A method for predicting the thickness of a modified layer in milling of a thin-walled part and a method for optimizing milling parameters

Abstract

The application discloses a thin-walled part milling metamorphic layer thickness prediction method and a milling machining parameter optimization method. The prediction method obtains the milling process parameters of the measured workpiece, and extracts surface ripple physical features from the machining surface; the milling process parameters and the surface ripple physical features are input into a metamorphic layer thickness prediction model, the metamorphic layer thickness prediction model comprises a CNN-SE feature extraction module, a bidirectional long short-term memory network layer, a sequence self-attention layer and an output layer connected in sequence, is used for extracting local correlation features, bidirectional correlation features of a cutting track and key feature position weights, and outputs a metamorphic layer thickness prediction value. The machining parameter optimization method takes candidate radial cutting depths and candidate feed speeds as to-be-optimized variables, takes the metamorphic layer thickness prediction value as fitness, and iteratively outputs a low-damage milling process parameter combination. The application can realize thin-walled part milling metamorphic layer thickness prediction by using surface topography features, and improve low-damage milling parameter screening efficiency.

Description

Technical Field

[0001] This invention belongs to the field of metal thin-wall processing technology, specifically relating to a method for predicting the thickness of the modified layer during milling of thin-walled parts and a method for optimizing milling parameters. Background Technology

[0002] Thin-walled components are widely used in critical structures such as aero-engine blades, integral bladed disks, and aircraft skins. These components typically have thin walls, low stiffness, and are easily deformable. During milling, they are susceptible to the combined effects of cutting forces, cutting heat, and machining vibrations, leading to defects such as chatter marks, burrs, microcracks, or localized plastic flow on the machined surface. For thin-walled components of nickel-based superalloys such as GH4169, due to the material's high strength, low thermal conductivity, and significant work hardening tendency, the cutting heat generated during milling is not easily dissipated in time, resulting in large fluctuations in cutting loads and making it easier to induce surface quality deterioration and subsurface microstructure alteration.

[0003] The thickness of the subsurface altered layer is a crucial indicator for evaluating the surface integrity of milled thin-walled parts. The altered layer is typically caused by the thermo-mechanical coupling effect during machining, manifesting as grain stretching, grain refinement, increased dislocation density, changes in residual stress, and localized structural deformation. Although this type of altered layer lies below the machined surface and is difficult to identify directly through ordinary visual inspection, it affects the fatigue performance, service life, and reliability of the component. Therefore, accurately evaluating the thickness of the subsurface altered layer and selecting appropriate milling parameters accordingly is essential for improving the machining quality of thin-walled parts.

[0004] In existing technologies, the thickness of the subsurface modified layer is typically obtained through methods such as cross-sectional sample preparation, corrosion treatment, and scanning electron microscopy. While these methods can provide a relatively intuitive characterization of the microstructure of the processed cross-section, the testing process is destructive, sample preparation and microscopic measurement are cumbersome, and the testing cycle is long, making them unsuitable for rapid evaluation under a large number of process parameter combinations. Furthermore, since the thickness of the modified layer is influenced by multiple factors such as radial depth of cut, feed rate, tool path, machining vibration, and material microstructure, relying solely on a small number of cross-sectional test results for process judgment often fails to comprehensively reflect the damage variation patterns under different processing conditions.

[0005] To reduce testing costs, some existing technologies employ finite element simulation, theoretical surface morphology models, or data-driven models to analyze the milling process. Finite element simulation can reflect cutting stress, strain, and material removal processes; however, if the model is primarily based on idealized material structures or simplified boundary conditions, it cannot fully reflect the influence of real material grain structure, machining noise, and vibration of thin-walled parts on surface morphology and internal damage. Furthermore, high-precision simulation is computationally expensive and unsuitable for calculating a large number of candidate parameters one by one. Theoretical surface morphology models are computationally efficient, but they typically focus more on tool trajectory and geometric residual height, failing to adequately characterize the material's microscopic response, local vibration disturbances, and the formation of the subsurface altered layer during actual milling.

[0006] Furthermore, existing data-driven prediction methods often use machining parameters, surface roughness, or single morphological indicators as model inputs, making it difficult to fully utilize physically meaningful surface texture features such as the length, width, and depth of surface grooves. For milling of weakly rigid thin-walled parts, there is a complex nonlinear relationship between surface grooves and the subsurface altered layer. Simple regression models or ordinary neural network models are prone to insufficient generalization ability under small sample sizes, high noise levels, and various operating conditions. Some process parameter optimization methods mainly focus on surface roughness, material removal rate, or machining efficiency as optimization targets, failing to adequately consider the implicit damage indicator of the subsurface altered layer thickness. This results in optimized parameter combinations that may not simultaneously achieve surface quality and internal damage control.

[0007] Therefore, existing milling evaluation and parameter optimization techniques for thin-walled parts still have problems such as reliance on destructive testing for subsurface altered layer thickness detection, limited coverage of single data sources, insufficient modeling of the correlation between surface texture features and internal altered layer thickness, and low efficiency in screening low-damage machining parameters. Summary of the Invention

[0008] The purpose of this invention is to provide a method for predicting the thickness of the altered layer in the milling of thin-walled parts and a method for optimizing milling parameters.

