Contour processing method and device for multi-axis linkage machine tool

Abstract

The present invention relates to the technical field of multi-axis linkage processing, and discloses a contour processing method and device for a multi-axis linkage machine tool. The method comprises: performing multi-angle image acquisition and three-dimensional convolution feature extraction on a rough workpiece to obtain a three-dimensional rough workpiece feature map; performing layered enhancement processing of contour features to obtain a multi-layer enhanced feature map; performing spatial adaptive transformation and feature correction to obtain a target correction feature map; performing iterative contour reconstruction and refinement processing to obtain a target contour model; performing multi-objective optimization calculation on processing parameters based on the target contour model to obtain a dynamic optimal processing parameter set; performing real-time workpiece deformation prediction and adaptive compensation control on the dynamic optimal processing parameter set and multi-source processing data collected in real time to obtain a target compensation instruction sequence. The present invention is used for accurate contour recognition, adaptive parameter optimization, and real-time deformation prediction and compensation control, thereby significantly improving processing accuracy and efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-axis linkage machining, and in particular to a contour machining method and device for a multi-axis linkage machine tool. Background Art

[0002] Multi-axis machine tools play a key role in the field of high-precision machining. However, traditional contour machining methods can no longer meet the growing demand for high precision and high efficiency. Existing technologies often have problems such as inaccurate contour recognition, insufficient optimization of machining parameters, and inadequate real-time compensation when processing complex surfaces, making it difficult to further improve machining accuracy and efficiency.

[0003] In addition, in the actual processing, the deformation of the workpiece, the wear of the tool and the influence of environmental factors will have a significant impact on the processing quality. Traditional methods often find it difficult to capture these dynamic changes in real time, resulting in the accumulation of errors during the processing, which ultimately affects the quality of the finished product. These problems are particularly prominent when processing large and complex curved workpieces. Summary of the invention

[0004] The present invention provides a contour processing method and device for a multi-axis linkage machine tool, which are used for accurate contour recognition, adaptive parameter optimization, and real-time deformation prediction and compensation control, thereby significantly improving processing accuracy and efficiency.

[0005] In a first aspect, the present invention provides a contour processing method for a multi-axis linkage machine tool, the contour processing method for a multi-axis linkage machine tool comprising:

[0006] Perform multi-angle image acquisition and three-dimensional convolution feature extraction on the rough workpiece to obtain a three-dimensional rough workpiece feature map;

[0007] Inputting the three-dimensional rough workpiece feature map into a multi-scale feature enhancement network to perform contour feature hierarchical enhancement processing to obtain a multi-layer enhanced feature map;

[0008] Performing spatial adaptive transformation and feature correction on the multi-layer enhanced feature map to obtain a target corrected feature map;

[0009] Inputting the target correction feature map into the cascaded deep transposed convolutional network and the graph convolutional network for contour iterative reconstruction and refinement processing to obtain a target contour model;

[0010] Based on the target contour model, multi-objective optimization calculation is performed on the processing parameters to obtain a dynamic optimal processing parameter set;

[0011] Real-time workpiece deformation prediction and adaptive compensation control are performed on the dynamic optimal machining parameter set and multi-source machining data collected in real time to obtain a target compensation instruction sequence.

[0012] In a second aspect, the present invention provides a contour processing device for a multi-axis linkage machine tool, the contour processing device for a multi-axis linkage machine tool comprising:

[0013] The acquisition module is used to perform multi-angle image acquisition and three-dimensional convolution feature extraction on the rough workpiece to obtain a three-dimensional rough workpiece feature map;

[0014] An enhancement module is used to input the three-dimensional rough workpiece feature map into a multi-scale feature enhancement network to perform contour feature hierarchical enhancement processing to obtain a multi-layer enhanced feature map;

[0015] A correction module, used for performing spatial adaptive transformation and feature correction on the multi-layer enhanced feature map to obtain a target corrected feature map;

[0016] A reconstruction module is used to input the target correction feature map into a cascaded deep transposed convolutional network and a graph convolutional network for contour iterative reconstruction and refinement processing to obtain a target contour model;

[0017] A calculation module, used for performing multi-objective optimization calculation on the processing parameters based on the target contour model to obtain a dynamic optimal processing parameter set;

[0018] The control module is used to perform real-time workpiece deformation prediction and adaptive compensation control on the dynamic optimal processing parameter set and multi-source processing data collected in real time to obtain a target compensation instruction sequence.

[0019] The third aspect of the present invention provides a contour processing device for a multi-axis linkage machine tool, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the contour processing device for the multi-axis linkage machine tool executes the above-mentioned contour processing method for the multi-axis linkage machine tool.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned contour processing method for a multi-axis linkage machine tool.

[0021] In the technical solution provided by the present invention, high-precision recognition of complex surface contours is achieved through multi-angle image acquisition and three-dimensional convolution feature extraction, combined with a multi-scale feature enhancement network. The spatial adaptive transformation and feature correction technology can dynamically adjust the feature processing strategy according to the characteristics of different workpieces, thereby improving the robustness and adaptability of contour recognition. The cascaded deep transposed convolutional network and the graph convolutional network are used for iterative contour reconstruction and refinement, which can capture tiny contour details and improve the accuracy of the contour model. The multi-objective optimization algorithm based on deep reinforcement learning can simultaneously consider multiple goals such as machining accuracy, efficiency, and tool life, and realize intelligent dynamic optimization of machining parameters. Combined with the self-correcting finite element model and the recursive neural network predictor, real-time prediction of the deformation of the workpiece during the machining process is realized. By constructing a compensation function and adjusting the tool path in real time, dynamic compensation for errors during the machining process is realized, which significantly improves the machining accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 A schematic flow chart of a contour processing method for a multi-axis linkage machine tool provided in an embodiment of the present application;

[0024] Figure 2 A schematic block diagram of the structure of a contour processing device for a multi-axis linkage machine tool provided in an embodiment of the present application;

[0025] Figure 3 A schematic block diagram of the structure of a contour processing device for a multi-axis linkage machine tool provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change based on actual conditions.

