete unit and the maximum deformation position is used as the initial hidden layer weight of the multilayer perceptron model, and the overall deformation amount corresponding to the thin-walled part is used as the output data of the multilayer perceptron model. Figure 3 In the formula, matrix W1 is the weight from the input layer to the hidden layer, w nn represents the weight of the n-th neuron of the hidden layer corresponding to the deformation amount of the n-th unit in the input layer to hidden layer calculation; matrix b1 is the bias of the input layer to hidden layer calculation; matrix W2 is the weight from the hidden layer to the output layer, represents the weight of the n-th neuron of the hidden layer corresponding to the output layer dimension calculation; b2 and are the bias of the hidden layer neuron corresponding to the output layer dimension calculation; W3 is the weight of the output parameter of the output layer, which only contains one ; b3 is the bias of the output parameter of the output layer. Finally, the multilayer perceptron model is trained, and the hidden layer weight vector of the trained multilayer perceptron model is output and saved.

[0065] In step S4, based on the U-net model, a residual block structure is introduced to establish a deep learning network model (URes-net model); the input of the deep learning network model is a one-dimensional vector x composed of the deformation amounts of all discrete units of the parts. n The output is the hidden layer weight vector w of the MLP model. n Similarly, it is a one-dimensional vector. The URes-net model takes a one-dimensional vector x as input. n The processing steps are as follows: Input data - one residual module - pooling layer - one residual module - pooling layer - one residual module - pooling layer - one residual module - pooling layer, thus completing the data encoding stage (i.e., feature extraction); then through a convolutional layer - one residual module - transposed pooling layer - one residual module - transposed pooling layer - one residual module - transposed pooling layer - one residual module - transposed pooling layer - one residual module - one residual module, thus completing the data decoding stage and outputting the predicted data, i.e., the hidden layer weight vector w of the MLP model. n For the structure of deep learning network models, see Figure 4 The residual module is configured as follows: a 1×1 convolutional kernel, a 3×1 convolutional kernel, and a LeakyReLu activation function.

[0066] The training process for the URes-net model is as follows:

[0067] Step A1: Create a second training dataset. Specifically, first, use an MLP model to collect hidden layer weight vectors, and combine them with the deformation of discrete elements of the part output by the residual stress field deformation mechanism model to form a second training dataset and a second test set, which are used to train the URes-net model.

[0068] Step A2: Initialize the hyperparameters of the URes-net model;

[0069] Step A3: Read the data from the second training dataset and calculate the training loss;

[0070] Step A4: Read the second test set data and calculate the test loss;

[0071] Step A5: Repeat steps A3 and A4 until the number of iterations meets the preset maximum number of iterations.

[0072] The URes-net model is trained using a CNN kernel in PyTorch, implemented in Python and CUDA. The mean squared error is used as the loss function for the entire URes-net model.

[0073] ;

[0074] Among them, B s This refers to the batch size, v=1,2,…,B s， is the hidden layer weight vector of the multilayer perceptron model in the second training data set, is the hidden layer weight vector predicted for the data in the second training data set; the deep learning network model is evaluated by the following evaluation function:

[0075]

[0076] wherein N is the number of training periods, is the hidden layer weight vector of the multilayer perceptron model in the second test set, is the hidden layer weight vector of the multilayer perceptron model predicted for the data in the second test set.

[0077] As a preferred example, the application can also quantify and visualize the influence degree of local stress on overall deformation. The specific steps are: (1) input the discrete element deformation data of the part into the URes-net model; (2) output the hidden layer weight vector w n of the MLP model; (3) save the hidden layer weight vector w n as the quantification result of the influence degree of local stress on overall deformation. Preferably, the extracted hidden layer weight vector w n is corresponding to the center point coordinates of the discrete elements of the part, and the color is given to w n according to the HSV color value, and a three-dimensional cloud chart is drawn combined with the corresponding coordinates to realize the visualization of the quantification result of the influence degree of local stress on overall deformation.