[0009] In a first aspect, the present invention provides a method for predicting the thickness of the modified layer during milling of thin-walled parts, comprising:

[0010] The milling process parameters of the workpiece under test are obtained, as well as the surface vibration physical features extracted from the three-dimensional morphology of the machined surface; the surface vibration physical features include vibration length, vibration width and vibration depth.

[0011] A model for predicting the thickness of a modified layer is constructed and trained. The model for predicting the thickness of a modified layer includes a CNN-SE feature extraction module, a bidirectional long short-term memory network layer, a sequence self-attention layer, and an output layer connected in sequence.

[0012] The milling process parameters and the surface vibration physical features are input into the CNN-SE feature extraction module. The CNN-SE feature extraction module extracts the local correlation features between the milling process parameters and the surface vibration physical features, and assigns weights to different feature channels to obtain channel-weighted fusion features.

[0013] After the channel weighted fusion features are converted into a data structure adapted for time-series modeling, they are input into the bidirectional long short-term memory network layer, which extracts the bidirectional correlation features in the cutting trajectory sequential direction and reverse direction.

[0014] The bidirectional correlation features are input into the sequence self-attention layer, which assigns attention weights based on the contribution of different feature positions to the thickness of the metamorphic layer, and performs weighted aggregation on the bidirectional correlation features.

[0015] The weighted and converged features are regressed through the output layer to output the predicted value of the metamorphic layer thickness.

[0016] As a preferred option, the training samples used when training the modified layer thickness prediction model are obtained by fusing actual milled surface topography data, finite element simulation surface data, and theoretically generated surface topography data.

[0017] Preferably, the finite element simulation surface data is obtained through the following method:

[0018] Based on the EBSD experimental data of the processed material, information on grain size, grain orientation and grain distribution is obtained, and a grain size gradient function is established according to the distribution law of grain size along the depth direction of the material.

[0019] Generate a representative volumetric unit model with a polycrystalline grain structure based on the grain size gradient function;

[0020] Using the representative volume element model as the workpiece model, micro-milling finite element simulation was performed to obtain the three-dimensional morphology data of the machined surface.

[0021] The texture length, texture width, and texture depth are extracted from the three-dimensional morphology data of the processed surface.

[0022] Preferably, in the representative volume element model, different grain regions are treated as independent geometric regions and assigned solid cross-sectional properties and material properties respectively.

[0023] Preferably, the theoretically generated surface morphology data is obtained through the following method:

[0024] Establish a tooth sweep trajectory that takes into account the tool rotation motion, tool feed motion, and milling vibration;

[0025] The surface to be processed is discretized into a two-dimensional regular mesh, and the surface height of the mesh nodes is updated according to the cutting envelope height corresponding to the sweeping trajectory of the cutting teeth to obtain the theoretically generated surface morphology data.

[0026] Preferably, the process of updating the surface height of the mesh nodes includes:

[0027] When the cutting envelope height is lower than the current surface height of the corresponding mesh node, the surface height of the mesh node is updated to the cutting envelope height;

[0028] When the cutting envelope height is not lower than the current surface height of the corresponding mesh node, the surface height of the mesh node remains unchanged.

[0029] As a preferred embodiment, the process of extracting the physical characteristics of surface vibration patterns from the three-dimensional morphology of the processed surface includes:

[0030] Multiple reference points are selected within the point cloud projection plane of the processed surface, and local regions are divided based on the multiple reference points;

[0031] Determine the peaks and troughs of the ripples in each local area;

[0032] The ripple length, ripple width, and ripple depth are determined based on the spatial distribution between the ripple peaks and ripple troughs.

[0033] Preferably, the physical characteristics of the surface vibration patterns can also be obtained through a theoretical model of the milled surface morphology.

[0034] Preferably, the CNN-SE feature extraction module includes a convolutional processing unit, a main feature extraction branch, and a channel attention branch;

[0035] The convolution processing unit is used to perform convolution processing on the input milling process parameters and surface texture physical features to obtain intermediate features;

[0036] The main feature extraction branch is used to perform convolution processing and nonlinear transformation on the intermediate features to obtain the main feature map;

[0037] The channel attention branch is used to perform global average pooling on the intermediate features to obtain a statistical description of each feature channel, and generate the channel weights corresponding to each feature channel based on the statistical description.

[0038] The CNN-SE feature extraction module is used to perform channel-by-channel weighted fusion of the channel weights and the main feature map to enhance the feature channels related to the thickness of the metamorphic layer and suppress redundant feature channels, thereby obtaining the channel-weighted fused features.

[0039] Secondly, the present invention provides a method for optimizing low-damage milling parameters for thin-walled parts, comprising:

[0040] The candidate radial depth of cut and / or candidate feed rate are used as variables to be optimized.

[0041] Using candidate radial depth of cut and candidate feed rate as inputs, the modified layer thickness prediction method for milling thin-walled parts according to any one of claims 1 to 8 obtains the predicted value of modified layer thickness as fitness, iteratively optimizes the candidate radial depth of cut and candidate feed rate, and outputs the combination of milling process parameters that minimizes the predicted value of modified layer thickness.