[0028] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0029] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0030] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0031] See also Figure 1 , Figure 1 A schematic flow chart of a contour processing method for a multi-axis linkage machine tool provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the contour processing method for a multi-axis linkage machine tool provided in an embodiment of the present application includes steps S100 to S600.

[0032] Step S100, performing multi-angle image acquisition and three-dimensional convolution feature extraction on the rough workpiece to obtain a three-dimensional rough workpiece feature map;

[0033] It is understandable that the execution subject of the present invention may be a contour processing device for a multi-axis linkage machine tool, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0034] Specifically, the rough workpiece is rotated 360 degrees at equal intervals, and high-resolution images are collected at each angle to fully capture the external structure and detail information of the workpiece. In this process, image collection is performed at a small angle interval to achieve sufficient coverage, and a multi-angle original image set containing at least 24 images is obtained. Each image in the multi-angle original image set is preprocessed to improve the image quality and the accuracy of subsequent three-dimensional reconstruction. Each image is denoised by Gaussian filtering to reduce the random noise introduced during the acquisition process. Histogram equalization is performed to enhance the contrast of the image and make the detailed features of the workpiece more prominent. The distortion caused by the camera shooting angle is removed by perspective transformation correction to ensure that the shape of the workpiece in each image is closer to the actual contour. After these processes, a preprocessed multi-angle image set is obtained. The preprocessed multi-angle image set is subjected to three-dimensional reconstruction based on structured light. Structured light is a technology that calculates depth by projecting known light patterns onto the surface of an object and collecting its deformation. The three-dimensional structure of the workpiece is accurately reconstructed by this technology. Using the principle of triangulation, the depth information of each pixel is calculated to generate an initial three-dimensional point cloud model of the workpiece. The point cloud model contains the coordinates of each pixel in three-dimensional space, thus describing the surface shape of the workpiece in detail. Next. The initial three-dimensional point cloud model is voxelized. The continuous point cloud data is discretized and converted into voxel form, and the size of the voxel is set to 0.5mm to ensure the accuracy of the data and the rationality of the discretization, generating a voxelized three-dimensional model of size 128x128x128. The voxelized three-dimensional model is input into the three-dimensional convolutional neural network for feature extraction. The three-dimensional convolutional neural network contains 5 three-dimensional convolutional layers, and the convolution kernel size of each layer is set to 3x3x3, the step size is 1, the padding mode is 1, and the activation function selects the ReLU function. The ReLU activation function can effectively introduce nonlinear characteristics and improve the model's ability to express complex features. In the convolutional network, the maximum pooling layer is added after the 2nd and 4th layers, and the size of the pooling kernel is 2x2x2. The size of the feature map is gradually reduced through the pooling operation, while retaining important features, reducing the amount of calculation, and preventing the model from overfitting, and obtaining a preliminary three-dimensional feature map. The preliminary three-dimensional feature map is input into 3 residual blocks. Each residual block consists of two 3D convolutional layers and a skip connection. Skip connection is a connection method that directly transfers input features to the output. It fuses low-level and high-level features by element-by-element addition, effectively alleviating the gradient vanishing problem in deep convolutional networks and improving the network's ability to express multi-scale features. After being processed by the residual block, a multi-scale fused 3D feature map is obtained. A global average pooling operation is performed on the multi-scale fused 3D feature map to compress the feature map to a size of 1x1x1, which greatly reduces the dimension of the feature map while still retaining the global feature information.The feature dimension is adjusted through the fully connected layer to obtain a feature vector with a dimension of 256, and the feature vector is L2 normalized to ensure that the features are on the unit sphere, so that the distances between features have good comparability, and finally the feature map of the three-dimensional rough workpiece is obtained.

[0035] Step S200, inputting the three-dimensional rough workpiece feature map into a multi-scale feature enhancement network to perform contour feature hierarchical enhancement processing to obtain a multi-layer enhanced feature map;

[0036] Specifically, the feature map of the three-dimensional rough workpiece is decomposed at multiple scales. Through Gaussian pyramid downsampling, the original feature map is subjected to multiple Gaussian blur and downsampling operations to gradually obtain three feature maps with different resolutions, namely the original resolution, 1 / 2 resolution and 1 / 4 resolution. Edge detection is performed on the feature map of each resolution to highlight the geometric features such as the contour and corner of the workpiece. The edge feature map at the corresponding scale is extracted from the feature maps of different resolutions to obtain three edge feature maps of different scales. The edge feature maps of the three different scales are fused with the original feature maps of their corresponding resolutions respectively, and the edge features are combined with the original feature maps by element-by-element addition to obtain three enhanced feature maps. The three enhanced feature maps are subjected to nonlinear transformation to improve the distinguishability and expression ability of the features. The nonlinear transformation enhances the nonlinear characteristics of the feature map by introducing an activation function to obtain three nonlinear enhanced feature maps. The three nonlinear enhanced feature maps are input into the feature pyramid network for multi-scale feature fusion. The function of the feature pyramid network is to integrate feature maps from different scales so that the final feature map can simultaneously contain effective information of each scale. Through the multi-scale fusion operation of the feature pyramid, the preliminary fused feature map can simultaneously have the details of high-resolution features and the global contour information of low-resolution features. The preliminary fused feature map is enhanced by residual features. By introducing a residual connection structure, the input features are added to the output features after several convolution operations to achieve the purpose of retaining the original feature information and enhancing the feature expression ability at the same time. Through residual enhancement, the gradient vanishing problem in the deep network is alleviated, and the richness of features is increased to obtain a residual enhanced feature map. The feature channels of the residual enhanced feature map are rearranged. By reorganizing the channel order of the feature map, the complementary information between different channels is better combined. Feature channel rearrangement reorders the features according to their importance in an adaptive manner, so that the most critical features can play a better role, and finally a multi-layer enhanced feature map is obtained.