[0078] Example

[0079] This embodiment takes a thin-walled blade-shaped part as the object, the material is aluminum alloy, and the machining is carried out on a WFL-M35 machine tool. The tool is a Walter four-blade carbide end mill. For each parameter of the deformation measurement region 5, the cutting force is measured by a wireless rotary dynamometer 1, the clamping force is measured by a thin film pressure sensor 2, the acceleration signal is measured by a PCB three-axis acceleration sensor 3, and the deformation is measured by a Keyence laser displacement sensor 4. The machining system and sensor arrangement are shown in Figure 5 . The training and testing of the deep learning network model are carried out on a workstation, and the workstation environment is as follows: Intel Core i9-13900KF CPU, NVIDIA GEFORCE RTX 4080 GPU, 128GB RAM, running Windows 10, Python 3.9 and Pycharm.

[0080] ​A total of 1000 machining data, including cutting force, acceleration, clamping force, spindle power and deformation, were collected from finite element simulation model and machining test. The data set was expanded to 40000 by data rotation, Gaussian noise and data enhancement. In the subsequent training process of MLP model and deep learning network model, the training data used by the two were different, but were based on the data set. Preferably, in the training process of the two models, the sample data can be divided into training set and test set according to the division principle of 70% and 30%.

[0081] The ultrasonic stress measurement method is used to measure the internal stress field of the thin-walled part. After measuring the stress values at multiple positions, the stress field is reconstructed by importing the finite element model to obtain the residual stress field distribution state before finishing. The initial stress field data and the collected machining process signals are input into the IncepU-net model to obtain the residual stress field distribution data of the thin-walled part during machining.

[0082] The residual stress field distribution data is input into the residual stress deformation mechanism model to obtain the deformation data r of each discrete element of the thin-walled part. i According to the position of the maximum deformation collected by the laser displacement sensor in the actual machining process, the distance between the center point of each discrete element of the part and the maximum deformation position is calculated, and the distance value is saved to d i . In this way, the first training set D M {r i , d i , dmax i} is formed for MLP training, and dmax i is the overall deformation of the thin-walled part.

[0083] The MLP model is built, d i is set as the hidden layer weight of the MLP model initialization, the input is r i , and the label is d maxi . The hyperparameters are set as follows: batch size is 32; learning rate is 0.0001; training iteration number is 10. Each data is trained once for the MLP model, and the training loss is shown in Figure 6a . After the MLP model is trained, its weight w i is extracted. r i and w i are combined to form the second training set W U {r i , w i}, which is used to train the Ures-net model. The total model training time is about 22 hours.

[0084] The Ures-net model is built, the input is r i , and the output is w iHyperparameter settings: learning rate 0.0001, batch size 64, number of iterations 100, optimizer Adam. The Ures-net model training loss is as follows: Figure 6b As shown. The model training time is approximately 1 day and 13 hours.

[0085] Two sets of cutting data were taken from the validation set. The mechanism-data driven method proposed in this invention was used to predict the machining deformation, and the results were compared with the deformation data collected by the laser displacement sensor. The comparison results are as follows: Figure 7 As shown, where, Figure 7 (a) in the text corresponds to the first set of cutting data. Figure 7 (b) in the figure corresponds to the second set of cutting data. During the machining process, the tool path is from the top of the thin-walled part to the root of the thin-walled part (the root is fixed by the caliper). The infeed section has poor rigidity, and the milling deformation increases at the beginning of the machining. As the machine reaches the root with better rigidity, the deformation first decreases and then tends to stabilize. Figure 7 As can be seen, as processing progresses, the predicted deformation amount by the mechanism-data driven method proposed in this invention exhibits a consistent trend with the actual deformation amount. Meanwhile, Figure 7 The bars at the bottom of the chart represent the deviation between the predicted and actual values. Statistically, the average prediction accuracy of the two sets of data is 98.41%. The maximum deviation is 1.8 × 10⁻⁶. -2 The proposed model exhibits good prediction accuracy. Furthermore, by extracting the weights of the MLP model and combining them with element coordinates to create a 3D contour plot, the influence of stress distribution on overall deformation is expressed. Figure 8 As shown.