[0042] Preferably, the iterative optimization process includes:

[0043] For each candidate individual, based on the candidate radial depth of cut and candidate feed rate, a corresponding three-dimensional morphology of the machined surface is generated using a milling surface morphology theoretical model. This model generates a tooth sweep trajectory based on the tool rotation, tool feed motion, and milling vibrations in the feed and radial depth of cut directions. The surface to be machined is discretized into a two-dimensional regular mesh, and the surface height of the mesh nodes is updated according to the cutting envelope height corresponding to the tooth sweep trajectory. The texture length, texture width, and texture depth corresponding to the candidate individual are extracted from the three-dimensional morphology of the machined surface.

[0044] Input the candidate radial cutting depth, candidate feed rate, ripple length, ripple width, and ripple depth corresponding to the candidate individual into the metamorphic layer thickness prediction model to obtain the predicted value of the metamorphic layer thickness corresponding to the candidate individual.

[0045] The predicted thickness of the metamorphic layer is used as the basis for fitness evaluation, and the candidate radial cutting depth and candidate feed rate are updated by a genetic algorithm.

[0046] The present invention has the following beneficial effects.

[0047] 1. This invention combines multi-source sample construction, mesh height mapping theory morphology generation, metamorphic layer thickness prediction model and genetic algorithm parameter optimization to form a complete closed loop from surface feature acquisition, internal damage prediction to low-damage process optimization, providing a reliable process decision basis for milling of weakly rigid thin-walled parts.

[0048] 2. By integrating actual milling experimental data, finite element simulation surface data, and theoretically generated surface morphology data, this invention enables the sample to simultaneously contain real machining noise, material microscopic response, and extended working condition information. This can alleviate the problem of insufficient coverage from a single data source and improve the model's generalization ability and stability under different machining conditions.

[0049] 3. This invention extracts the ripple length, ripple width, and ripple depth from the three-dimensional morphology of the processed surface, transforming the macroscopically measurable surface texture into structural features with clear physical meaning. It can predict the thickness of the subsurface metamorphic layer based on non-destructive surface measurement results, reducing the reliance on cross-sectional destructive testing.

[0050] 4. This invention constructs a composite deep regression model by using convolutional neural networks, channel attention, bidirectional long short-term memory networks, and sequential self-attention layers. This model can simultaneously extract local spatial features, bidirectional correlation of cutting trajectories, and weights of key damage-sensitive areas, thereby improving the prediction accuracy of the nonlinear mapping between surface texture features and the thickness of the altered layer.

[0051] 5. This invention uses the trained modified layer thickness prediction model as the fitness evaluator of the genetic algorithm, and optimizes the radial depth of cut and feed rate with the goal of minimizing the subsurface modified layer thickness. This can reduce a large number of solid cutting experiments and high-cost simulations, improve the efficiency of low-damage parameter screening, and improve the surface integrity of thin-walled nickel-based superalloy components. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method for predicting the thickness of the modified layer in milling of thin-walled parts provided in Embodiment 1 of the present invention.

[0053] Figure 2 This is a schematic diagram of the milled surface texture distribution in Embodiment 1 of the present invention.

[0054] Figure 3 This is a schematic diagram of the simulation process of the theoretical surface morphology of milling in Embodiment 1 of the present invention.

[0055] Figure 4 This is a schematic diagram of the network structure of the metamorphic layer thickness prediction model in Embodiment 1 of the present invention.

[0056] Figure 5 The image shows the prediction results of the thin-walled part milling modified layer thickness prediction method provided in Embodiment 1 of the present invention for the training set.

[0057] Figure 6 The image shows the prediction results of the thin-walled part milling modified layer thickness prediction method provided in Embodiment 1 of the present invention for the test set.

[0058] Figure 7 These are measured cross-sectional morphology images of the workpiece at three different positions obtained by processing with optimized processing parameters in Embodiment 2 of the present invention. Detailed Implementation

[0059] The present invention will be further described below with reference to the accompanying drawings.

[0060] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0061] Example 1

[0062] A method for predicting the thickness of the modified layer in milled thin-walled parts is provided for evaluating the surface integrity and predicting the thickness of the subsurface modified layer in thin-walled components made of difficult-to-machine high-temperature alloys such as GH4169. Figure 1 As shown, the method includes the following steps:

[0063] Step S1: Construct a milling sample dataset with multi-source feature fusion

[0064] (1) Obtain the measured data.

[0065] Side milling experiments were conducted on thin-walled parts to obtain machined samples under different combinations of milling process parameters. The milling process parameters included at least the radial depth of cut. and feed rate In this embodiment, the feed rate exist to The radial cutting depth can be set to multiple different values ​​within the range. exist to The range is set to multiple different values, with the axial cutting depth fixed at [value]. The spindle speed is fixed at After processing, three-dimensional topographic point cloud data of the processed surface is acquired using a surface contour acquisition device, and surface vibration physical features are extracted from the three-dimensional topographic point cloud data. Simultaneously, the processed cross-section is microscopically characterized using a scanning electron microscope to measure the thickness of the subsurface metamorphic layer corresponding to the surface vibration physical features. and the thickness of the subsurface metamorphic layer As a label for supervised learning.