[0037] Step S300, performing spatial adaptive transformation and feature correction on the multi-layer enhanced feature map to obtain a target corrected feature map;

[0038] Specifically, the multi-layer enhanced feature map is subjected to feature statistical analysis, the feature mean and variance of each spatial position are calculated, and a feature statistical map is generated to describe the distribution of features at different spatial positions and reflect the feature differences in different regions. Based on these statistical information, the feature space is divided into high-frequency regions and low-frequency regions to obtain a region segmentation map. The high-frequency region corresponds to the part with richer details, such as the edge and corner of the workpiece, while the low-frequency region corresponds to the part with smoother changes. The region segmentation map is subjected to morphological processing to obtain an optimized region segmentation map. Morphological processing further optimizes the segmented region through operations such as expansion, corrosion, opening and closing operations to eliminate the existing small noise regions and isolated points, making the region segmentation more reasonable and continuous. The optimized region segmentation map is weighted fused with the multi-layer enhanced feature map, in which the high-frequency region is given a higher weight and the low-frequency region is given a lower weight. Through weighted fusion, while maintaining important details, the interference of unnecessary smooth parts is reduced to obtain a preliminary adaptive guidance map. The preliminary adaptive guidance map can perform differentiated weighted processing in different regions according to their frequency characteristics, so that the retention and enhancement of important features are more prominent. The preliminary adaptive guidance map is processed by applying an edge-preserving filter to effectively preserve the edge information and smooth the interior of the region. The edge-preserving filter is a filter that can smooth the non-edge region while preserving the edge information of the image, thereby preserving the contour features in the detail area and reducing noise and unnecessary details in the smooth area to obtain a refined adaptive guidance map. Based on the refined adaptive guidance map, the spatial transformation parameter set required by the spatial transformation network is constructed. The main purpose of the spatial transformation network is to spatially adjust the feature map to adapt to different processing requirements. By refining the regional information and edge information contained in the adaptive guidance map, a spatial transformation parameter set that adapts to the current workpiece features is constructed. These parameters include transformation operations such as rotation, translation, and scaling, which are used to flexibly adjust the feature map in space so that the feature map can more accurately represent the structural features of the workpiece. According to the spatial transformation parameter set, the multi-layer enhanced feature map is spatially adaptively transformed to obtain the feature map after spatial transformation. The transformation process adjusts the position and structure of the feature map so that important features can be expressed more prominently, while unimportant parts are appropriately compressed or ignored. The feature map after spatial transformation is adaptively normalized. According to the statistical information of the feature map at different positions, the feature values ​​are redistributed so that the final feature map is kept within a certain range in terms of value, while enhancing the contrast between features. Through adaptive normalization, the target correction feature map is finally obtained.

[0039] Step S400: input the target correction feature map into the cascaded deep transposed convolutional network and the graph convolutional network for contour iterative reconstruction and refinement processing to obtain a target contour model;

[0040] Specifically, the target correction feature map is input into the first layer of the cascaded deep transposed convolutional network, which contains 128 transposed convolution kernels of size 3x3, with a step size of 2 and a padding mode of 1. The ReLU activation function and batch normalization operation are used at the same time. The transposed convolution operation of this layer is regarded as an upsampling process. The spatial resolution of the feature map is enlarged to generate a preliminary upsampled feature map, and nonlinearity is introduced through the ReLU activation function to enhance the expressiveness of the model. Batch normalization helps to stabilize the training process of the network and reduce the risk of gradient disappearance or explosion. The preliminary upsampled feature map is residually connected and added element by element with the target correction feature map to obtain a fused feature map. The introduction of residual connection can retain the original input features and directly add them to the upsampled features, enhance the network's ability to retain the original detail features, and improve the network's performance in complex feature reconstruction tasks, and obtain a more representative fused feature map. The fused feature map is input to the second layer of the cascaded deep transposed convolutional network. The second layer contains 64 3x3 transposed convolution kernels, with a stride of 2 and a padding of 1. The ReLU activation function and batch normalization are used. The operation of this layer is to further upsample the fused feature map so that the spatial resolution of the feature map is further improved to generate a secondary upsampled feature map. Through two transposed convolution operations, the original low-resolution feature map is gradually enlarged, and its detailed features are enhanced by fusion and nonlinear activation, so that the reconstructed feature map can better describe the complex contour information of the workpiece. The secondary upsampled feature map is converted into graph structure data. In this process, each pixel is regarded as a node in the graph, and edge connections are established between adjacent pixels to form an initial graph structure. The initial graph structure is input to the first layer of the graph convolutional network, which contains 32 graph convolution kernels. Each graph convolution kernel performs a convolution operation on the node and its first-order neighbors. The features of each node are updated by the features of itself and its adjacent nodes, resulting in a richer local feature enhancement map. By using the ReLU activation function, nonlinearity is effectively introduced, making the expression of local features more flexible and diverse. The local feature enhancement map effectively captures the relationship between each node and its direct neighbors, thereby enhancing the local contour information of the workpiece. In order to improve the expression ability of global features, the local feature enhancement map is input into the second layer of the graph convolutional network. This layer contains 16 graph convolution kernels, each of which performs convolution operations on the node and its second-order neighbors. The features of each node are not only affected by its direct neighbors, but also by more distant neighbor nodes. In this way, global features in a larger range are captured and a global feature enhancement map is obtained. Global feature enhancement helps to understand the overall structure of the workpiece, ensure the coordination between the various parts during the reconstruction process, and make the reconstructed target contour model perform better in global consistency. The edge refinement processing is performed on the global feature enhancement map. The nodes in the graph are labeled and optimized using conditional random fields.Conditional random fields are a probabilistic graph model that can globally optimize node labels in a graph, smoothing the interior of a region while keeping edges sharp. Through the optimization of conditional random fields, edge blur or jaggedness can be effectively reduced, resulting in a clearer boundary and ultimately generating a target contour model.