[0086] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0087] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A mechanism-data hybrid driven online prediction method for machining deformation, characterized in that, The method includes the following steps: S1. Conduct milling tests on thin-walled parts. Perform layered cutting on several thin-walled parts and collect the cutting force, clamping force, acceleration, overall deformation and the location of the maximum deformation of each thin-walled part during the machining process. S2, perform finite element mesh generation for each thin-walled part in sequence to obtain several discrete elements of the part; predict the internal residual stress state of the thin-walled part based on the collected cutting force, clamping force and acceleration, and obtain the residual stress value of each discrete element of the part; establish a residual stress field deformation mechanism model, and calculate the deformation amount of each discrete element of the part based on the input residual stress value of each discrete element of the part. S3. Integrate the deformation of all discrete units of each thin-walled component, the distance between the center point of each discrete unit and the maximum deformation position, and the overall deformation to generate the first training dataset and the first test set. Use the deformation of all discrete units of any thin-walled component as the input data of the multilayer perceptron model, the distance vector formed by the distance between the center point of each discrete unit and the maximum deformation position as the initial hidden layer weight vector of the multilayer perceptron model, and the overall deformation of the thin-walled component as the output data. Use the first training dataset and the first test set to train and validate the multilayer perceptron model. Output and save the hidden layer weight vector of the trained multilayer perceptron model. S4. Introduce a residual module into the U-net model to construct a deep learning network model; integrate the deformation of all discrete units of each thin-walled component with the hidden layer weight vector of the corresponding multilayer perceptron model to generate a second training dataset and a second test set; use the deformation of all discrete units of each thin-walled component as the input data of the deep learning network model, and the hidden layer weight vector of the corresponding multilayer perceptron model as the output data of the deep learning network model; use the second training dataset and the second test set to train and validate the deep learning network model. S5. For a thin-walled part being milled, the cutting force, clamping force, and acceleration of the thin-walled part during the machining process are collected online to obtain the residual stress value of each discrete unit of the thin-walled part. Then, the deformation of all discrete units of the thin-walled part is calculated using the residual stress field deformation mechanism model. The calculated deformation of all discrete units of the thin-walled part is input into the trained deep learning network model, and the hidden layer weight vector of the multilayer perceptron model is output. The output hidden layer weight vector of the multilayer perceptron model and the deformation of all discrete units of the thin-walled part are input together into the multilayer perceptron model to output the overall deformation of the thin-walled part.

2. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 1, characterized in that, Step S1 further includes: A thin-walled part milling experiment was conducted, in which material was removed layer by layer, and the cutting force and acceleration during the thin-walled part milling process were recorded; during the thin-walled part milling process, a thin-film pressure sensor was used to collect clamping force data. A laser displacement sensor is fixed to the bottom of a thin-walled part, and a line laser beam is aimed at the weak area of ​​the part to collect deformation data during the milling process. The deformation data includes the overall deformation and the location of the maximum deformation.

3. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 1, characterized in that, In step S2, the thin-walled part is divided into m layers. The deflection of the neutral layer of the thin-walled part is set to zero. A mechanism model of the deformation caused by residual stress after material removal is established using the thin plate bending theory. Combined with the cantilever beam theory, a residual stress field deformation mechanism model is formed. ； In the formula, This represents the change in deflection caused by the change in residual stress field after the removal of the i-th layer of material in the discrete element of the part. This represents the deflection change caused by the tool applying a concentrated load F to the discrete elements of the part during the machining of thin-walled parts, according to the cantilever beam theory. It represents the deformation of a discrete element of a part.

4. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 3, characterized in that, The deflection change at the center point caused by the change in residual stress field after the removal of the i-th layer of material in the discrete element of the part is calculated using the following formula. : ； In the formula, t is the thickness of each layer of the thin-walled component, i is the current number of layers to be removed, and h i and h i+1 Let μ be the remaining thickness of the thin-walled component before removing the i-th and (i+1)-th layers, respectively, and let E be the elastic modulus. x,i-1,i To remove the residual stress in the X direction of the i-th thin-walled component after the (i-1)-th layer, σ y,i-1,i To remove the residual stress in the Y direction of the i-th thin-walled component after the (i-1)-th layer, σ y,i-1,1 To remove the residual stress in the Y direction of the first thin-walled component after the (i-1)th layer, σ x,i-1,1 To remove the residual stress in the X direction of the first thin-walled component after the (i-1)th layer; The deflection change value at the center point The deflection change value caused by the change in residual stress field after the removal of the i-th layer of material in the discrete element of the part. .

5. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 3, characterized in that, The deflection change w of a part after the tool applies a concentrated load F to the part during the machining of thin-walled parts. B The calculation process includes the following steps: The discrete unit of the part is considered as a cantilever beam, and the contact process between the tool and the cantilever beam is considered as the application of a concentrated load F at the end of the cantilever beam, causing the cantilever beam to undergo bending deformation as a whole, with a bending moment M on any cross-section: ； In the formula, This indicates the concentrated load applied by the tool to the workpiece. The distance between the moment section and the free end is represented by l, where l is the length of the cantilever beam; the x-coordinate of the moment section B is used as the coordinate. Substituting the values ​​yields the deflection change. for: ； In the formula, I is the moment of inertia of the cross section. It represents the elastic modulus.

6. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 1, characterized in that, In step S3, the multilayer perceptron model includes an input layer, a hidden layer, a fully connected layer, and an output layer connected in sequence; In this model, the input layer of the multilayer perceptron is a one-dimensional vector data, the length of which is equal to the number of discrete units n of the part; the number of neurons in the hidden layer is set to the number of discrete units n of the part, and the initial weight vector of the hidden layer is a distance vector composed of the distance between each discrete unit of the part and the position of maximum deformation; the fully connected layer uses the LeakyReLU function as the neuron activation function.

7. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 1, characterized in that, In step S4, a residual module is introduced into the U-net model to construct a deep learning network model. Both the input and output of this model are one-dimensional vectors. The input of the deep learning network model is the deformation of all discrete units of the thin-walled part, and the output is the hidden layer weight vector w of the trained multilayer perceptron model. n ; The deep learning network model includes an encoding module and a decoding module. The encoding module includes a first residual module, a first pooling layer, a second residual module, a second pooling layer, a third residual module, and a third pooling layer, and extracts features from the deformation of the input discrete units of the part. The decoding module includes a convolutional layer, a fourth residual module, a first transposed pooling layer, a fifth residual module, a second transposed pooling layer, a sixth residual module, a third transposed pooling layer, a seventh residual module, a fourth transposed pooling layer, an eighth residual module, and a ninth residual module, and decodes the features extracted by the encoding module, outputting the predicted hidden layer weight vector w. n ; The residual module consists of a 1×1 convolutional kernel, a 3×1 convolutional kernel, and a LeakyReLu activation function connected in sequence.

8. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 1, characterized in that, In step S5, the mean squared error (MSE) is used as the loss function during the training of the deep learning network model. ； Among them, B s This refers to the batch size, v=1,2,…, B s , These are the hidden layer weight vectors of the multilayer perceptron model in the second training set. These are the hidden layer weight vectors predicted using the second training set data; evaluated using the following function. Evaluating deep learning network models: ； Where N is the number of training periods, These are the hidden layer weight vectors of the multilayer perceptron model in the second test set. It is the hidden layer weight vector of the multilayer perceptron model predicted from the data in the second test set.

9. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 1, characterized in that, The method further includes: The deformation of the discrete elements of the part is input into a deep learning network model, and the hidden layer weight vector w of the multilayer perceptron model is output. n ; Set the hidden layer weight vector w n This serves as a quantitative result of the degree to which local stress affects overall deformation.

10. The mechanism-data hybrid driven online prediction method for machining deformation according to claim 9, characterized in that, The extracted hidden layer weight vector w n Corresponding to the center point coordinates of the discrete unit of the part, the hidden layer weight vector w is based on the HSV color value. n By assigning colors and combining them with corresponding coordinates, a 3D cloud map is drawn, which visualizes the quantitative results of the influence of local stress on the overall deformation.

Images