[0066] The physical characteristics of the surface ripples include ripple length. ripple width and ripple depth Specifically, such as Figure 2 As shown, multiple reference points are selected within the point cloud projection plane to divide the machined surface into several local regions. Within each local region, the peaks and troughs of the ripple are determined through local extremum search. The ripple length is determined by the spatial distribution between the peaks and troughs. ripple width and ripple depth This allows macroscopically measurable surface textures to be transformed into structured features that can be input into neural networks. This processing method transforms the processed surface morphology from simply being used as an image or point cloud for direct modeling into ripple length, ripple width, and ripple depth with clear physical meaning, thereby improving the ability of subsequent prediction models to identify damage formation factors.

[0067] To expand the sample dataset, this embodiment also constructs a milling simulation model and a milling surface morphology theoretical model, as described in sections (2) and (3).

[0068] (2) The milling simulation model is used to obtain the simulated morphology data of the machined surface through polycrystalline micro-cutting finite element machining simulation. Specifically, the grain size, grain orientation, and grain distribution information are first obtained based on the EBSD experimental data of GH4169 material, and a grain size gradient function is established according to the distribution law of the grain size along the material depth direction; then, ABAQUS is further developed using Python scripts, and a representative volume element model is automatically generated according to the grain size gradient function, material geometry, and target grain number, so that the representative volume element model has a polycrystalline grain structure with continuous transition along the depth direction. In the representative volume element model, different grain regions are treated as independent geometric regions, and each grain region is assigned solid cross-sectional properties and material properties, thereby forming a polycrystalline finite element model for characterizing the real crystal structure of the material.

[0069] When establishing the milling simulation model, the polycrystalline finite element model is used as the workpiece model, and the tool-workpiece cutting relationship is established according to preset milling parameters. The tool model can be set as a rigid body model to reduce computational load and improve simulation convergence; the workpiece material model can adopt a constitutive model and damage criteria capable of describing high strain rate cutting deformation and damage evolution. Subsequently, the contact relationship between the tool and workpiece, tool rotation motion, tool feed motion, workpiece constraints, and mesh generation conditions are set, and a micro-milling finite element simulation of a weakly rigid component is run to obtain stress distribution, material removal status, and three-dimensional surface morphology data after machining.

[0070] The three-dimensional morphology data of the processed surface includes the height distribution, periodic residual height, groove spacing, local plastic accumulation, and surface micro-defect information of the simulated processed surface. Further, surface texture features are extracted from the three-dimensional morphology data of the processed surface to obtain the groove length, groove width, and groove depth of the simulated surface, which are then used as finite element simulation surface feature data. By introducing the finite element simulation surface feature data, surface texture samples reflecting the heterogeneity of grain structure, cutting stress concentration, and the influence of material plastic removal can be added to the sample dataset, thereby improving the prediction stability and generalization ability of the subsequent composite depth regression model for the thickness of the modified layer in milled weakly rigid thin-walled parts.

[0071] (3) such as Figure 3 As shown, the theoretical model for milling surface morphology considers the influence of the cutter tooth motion trajectory and milling vibration on the surface morphology, and uses Z-map to represent the machined three-dimensional surface. Taking a four-flute end mill as an example, let... The diameter of the milling cutter. For blade length, Let be the helix angle, then the total development angle of the knife tooth helix is... satisfy:

[0072]

[0073] set up Number the cutting edge of the milling cutter. ; For the first The blade is in The absolute rotation angle at any given moment; The angle corresponding to the position of the cutting tooth on the helical path; For the first The initial phase offset of each blade; Let be the rotational speed of the milling cutter, then:

[0074]

[0075]

[0076] Considering the rotational and feed motions of the tool, at any point on any cutting edge... Spatial coordinates of time It can be represented as:

[0077]

[0078]

[0079]

[0080] in, The radius value corresponding to any point on the cutting edge. This represents the tool feed rate. Furthermore, to simulate the minute vibrations during actual milling, simple harmonic vibration components are added in the feed direction and the radial depth of cut direction:

[0081]

[0082]

[0083] The coordinates of the cutting edge after adding vibration are:

[0084]

[0085]

[0086]

[0087] in, and These represent the vibration amplitudes in the corresponding directions. and These represent the vibration frequencies in the corresponding directions. and These represent the initial phases in the corresponding directions. Using the above trajectory equations, the sweeping trajectories of the cutting teeth under different process parameters and vibration states can be generated.

[0088] In the Z-map surface topography generation process, the workpiece surface to be processed is discretized into a two-dimensional regular mesh, and a surface height value is set for each mesh node. Let the first node be... Each blade tooth The cutting envelope point at time t is mapped to the first time. line, number The column mesh nodes, and the coordinates of the cutting envelope point in the height direction are: Let the first Before the update, the surface height of this mesh node was: , No. The surface height after the next update is: .when Less than When the cutting envelope of the cutting edge is lower than the current surface height, material removal occurs at that mesh node, and the surface height of that mesh node is updated. ;when Greater than or equal to When the cutting envelope of the cutting edge is not lower than the current surface height, no material removal occurs at this mesh node, and the original surface height remains unchanged. The above height update relationship can be expressed as:

[0089]

[0090] in, Indicates the first Before the next update line, number Surface height of column grid nodes, Indicates the first The surface height after the next update Indicates the first Each blade tooth The time corresponds to the cutting envelope height of the mesh node. By sequentially traversing the cutting envelope points of each tooth at different times and updating the surface height of the corresponding mesh node in the manner described above, the theoretical surface morphology after milling can be obtained.