[0041] Step S500: Based on the target contour model, multi-objective optimization calculation is performed on the processing parameters to obtain a dynamic optimal processing parameter set;

[0042] Specifically, the target contour model is adaptively divided into multiple layers, and the contour is divided based on the curvature change rate. The contour is divided into multiple sub-regions, and the curvature change rate of each sub-region does not exceed the preset threshold to ensure that the contour of each region is relatively smooth. The complex target contour is decomposed into several simple regions to obtain a block contour model. A local coordinate system is constructed for each sub-region in the block contour model, and the global contour information is converted into a local feature representation to eliminate the complexity caused by the scale difference between different regions, so that the features of each sub-region can be described in a unified local coordinate system to obtain a local feature set. The local feature set is input into a pre-trained deep reinforcement learning agent network for optimization. The deep reinforcement learning agent network contains 4 fully connected layers, each layer uses the ReLU activation function, and the output layer uses the Softmax function to obtain the initial processing parameter action value. The role of the deep reinforcement learning agent network is to generate a set of processing parameter action values ​​based on the input local feature set. These parameters include variables such as feed speed and cutting depth that directly affect the processing effect. By using the ReLU activation function, the network effectively introduces nonlinearity, making the expression of complex features more flexible, while the Softmax function probabilistically processes the output action value to select the optimal processing action. Based on the initial processing parameter action value, a multi-objective reward function is constructed to comprehensively evaluate the current processing parameters. The multi-objective reward function includes three objectives: processing accuracy, processing efficiency, and tool life. By weighted summation of these three objectives, the performance of the current action value is comprehensively evaluated to obtain a comprehensive reward value. In order to improve the performance of deep reinforcement learning, the comprehensive reward value is input into the advantage function estimator of the deep reinforcement learning agent network. The advantage function estimator consists of two fully connected layers and uses the Tanh activation function to calculate the advantage of the action value in the current state. The main function of the advantage function is to measure the superiority of the current action relative to the average level, so as to help the network better focus on those outstanding actions in policy updating. By calculating the advantage function value, more detailed feedback is provided to the network, so as to more effectively improve the quality of the processing parameters in the policy optimization process. Based on the advantage function value, the policy gradient algorithm is used to update the parameters of the deep reinforcement learning agent network. The policy gradient algorithm continuously adjusts the network parameters to increase the probability of obtaining a high reward value in a given state, so that the network can gradually learn better processing parameters. In this process, the network continuously iterates and optimizes, and the processing parameters are gradually improved to obtain an optimized processing parameter set. The feasibility of the optimized processing parameter set is verified, and the processing parameters are checked through kinematic and dynamic constraints. Parameter combinations that do not meet the physical limitations of the machine tool are eliminated, such as those that cause the machine tool to exceed its range of motion or withstand excessive force, to obtain a feasible processing parameter set.In order to ensure the continuity and smoothness of the machining parameters between adjacent sub-regions, the parameters in the feasible machining parameter set are interpolated and smoothed. The main purpose of interpolation smoothing is to avoid transition discontinuity problems caused by sudden changes in parameters during machining, thereby reducing the impact on machine tools and tools and improving the surface quality of machining. After interpolation smoothing, the dynamic optimal machining parameter set is finally obtained.

[0043] Step S600: Perform real-time workpiece deformation prediction and adaptive compensation control on the dynamic optimal processing parameter set and the multi-source processing data collected in real time to obtain a target compensation instruction sequence.

[0044] Specifically, the dynamic optimal machining parameter set is input into the self-correcting finite element model, and the initial geometric model and material property model of the workpiece are constructed to obtain the initial finite element model. The model contains the geometric characteristics of the workpiece and the mechanical properties of the material. The initial finite element model is meshed and the boundary conditions are set. The adaptive mesh refinement algorithm is used to locally encrypt the high stress area to obtain the optimized finite element mesh model. The resolution of the stress concentration area is improved by increasing the mesh density. The multi-source machining data collected in real time, including cutting force, temperature and vibration data, are input into the optimized finite element mesh model for static and dynamic analysis to obtain the stress distribution and deformation field of the workpiece, describing the internal stress state of the workpiece during the machining process and the deformation caused by these stresses. In order to reduce the complexity of the calculation and extract the most critical features for deformation prediction, feature extraction and feature dimensionality reduction are performed on the stress distribution and deformation field to obtain a low-dimensional feature vector. The low-dimensional feature vector is input into a recurrent neural network (RNN) predictor, which contains a long short-term memory (LSTM) layer and two fully connected layers, and uses a tanh activation function. The LSTM layer is suitable for processing time-dependent sequence data, and can effectively predict the deformation of the workpiece at future moments to obtain the predicted deformation field. Based on the predicted deformation field, the back propagation algorithm is used to update the parameters of the self-correcting finite element model. The parameter update includes the adjustment of material properties and boundary conditions, so that the finite element model can more accurately reflect the state changes of the workpiece during the processing process. Through back propagation, the model adjusts the parameters according to the deviation between the prediction and the actual after each iteration, so that its prediction results gradually approach the actual physical response, and the corrected finite element model is obtained. The output of the corrected finite element model is compared with the actual measured deformation data, the error is calculated and the compensation function is constructed. The role of the compensation function is to map the prediction error to the correction amount of the tool path, so as to generate the corresponding compensation strategy. Based on the compensation strategy, the real-time adjustment instructions of the tool path are generated. Combined with the kinematic inverse solution algorithm of the machine tool, the correction amount of the tool path is converted into the motion instructions of each axis. The kinematic inverse solution algorithm of the machine tool is used to decompose the global tool position and posture requirements into displacement instructions of each axis to ensure the accurate implementation of the compensation strategy on the physical machine tool. Through this process, the target compensation instruction sequence is finally obtained. These instructions can guide the machine tool axes to perform real-time compensation actions, thereby ensuring that the workpiece always maintains the expected contour and shape during the machining process.

[0045] In the embodiment of the present invention, high-precision recognition of complex surface contours is achieved through multi-angle image acquisition and three-dimensional convolution feature extraction, combined with a multi-scale feature enhancement network. By using spatial adaptive transformation and feature correction technology, the feature processing strategy can be dynamically adjusted according to the characteristics of different workpieces, thereby improving the robustness and adaptability of contour recognition. The use of cascaded deep transposed convolutional networks and graph convolutional networks for iterative contour reconstruction and refinement can capture tiny contour details and improve the accuracy of the contour model. The multi-objective optimization algorithm based on deep reinforcement learning can simultaneously consider multiple goals such as machining accuracy, efficiency, and tool life, and realize intelligent dynamic optimization of machining parameters. Combined with a self-correcting finite element model and a recursive neural network predictor, real-time prediction of deformation of the workpiece during machining is achieved. By constructing a compensation function and adjusting the tool path in real time, dynamic compensation for errors during machining is achieved, significantly improving machining accuracy.

[0046] In a specific embodiment, the process of executing step S100 may specifically include the following steps:

[0047] The blank workpiece is rotated 360 degrees at equal intervals, and a high-resolution image is collected at each angle to obtain a multi-angle original image set containing at least 24 images;

[0048] Perform Gaussian filtering denoising, histogram equalization contrast enhancement and perspective transformation correction on each image in the multi-angle original image set to obtain a preprocessed multi-angle image set;

[0049] The pre-processed multi-angle image set is reconstructed based on structured light, and the depth information of each pixel is calculated through the triangulation principle to obtain the initial three-dimensional point cloud model;

[0050] The initial 3D point cloud model was voxelized, the voxel size was set to 0.5 mm, and the continuous point cloud data was converted into discrete voxel data to obtain a voxelized 3D model of 128x128x128 size;

[0051] The voxelized 3D model is input into the 3D convolutional neural network. The 3D convolutional neural network contains 5 3D convolutional layers. The convolution kernel size of each layer is 3x3x3, the step size is 1, the padding is 1, and the activation function is ReLU. The maximum pooling layer is added after the 2nd and 4th layers. The pooling kernel size is 2x2x2 to obtain a preliminary 3D feature map.

[0052] The preliminary 3D feature map is input into three residual blocks. Each residual block contains two 3D convolutional layers and a skip connection. The low-level and high-level features are fused by element-by-element addition operation to obtain a multi-scale fused 3D feature map.

[0053] The multi-scale fused three-dimensional feature map is globally averaged and pooled to compress the feature map to a size of 1x1x1. The feature dimension is then adjusted to 256 through a fully connected layer, and finally L2 normalization is performed to obtain a three-dimensional blank workpiece feature map.

[0054] Specifically, the blank workpiece is subjected to omnidirectional image acquisition. The workpiece is fixed on a rotating table, which rotates 360 degrees at equal intervals. A high-resolution image is acquired at each angle, and at least 24 images are acquired in total. Assume that the rotation interval angle is ,in Represents the number of collected images to ensure that the full view of the workpiece can be captured. These images constitute a multi-angle original image set. Each image in the multi-angle original image set is preprocessed to remove noise and enhance the image contrast. Each image is denoised by Gaussian filtering, and the kernel function of the Gaussian filter is:

[0055] ;

[0056] in, and are spatial coordinates, represents the standard deviation of the Gaussian filter. , control the degree of denoising, and strike a balance between maintaining image details and removing noise. Perform histogram equalization on the image to enhance contrast. Histogram equalization redistributes pixel intensity values ​​to make the brightness range of the image more balanced. Correct the image through perspective transformation to eliminate the distortion caused by the camera angle, so that the geometric shape of the workpiece in each image is closer to reality. After these three steps, a preprocessed multi-angle image set is obtained. The preprocessed multi-angle image set is reconstructed in three dimensions using a three-dimensional reconstruction method based on structured light. Structured light is a technology that restores the three-dimensional shape of an object by projecting a known grating pattern on the surface of the workpiece and capturing its deformed image. The depth information of each pixel is calculated using the principle of triangulation. Assuming that a triangle is formed between the projector, the camera, and the workpiece, the positions of the projector and the camera and the angle of the projected light are known, and the depth information is calculated based on the intersection position of the light and the surface of the object. Assuming that the baseline distance between the projector and the camera is , the projection angle is , the camera’s viewing angle is , then the depth of a certain point of the workpiece Calculated by triangulation formula:

[0057] ;

[0058] By performing similar depth calculations on each pixel, an initial 3D point cloud model is obtained. This point cloud model is a dense sampling of the workpiece surface, including the spatial position of each pixel. The initial 3D point cloud model is voxelized. The 3D space is divided into voxels, which are similar to pixels in a 2D image, and the continuous point cloud is represented as a discrete 3D voxel grid. The voxel size is set to 0.5 mm, that is, each voxel represents 0.5 mm of 3D space, and then the entire point cloud model is converted into a 3D grid of size The voxelized 3D model stores the structural information of the workpiece in the form of a 3D matrix. The voxelized 3D model is input into the 3D convolutional neural network for deep feature extraction. The 3D convolutional neural network contains 5 3D convolutional layers, and the convolution kernel size of each layer is , the step size is 1, the padding is 1, and the activation function is ReLU. The activation function ReLU is defined as:

[0059] ;

[0060] It can effectively introduce nonlinearity, making the neural network more expressive. A maximum pooling layer is added after the 2nd and 4th layers, and the pooling kernel size is , gradually reducing the size of the feature map through pooling operations while retaining the most significant features. After convolution and pooling operations, a preliminary three-dimensional feature map is obtained. The preliminary three-dimensional feature map is input into three residual blocks. Each residual block contains two three-dimensional convolutional layers and a jump connection, which fuses low-level and high-level features through element-by-element addition operations. The introduction of residual blocks effectively alleviates the gradient vanishing problem in deep networks. At the same time, low-level features are retained through direct jump connections, which improves the network's ability to express features of different scales and obtains a multi-scale fused three-dimensional feature map. Global average pooling is performed on the multi-scale fused three-dimensional feature map to compress the dimension of the feature map to The process of global average pooling is to take the average value of each channel in the feature map to obtain a compact feature vector. The feature dimension is adjusted to 256 dimensions through the fully connected layer, and finally L2 normalization is performed to ensure that the norm of the feature vector is 1. The formula for L2 normalization is:

[0061] ;

[0062] in, Represents the feature vector The L2 norm of Through L2 normalization, the feature vector is ensured to be on the unit sphere, which is convenient for similarity measurement and subsequent analysis. The final three-dimensional blank workpiece feature map is a compact and standardized feature vector that can effectively represent the overall shape and detail information of the workpiece.