[0091] In this embodiment, the processing method of the milling surface topography theoretical model does not require storing complex mesh adjacency relationships, and can quickly characterize the material removal process between the tool and the workpiece, which is beneficial for generating a large number of surface texture samples and expanding the training data coverage.

[0092] In this embodiment, the sample dataset is a fusion of actual milled surface topography data, finite element simulation surface data based on crystal structure, and theoretically generated surface topography data. The actual milled surface topography data retains the real machining noise and actual vibration effects; the finite element simulation surface data introduces the heterogeneity effect of the material's crystal structure; and the theoretically generated surface topography data covers more boundary conditions. These three types of data together form sample pairs containing process parameters, surface vibration characteristics, and the thickness of the subsurface altered layer. This multi-source data fusion method can alleviate the problems of limited sample size in a single experimental model, lack of realistic noise in a single theoretical model, and limited coverage of a single finite element simulation, thereby improving the generalization ability and robustness of the prediction model.

[0093] Step S2: Preprocess the sample data and convert it into a tensor format suitable for convolution processing.

[0094] The dataset obtained in step S1 is constructed into a two-dimensional feature matrix. The two-dimensional feature matrix It includes multiple samples, each containing process parameters and surface texture physical characteristics. In one embodiment, a two-dimensional feature matrix... for 3D data matrix, where For the sample size, The number of features for each sample. Arrange the dataset according to... The dataset is divided into training and validation sets, and then each feature is normalized.

[0095]

[0096] in, For the first in the dataset The first sample The original values ​​of each feature, For the first The minimum value of a feature across all samples. For the first The maximum value of a feature across all samples. These are the normalized feature values. After normalization, the two-dimensional texture feature data is converted into a three-dimensional tensor format, enabling it to be input into a convolutional neural network for local spatial feature extraction. This preprocessing method reduces the impact of numerical differences between features of different dimensions on network training, giving the model a more stable convergence direction in the early stages of training.

[0097] Step S3: Construct a model for predicting the thickness of the metamorphic layer

[0098] like Figure 4 As shown, the variation layer thickness prediction model in this embodiment is a CNN-biLSTM-Attention composite deep regression model, which includes an input layer, a CNN-SE feature extraction module, a sequence transformation module, a bidirectional long short-term memory network layer, a sequence self-attention layer, a fully connected layer, and an output layer.

[0099] The input layer is used to receive the current process parameters and the ripple length of each region on the machined surface. ripple width ripple depth The resulting multidimensional physical feature vector.

[0100] The CNN-SE feature extraction module includes a convolutional processing unit, a main feature extraction branch, and a channel attention branch. The convolutional processing unit performs convolutional processing on the input features within their local receptive field, extracting multi-dimensional local structural features between different surface texture features and processing boundaries. In some embodiments, the convolutional processing unit employs two sets of... The convolutional kernel performs feature extraction twice, and intermediate features are obtained through the ReLU activation function.

[0101] The intermediate features output by the convolutional processing unit are simultaneously input into the main feature extraction branch and the channel attention branch. The main feature extraction branch is used to further extract spatial features and perform nonlinear transformations on the intermediate features output by the convolutional processing unit. Specifically, the main feature extraction branch performs continuous convolution processing on the intermediate features output by the convolutional feature extraction module to obtain the main features used for subsequent channel-weighted fusion.

[0102] Let the intermediate features output by the convolutional feature extraction module be the features. The convolution kernel weights in the main feature extraction branch are: The bias term is The main feature extracted by the main feature extraction branch will then be the main feature. It can be represented as:

[0103]

[0104] in, This indicates that the main feature output by the main feature extraction branch is in the th position. line, number Column, No. Feature values ​​at each channel; Representation of features Within the convolution window, the first line, number Column, No. Feature values ​​at each channel; Indicating the main feature extraction branch The convolution kernel in the th line, number Columns, Input Channels With output channel Weight parameters between them; Indicates the first The bias term corresponding to each output channel; This represents a non-linear activation function.

[0105] As can be seen from the above expression, the main feature extraction branch utilizes... The local convolutional window weights and aggregates information from adjacent spatial locations and 64 input channels, and then forms the main feature map after ReLU activation. Through this calculation process, the main feature extraction branch can further extract deep local spatial correlation information between process parameters, ripple length, ripple width, and ripple depth, providing backbone features for subsequent fusion with the channel weights output by the channel attention branch.

[0106] The channel attention branch is used to perform channel dependency modeling on the intermediate features output by the convolutional feature extraction module and to calculate the importance weight of each feature channel.

[0107] Specifically, the channel attention branch first obtains the statistical description of each channel through global average pooling:

[0108]

[0109] in, and These are the feature map height and width, respectively. For the first Each channel is located in The eigenvalue at that location.