[0063] In a specific embodiment, the process of executing step S200 may specifically include the following steps:

[0064] The feature map of the 3D rough workpiece is decomposed at multiple scales, and three feature maps with different resolutions are obtained through Gaussian pyramid downsampling, namely original resolution, 1 / 2 resolution and 1 / 4 resolution.

[0065] Perform edge detection on the feature map of each resolution to obtain three edge feature maps of different scales, and perform feature fusion on the three edge feature maps of different scales with the original feature map of the corresponding resolution to obtain three enhanced feature maps by element-by-element addition.

[0066] The three enhanced feature maps are transformed nonlinearly to obtain three nonlinear enhanced feature maps, and the three nonlinear enhanced feature maps are input into the feature pyramid network for multi-scale feature fusion to obtain a preliminary fused feature map;

[0067] The residual feature enhancement is performed on the preliminary fusion feature map to obtain a residual enhanced feature map, and the feature channels of the residual enhanced feature map are rearranged to obtain a multi-layer enhanced feature map.

[0068] Specifically, the feature map of the three-dimensional rough workpiece is decomposed at multiple scales, and three feature maps with different resolutions are generated using Gaussian pyramid downsampling. Gaussian pyramid is a multi-scale representation method that gradually reduces the resolution of an image. It constructs a series of scale images by applying Gaussian blur filtering to the original image and then downsampling. Suppose the original feature map is , through the Gaussian blur kernel Perform convolution on the feature map and then sample every two pixels to get Feature map of resolution Similarly, Gaussian blur and downsampling are performed again to obtain Feature map of resolution The formula is as follows:

[0069] , ;

[0070] in, Indicates The feature map of the layer, is a Gaussian kernel, * denotes a convolution operation, and Represents the coordinate position after downsampling. Through Gaussian pyramid downsampling, three feature maps of different scales are obtained: original resolution , Resolution and Resolution , which respectively describe the detailed information and overall structure of the workpiece at different scales. Perform edge detection on the feature map of each resolution to highlight the contour information. The purpose of edge detection is to identify the parts with drastic changes in the feature map, which represent important geometric features such as the edges and corners of the workpiece. Assume that the Sobel operator is used for edge detection. The Sobel operator detects edges based on calculating the image gradient. The gradient calculation formula is:

[0071] ;

[0072] in, Represents pixel The edge strength, and are the Sobel operators in the horizontal and vertical directions respectively, For the Layer feature map. By performing edge detection on the feature map of each resolution, three edge feature maps of different scales are obtained. . Perform feature fusion on the edge feature map of each resolution with the original feature map of the corresponding resolution. Combine the detail information in the original feature map with the contour information in the edge feature map to obtain an enhanced feature representation. Fuse the edge feature map and the original feature map together by element-by-element addition to obtain an enhanced feature map:

[0073] ;

[0074] in, Indicates The three enhanced feature maps are transformed nonlinearly to improve their feature expression capabilities. The model’s ability to express complex geometric features is enhanced by introducing nonlinear operations, which is achieved through activation functions (such as ReLU functions). is the ReLU activation function, then each enhanced feature map is transformed into a nonlinear enhanced feature map through nonlinear transformation ,in . Through nonlinear transformation, the nonlinear features of the enhanced feature map are improved. The three nonlinear enhanced feature maps are input into the feature pyramid network for multi-scale feature fusion. The features from different scales are effectively fused so that the final feature map contains both the detail information of the high-resolution features and the global information of the low-resolution features. Through feature pyramid fusion, a preliminary fused feature map is obtained, which can effectively describe the geometric features of the workpiece at different scales. In order to improve the expressive power of the features, the preliminary fused feature map is enhanced by residual features. Residual feature enhancement introduces a residual block to add the input and the convolved output of the preliminary fused feature map element by element to retain the original features and enhance its expressive power. Suppose the preliminary fused feature map is , after the convolution operation, the output feature map is , then the residual enhanced feature map is:

[0075] ;

[0076] in, Represents the feature map after residual enhancement. Through residual connection, the important information in the preliminary fusion features is effectively retained and further enhanced, making the feature map more stable when expressing complex geometric features. The feature channels of the residual enhanced feature map are rearranged to optimize the expression of features. By reorganizing the order and relationship between feature channels, the complementarity between features of different channels is better utilized. Channel rearrangement is achieved through the attention mechanism. By calculating the importance weight of each channel, the channels are reordered to obtain a multi-layer enhanced feature map.

[0077] In a specific embodiment, the process of executing step S300 may specifically include the following steps:

[0078] Perform feature statistical analysis on the multi-layer enhanced feature map, calculate the feature mean and variance of each spatial position, obtain a feature statistical map, and based on the feature statistical map, divide the feature space into a high-frequency area and a low-frequency area to obtain a region segmentation map;

[0079] Perform morphological processing on the region segmentation map to obtain an optimized region segmentation map, and perform weighted fusion on the optimized region segmentation map and the multi-layer enhanced feature map, assigning a higher weight to the high-frequency region and a lower weight to the low-frequency region, to obtain a preliminary adaptive guidance map;

[0080] Applying an edge-preserving filter to the preliminary adaptive guided map to process edge information and smooth the interior of the region to obtain a refined adaptive guided map, and constructing a spatial transformation parameter set of a spatial transformation network based on the refined adaptive guided map;

[0081] According to the spatial transformation parameter set, the multi-layer enhanced feature map is spatially adaptively transformed to obtain a feature map after spatial transformation, and the feature map after spatial transformation is adaptively normalized to obtain a target correction feature map.

[0082] Specifically, the feature mean and variance are calculated for each spatial position in the feature map to form a feature statistical map. The purpose of feature statistical analysis is to extract the basic statistical characteristics of each spatial position to help identify the feature differences in different regions. Suppose the feature map is ,in Represents the three-dimensional spatial position. Feature mean and characteristic variance Calculated by the following formula:

[0083] ;

[0084] ;

[0085] in, represents the number of feature channels, Represents the feature map Channels at position The value at . By calculating the feature mean and variance, the feature statistics graph is obtained. Based on the feature statistics graph, the feature space is divided into high-frequency area and low-frequency area to obtain a region segmentation graph. The high-frequency area usually corresponds to the part where the feature changes more dramatically, such as edges and corners, while the low-frequency area corresponds to the part where the feature changes more smoothly, such as a more uniform surface. Set the variance threshold, assuming that the threshold is set to , then the high frequency region satisfies , while the low-frequency region satisfies . The feature map is divided into two different regions according to the size of the feature variance to obtain a region segmentation map. The region segmentation map is morphologically processed to remove the noise and isolated points in the segmentation, making the region segmentation more continuous and smooth. Morphological processing includes operations such as dilation, erosion, opening and closing operations, which are used to fill small holes or eliminate small protrusions to optimize the morphological structure of the region. Through morphological processing, an optimized region segmentation map is obtained. The optimized region segmentation map is weightedly fused with the multi-layer enhanced feature map to generate a preliminary adaptive guidance map. Assume that the weight of the high-frequency area is , the weight of the low-frequency region is , then the fused preliminary adaptive guidance map It is expressed as:

[0086] ;

[0087] in, is the binary mask of the region segmentation map, and the high-frequency region , in the low frequency region . Through weighted fusion, the features of the high-frequency area are more prominently displayed in the preliminary adaptive guided map, so that the feature map has a stronger description ability for the complex shape and edge parts of the workpiece. The preliminary adaptive guided map is processed by an edge-preserving filter to smooth the interior of the area while maintaining the edge information. The purpose of the edge-preserving filter is to reduce the noise in the smooth area without losing the edge features, such as using a bilateral filter or a guided filter. Suppose the preliminary adaptive guided map is , after being processed by the edge-preserving filter, a refined adaptive guidance map is obtained , where the edge part is retained and other smooth areas are noise suppressed. Based on the refined adaptive guidance map, the spatial transformation parameter set of the spatial transformation network is constructed. The purpose of the spatial transformation network is to spatially adjust the multi-layer enhanced feature map to better meet the processing requirements and the geometric characteristics of the actual workpiece. The refined adaptive guidance map is extracted through convolution operations, and a set of spatial transformation parameters is generated, such as rotation, translation and scaling parameters. Assume that the spatial transformation parameter set is ,in represents the rotation matrix, represents the translation vector, Represents the scaling factor, and these parameters are used to perform geometric transformation on the feature map. According to the constructed spatial transformation parameter set, the multi-layer enhanced feature map is spatially adaptively transformed to obtain the feature map after spatial transformation. The spatial transformation operation includes geometric transformations such as rotation, translation and scaling of the feature map, so that the feature map is more consistent with the actual shape of the workpiece in terms of spatial structure. This process is expressed as:

[0088] ;

[0089] in, is the feature map after spatial transformation, is the position coordinate before transformation, , and are rotation, translation and scaling parameters respectively. Through this spatial transformation, the geometric structure of the feature map is adjusted to be more in line with the actual workpiece shape. The feature map after spatial transformation is adaptively normalized to obtain the target correction feature map. The values ​​in the feature map are readjusted so that they have consistent distribution characteristics within a specific range. Adaptive normalization calculates the mean and standard deviation for each channel and then normalizes the feature map so that the distribution of the feature map has better statistical characteristics. Suppose the feature map after spatial transformation is , then the target correction feature map after normalization is:

[0090] ;

[0091] in, and Represent the mean and standard deviation of the feature map respectively. Through normalization, the distribution of the target correction feature map is made within the unit range, enhancing the contrast and comparability between features.

[0092] In a specific embodiment, the process of executing step S400 may specifically include the following steps:

[0093] The target correction feature map is input into the first layer of the cascaded deep transposed convolutional network. The first layer of the cascaded deep transposed convolutional network contains 128 3x3 transposed convolution kernels with a step size of 2 and a padding of 1. The ReLU activation function and batch normalization are used to obtain a preliminary upsampled feature map.

[0094] Perform residual connection on the preliminary upsampled feature map and add it element by element with the target correction feature map to obtain a fused feature map;

[0095] The fused feature map is input into the second layer of the cascaded deep transposed convolutional network. The second layer of the cascaded deep transposed convolutional network contains 64 3x3 transposed convolution kernels with a step size of 2 and a padding of 1. The ReLU activation function and batch normalization are used to obtain a secondary upsampled feature map.

[0096] The secondary upsampled feature map is converted into graph structure data, each pixel is used as a node of the graph, and edges are established between adjacent pixels to obtain the initial graph structure;

[0097] The initial graph structure is input into the first layer of the graph convolutional network. The first layer of the graph convolutional network contains 32 graph convolution kernels. Each graph convolution kernel performs a convolution operation on the node and its first-order neighbors, and uses the ReLU activation function to obtain a local feature enhancement graph.

[0098] The local feature enhancement map is input into the second layer of the graph convolutional network. The second layer of the graph convolutional network contains 16 graph convolution kernels. Each graph convolution kernel performs a convolution operation on the node and its second-order neighbors, and uses the ReLU activation function to obtain the global feature enhancement map.

[0099] The edge of the global feature enhancement graph is refined, and the nodes in the graph are labeled and optimized using conditional random fields to obtain the target contour model.