[0110] Subsequently, the channel attention branch obtains channel weights through two fully connected layers and a sigmoid activation function, and then compares these channel weights with the main feature map output by the main feature extraction branch. Channel-by-channel weighted fusion is performed. The channel attention branch can highlight key features such as ripple depth and ripple width that are sensitive to the thickness of the subsurface metamorphic layer, while suppressing redundant or noise features, thereby reducing local overfitting and error fluctuations.

[0111] After convolutional feature extraction and channel-weighted fusion, the fused features are unfolded into a data structure suitable for temporal modeling through a sequence defolding layer and a flattening layer, and then input into a bidirectional long short-term memory (LSTM) network layer. The bidirectional LSTM network layer includes forward LSTM units and backward LSTM units, used to simultaneously capture long-distance structural associations in both the sequential and reverse directions of the cutting trajectory. Because the tool periodically enters and exits the workpiece during milling, the texture morphology of a certain area on the surface is not only affected by the cutting state of the previous tooth, but may also be related to the entry state and vibration accumulation state of subsequent teeth. Therefore, using a bidirectional LSTM network can more completely capture the spatiotemporal dynamic evolution of milled surface texture during continuous machining.

[0112] The sequence self-attention layer is set after the bidirectional long short-term memory network layer and is used to dynamically allocate attention weights based on the contribution of different feature locations to the thickness of the subsurface metamorphic layer.

[0113] Figure 4 In the diagram, the multiple "C" modules at the bottom represent the CNN-SE feature extraction module, the yellow and green nodes in the middle represent the forward and backward LSTM hidden states, respectively, and the top node aggregates the multiple hidden states with attention weights before outputting the prediction result. This structure allows the model to focus on key damage source features such as deep grooves, abrupt peaks and valleys, and wide and deep anomalous regions, rather than averaging all texture features, thereby improving its ability to express nonlinear damage mapping relationships.

[0114] The output layer performs a nonlinear transformation on the high-order features output by the sequence self-attention layer through a fully connected layer, and outputs the predicted thickness of the subsurface metamorphic layer. Through the above structure, CNN is used to extract local spatial structural features of surface texture, channel attention branches are used to enhance key physical features, biLSTM is used to capture bidirectional long-range dependencies in the cutting trajectory direction, and sequential self-attention layers are used to locate key regions that contribute more to the thickness of the metamorphic layer. The synergistic effect of these modules enables a high-precision nonlinear mapping from surface texture features to the thickness of the subsurface metamorphic layer.

[0115] Step S4: Train and evaluate the metamorphic layer thickness prediction model.

[0116] This embodiment uses the Adaptive Moment Estimation Adam optimizer to train the model parameters, and uses the Mean Squared Error (MSE) as the loss function. The loss function is:

[0117]

[0118] in, The total number of samples, For the first The actual subsurface metamorphic layer thickness of each sample The model predicts the thickness of the subsurface metamorphic layer. MSE (Mean Sequence Scheme) penalizes larger prediction errors more severely, causing the model to focus more on fitting outlier samples during training, which helps improve overall prediction accuracy. In one embodiment, the initial learning rate is set to... Furthermore, a segmented descent strategy is adopted, reducing the learning rate every 100 training cycles to improve training stability and avoid oscillations in the later stages of training.

[0119] After the model training is completed, the coefficient of determination is used. Root mean square error Mean absolute error and remaining predicted residuals An evaluation will be conducted. The relevant indicators are:

[0120]

[0121]

[0122]

[0123]

[0124] in, This represents the average thickness of the actual subsurface metamorphic layer. The standard deviation is denoted as . The closer the model is to 1, the stronger its fitting ability. and The smaller the value, the lower the prediction error; A larger value indicates better model prediction stability.

[0125] In a specific training result, the model on the training set It is 3.999. It is 15.992. The value is 4.3492; on the test set. It is 3.9928. It is 15.9423. The value is 4.2264. Compared to CNN-GRU and CNN-LSTM models, the CNN-biLSTM-Attention model exhibits lower prediction error and higher prediction stability. The comparison of prediction results between the training and test sets shows that the curves representing actual values ​​and predicted values ​​exhibit a high degree of consistency in their overall trends. Multiple peak and valley positions can be synchronously tracked by the model, indicating that the composite model can learn the nonlinear correspondence between the ripple characteristics and the thickness of the subsurface modified layer. This result demonstrates that this embodiment, through the combination of convolutional spatial feature extraction, bidirectional temporal modeling, and attention weighting mechanisms, can reduce the error fluctuations of traditional data-driven models in predicting milling damage in high-temperature alloy weak-rigid components.

[0126] Example 2

[0127] A method for optimizing low-damage milling parameters for thin-walled parts is disclosed, used to optimize milling parameters for thin-walled, weakly rigid components made of difficult-to-machine high-temperature alloys such as GH4169. This embodiment focuses on the milling of thin-walled GH4169 precipitation-strengthened nickel-based high-temperature alloy workpieces. The workpiece used is a GH4169 cubic block with dimensions of 30mm × 20mm × 14mm; the cutting tool material used in the experiment is YG3 cemented carbide with a hardness of 66HRC. The tool's rake angle is 15°, clearance angle is 10°, tool width is 13mm, and tool tip radius is 0.04mm.

[0128] The method includes the following steps:

[0129] Step S1: Construct a global optimization model for process parameters based on genetic algorithm

[0130] After the model training is complete and the preset accuracy requirement is achieved, the weight-locked CNN-biLSTM-Attention composite deep regression model is used as the fitness evaluator for the genetic algorithm. The optimization objective of the genetic algorithm is to minimize the predicted subsurface metamorphic layer thickness. The objective function is:

[0131]

[0132] in, To train a good model for predicting the thickness of the metamorphic layer, , and These represent the surface texture length, surface texture width, and surface texture depth, respectively. To avoid the genetic algorithm's search results exceeding the actual processing range, the radial cutting depth is... and feed rate Set the following boundary constraints:

[0133]

[0134]

[0135] Meanwhile, the axial depth of cut and spindle speed are fixed as constant constraints. and The genetic algorithm selects candidate radial cutting depths. and candidate feed rate Encode individuals as chromosomes to construct the initial population.

[0136] Step S2: Iterative optimization to obtain the optimal parameters.

[0137] For each candidate individual in the population, the corresponding surface texture length is first generated by calling the milling surface morphology theoretical model in Example 1. Surface groove width and surface texture depth The generated surface vibration physical features are then input into the modified layer thickness prediction model with locked weights to obtain the corresponding predicted subsurface modified layer thickness. Genetic algorithms are based on predictions. Fitness is calculated and the population is iteratively updated through selection, crossover, and mutation operations.

[0138] In one embodiment, the genetic algorithm has a population size of 80, a crossover ratio of 0.8, and a maximum number of iterations of 100. The algorithm iteratively performs fitness calculation, selection, crossover, and mutation operations until the maximum number of iterations is reached or the convergence criterion is met, ultimately outputting the combination of process parameters that minimizes the predicted thickness of the subsurface metamorphic layer. This optimization method embeds a deep learning prediction model into the fitness calculation stage of the genetic algorithm, enabling the genetic algorithm to quickly evaluate the damage risk corresponding to each candidate parameter without requiring physical cutting experiments or high-cost finite element simulations for each set of candidate parameters, thereby improving global optimization efficiency.

[0139] In this embodiment, the optimal combination of milling parameters output by the genetic algorithm is the radial depth of cut. feed rate Under this parameter combination, the metamorphic layer thickness prediction model predicts the subsurface metamorphic layer thickness as follows: Compared to the optimal sample of damage layer thickness in the original dataset Reduced The optimal parameter combination was used for actual milling, and the machined cross-section was verified using a scanning electron microscope. The results are as follows: Figure 7 As shown. From Figure 7 As can be seen from the data, the thickness of the subsurface metamorphic layer at different locations is as follows: , and The experimental measurement results agree well with the model predictions, with a maximum relative error of [value missing]. The average error is .

[0140] In addition, combined Figure 7 It can be seen that the thickness distribution of the damaged layer on the workpiece section is relatively stable under the optimized parameter conditions, and no obvious deep-seated severe plastic deformation zone appears. From the surface morphology diagram after parameter optimization, it can be seen that the surface texture is generally continuous, clear, and relatively uniformly distributed, without obvious tearing, collapse, or severe plastic flow. This indicates that the optimized machining parameter combination obtained in this embodiment can reduce the thickness of the subsurface altered layer while maintaining stable cutting, thus improving the surface integrity of thin-walled parts after milling.

[0141] Building upon Example 1, Example 2 transforms internal subsurface damage, which is difficult to measure directly online, into a target quantity that can be indirectly predicted by surface morphology, enabling non-destructive assessment of milling damage in thin-walled parts and optimization of low-damage parameters. Through a combination of multi-source data fusion, Z-map theoretical surface morphology generation, modified layer thickness prediction model prediction, and GA global optimization, this example improves the accuracy of modified layer thickness prediction and provides process parameter basis for high-precision, low-damage milling of thin-walled components made of difficult-to-machine high-temperature alloys such as GH4169.

Claims

1. A method for predicting the thickness of the modified layer during milling of thin-walled parts, characterized in that, include: The milling process parameters of the workpiece under test are obtained, as well as the surface vibration physical characteristics extracted from the three-dimensional morphology of the machined surface. The physical characteristics of the surface texture include texture length, texture width, and texture depth; A model for predicting the thickness of a modified layer is constructed and trained. The model for predicting the thickness of a modified layer includes a CNN-SE feature extraction module, a bidirectional long short-term memory network layer, a sequence self-attention layer, and an output layer connected in sequence. The milling process parameters and the surface vibration physical features are input into the CNN-SE feature extraction module. The CNN-SE feature extraction module extracts the local correlation features between the milling process parameters and the surface vibration physical features, and assigns weights to different feature channels to obtain channel-weighted fusion features. After the channel weighted fusion features are converted into a data structure adapted for time-series modeling, they are input into the bidirectional long short-term memory network layer, which extracts the bidirectional correlation features in the cutting trajectory sequential direction and reverse direction. The bidirectional correlation features are input into the sequence self-attention layer, which assigns attention weights based on the contribution of different feature positions to the thickness of the metamorphic layer, and performs weighted aggregation on the bidirectional correlation features. The output layer performs regression processing on the weighted and converged features to output the predicted value of the metamorphic layer thickness.

2. The method for predicting the thickness of the modified layer in milling of thin-walled parts according to claim 1, characterized in that, The training samples used in training the modified layer thickness prediction model are obtained by fusing actual milled surface topography data, finite element simulation surface data, and theoretically generated surface topography data.

3. The method for predicting the thickness of the modified layer in milled thin-walled parts according to claim 2, characterized in that, The finite element simulation surface data was obtained through the following methods: Based on the EBSD experimental data of the processed material, information on grain size, grain orientation and grain distribution is obtained, and a grain size gradient function is established according to the distribution law of grain size along the depth direction of the material. Generate a representative volumetric unit model with a polycrystalline grain structure based on the grain size gradient function; Using the representative volume element model as the workpiece model, micro-milling finite element simulation was performed to obtain the three-dimensional morphology data of the machined surface. The texture length, texture width, and texture depth are extracted from the three-dimensional morphology data of the processed surface.

4. The method for predicting the thickness of the modified layer in milling of thin-walled parts according to claim 3, characterized in that, In the representative volume element model, different grain regions are treated as independent geometric regions and are assigned solid cross-sectional properties and material properties respectively.

5. The method for predicting the thickness of the altered layer in milling of thin-walled parts according to claim 2, characterized in that, The theoretically generated surface morphology data is obtained through the following methods: Establish a tooth sweep trajectory that takes into account the tool rotation motion, tool feed motion, and milling vibration; The surface to be processed is discretized into a two-dimensional regular mesh, and the surface height of the mesh nodes is updated according to the cutting envelope height corresponding to the sweeping trajectory of the cutting teeth to obtain the theoretically generated surface morphology data.

6. The method for predicting the thickness of the modified layer in milling of thin-walled parts according to claim 5, characterized in that, The process of updating the surface height of a mesh node includes: When the cutting envelope height is lower than the current surface height of the corresponding mesh node, the surface height of the mesh node is updated to the cutting envelope height; When the cutting envelope height is not lower than the current surface height of the corresponding mesh node, the surface height of the mesh node remains unchanged.

7. The method for predicting the thickness of the modified layer in milling of thin-walled parts according to claim 1, characterized in that, The process of extracting the physical characteristics of surface vibration marks from the three-dimensional morphology of the machined surface includes: Multiple reference points are selected within the point cloud projection plane of the processed surface, and local regions are divided based on the multiple reference points; Determine the peaks and troughs of the ripples in each local area; The ripple length, ripple width, and ripple depth are determined based on the spatial distribution between the ripple peaks and ripple troughs.

8. The method for predicting the thickness of the modified layer in milling of thin-walled parts according to claim 1, characterized in that, The CNN-SE feature extraction module includes a convolutional processing unit, a main feature extraction branch, and a channel attention branch; The convolution processing unit is used to perform convolution processing on the input milling process parameters and surface texture physical features to obtain intermediate features; The main feature extraction branch is used to perform convolution processing and nonlinear transformation on the intermediate features to obtain the main feature map; The channel attention branch is used to perform global average pooling on the intermediate features to obtain statistical descriptions of each feature channel, and generate channel weights corresponding to each feature channel based on the statistical descriptions. The CNN-SE feature extraction module is used to perform channel-by-channel weighted fusion of the channel weights and the main feature map to enhance the feature channels related to the thickness of the metamorphic layer and suppress redundant feature channels, thereby obtaining the channel-weighted fused features.

9. A method for optimizing low-damage milling parameters for thin-walled parts, characterized in that, include: The candidate radial depth of cut and / or candidate feed rate are used as variables to be optimized. Using candidate radial depth of cut and candidate feed rate as inputs, the modified layer thickness prediction method for milling thin-walled parts according to any one of claims 1 to 8 obtains the predicted value of modified layer thickness as fitness, iteratively optimizes the candidate radial depth of cut and candidate feed rate, and outputs the combination of milling process parameters that minimizes the predicted value of modified layer thickness.

10. The method for optimizing low-damage milling parameters for thin-walled parts according to claim 9, characterized in that, The iterative optimization process includes: For each candidate individual, based on the candidate radial depth of cut and candidate feed rate, a corresponding three-dimensional morphology of the machined surface is generated using a milling surface morphology theoretical model. This model generates a tooth sweep trajectory based on the tool rotation, tool feed motion, and milling vibrations in the feed and radial depth of cut directions. The surface to be machined is discretized into a two-dimensional regular mesh, and the surface height of the mesh nodes is updated according to the cutting envelope height corresponding to the tooth sweep trajectory. The texture length, texture width, and texture depth corresponding to the candidate individual are extracted from the three-dimensional morphology of the machined surface. Input the candidate radial cutting depth, candidate feed rate, ripple length, ripple width, and ripple depth corresponding to the candidate individual into the metamorphic layer thickness prediction model to obtain the predicted value of the metamorphic layer thickness corresponding to the candidate individual. The predicted thickness of the metamorphic layer is used as the basis for fitness evaluation, and the candidate radial cutting depth and candidate feed rate are updated by a genetic algorithm.

Images