[0100] Specifically, the target correction feature map is input into the first layer of the cascaded deep transposed convolutional network for feature upsampling. The purpose of the cascaded deep transposed convolutional network is to gradually increase the resolution of the feature map so that the features can be magnified and refined in space. The first deep transposed convolutional layer contains 128 layers of size The transposed convolution kernel is 2 with a stride of 2 and a padding of 1. The role of each convolution kernel is to expand the spatial dimension of the input feature map and introduce details through the convolution operation. The role of the transposed convolution (also known as deconvolution) is similar to upsampling. It expands the feature map by inserting zero elements and then performing convolution to increase the resolution of the feature map. Suppose the input feature map is , the output of the transposed convolution operation is , which is expressed by the following formula:

[0101] ;

[0102] Among them, ConvTranspose Represents the input feature map Using convolution kernel (Size is ) performs transposed convolution operation, step size ,filling BN stands for batch normalization, and ReLU is an activation function used to introduce nonlinear characteristics. After this layer of processing, the initial upsampled feature map is obtained. , whose resolution is twice that of the original feature map, and the details are enhanced. The preliminary upsampled feature map is subjected to residual connection operation and added element by element with the original target correction feature map to obtain the fused feature map. The purpose of residual connection is to increase the resolution of the feature map while retaining the original information of the input features, which helps to alleviate the problem of feature loss. Assume that the target correction feature map is , the initial up-sampled feature map is , then the fusion feature map It is expressed as:

[0103] ;

[0104] By adding elements one by one, the fused feature map contains both the upsampled detail information and the global information in the original features. The fused feature map is input to the second layer of the cascaded deep transposed convolutional network to increase the resolution of the feature map. This layer contains 64 layers of size The transposed convolution kernel has a step size of 2 and a padding of 1. Similar to the first layer, the purpose of this layer is to further expand the resolution of the fused feature map to generate a secondary upsampled feature map. . This process also includes transposed convolution operations, batch normalization, and ReLU activation functions. Through upsampling, the spatial dimension of the feature map is further expanded, and important detail information is retained in the process. The twice upsampled feature map is converted into a graph structure representation. Each pixel is regarded as a node in the graph, and edges are established between adjacent pixels to build the initial graph structure. Assume that the spatial dimension of the feature map is , then the graph structure contains nodes, each node corresponds to a pixel in the feature map. Edges are established between adjacent pixels to form an edge set of the graph. The graph structure representation effectively captures the spatial neighbor relationship. The initial graph structure is input into the first layer of the graph convolutional network for processing, and local features are enhanced on the graph structure. The first layer of the graph convolutional network contains 32 graph convolution kernels, each of which performs a convolution operation on a node and its first-order neighbors, and uses the ReLU activation function for nonlinear transformation. Suppose the initial graph structure is ,in represents a node set, represents the edge set, and the node features are , then the update formula of graph convolution is:

[0105] ;

[0106] in, Representation Node In the The characteristics of the layer, Representation Node The first-order neighbor of is the weight matrix of the graph convolution kernel, Represents the number of neighbor nodes, and updates the node features by weighted averaging the features of neighbor nodes. After the first layer of graph convolution processing, a local feature enhancement map is obtained, in which the features of each node not only contain its own information, but also integrate the information of its first-order neighbors to achieve local feature aggregation. The local feature enhancement map is input into the second layer of the graph convolution network for global feature enhancement. The second layer contains 16 graph convolution kernels, each of which performs convolution operations on the node and its second-order neighbors to capture a wider range of global information. In this way, the features of the node are not only related to its directly adjacent nodes, but also to its more distant neighbor nodes, resulting in a global feature enhancement map. The second layer of graph convolution also uses the ReLU activation function to further enhance the nonlinear representation ability of the features. After that. The global feature enhancement map is subjected to edge refinement processing, and the node labels in the graph are optimized using conditional random fields. Conditional random fields are a probabilistic graph model that can globally optimize the node labels in the graph to ensure that the details of the edge information are retained while smoothing the interior of the region. Let the global feature enhancement map be , the node label is , then the energy function of the conditional random field is expressed as:

[0107] ;

[0108] in, Representation Node The potential function is used to measure the matching degree between node labels and their features. Representation Node and nodes The potential function between is used to encourage adjacent nodes to have similar labels. By minimizing the energy function , perform global optimization on the node labels and obtain the target contour model.

[0109] In a specific embodiment, the process of executing step S500 may specifically include the following steps:

[0110] The target contour model is adaptively divided into multiple layers, and the contour is divided into multiple sub-areas based on the curvature change rate. The curvature change rate of each sub-area does not exceed a preset threshold, and a block contour model is obtained;

[0111] A local coordinate system is constructed for each sub-region in the block contour model, and the global contour information is converted into a local feature representation to obtain a local feature set;

[0112] The local feature set is input into the pre-trained deep reinforcement learning agent network. The pre-trained deep reinforcement learning agent network contains 4 fully connected layers. Each layer uses the ReLU activation function and the output layer uses the Softmax function to obtain the initial processing parameter action value.

[0113] Based on the initial machining parameter action value, a multi-objective reward function is constructed, including three objectives: machining accuracy, machining efficiency, and tool life. The performance of the current action value is comprehensively evaluated by weighted summation to obtain a comprehensive reward value.

[0114] The comprehensive reward value is input into the advantage function estimator of the deep reinforcement learning agent network. The advantage function estimator consists of two fully connected layers and uses the Tanh activation function to calculate the advantage of the action value in the current state to obtain the advantage function value.

[0115] Based on the advantage function value, the policy gradient algorithm is used to update the parameters of the deep reinforcement learning agent network, and the processing parameters are iteratively optimized to obtain the optimized processing parameter set;

[0116] The feasibility of the optimized machining parameter set is verified. Through kinematic and dynamic constraint checks, the parameter combinations that do not meet the physical limitations of the machine tool are eliminated to obtain a feasible machining parameter set.

[0117] The parameters in the feasible machining parameter set are interpolated and smoothed to ensure the continuity and smoothness of the parameters between adjacent sub-regions, and the dynamic optimal machining parameter set is obtained.

[0118] Specifically, the target contour model is adaptively divided into multiple layers, and the contour is divided into multiple sub-regions based on the curvature change rate to ensure that the curvature change rate of each sub-region does not exceed the preset threshold, so that the shape characteristics of each region are relatively smooth. The curvature change rate is used to measure the curvature of the contour curve and is defined as the change rate of the normal vector along the curve. Assume that the curvature on the curve is ,in represents the arc length of the curve, and the rate of change of curvature is expressed by calculating the difference between adjacent curvatures:

[0119] ;

[0120] in, represents the rate of change of curvature, is the arc length increment of the sampling. By setting a preset threshold , take the part where the curvature change rate does not exceed the threshold as a sub-region, ensure that each sub-region is easy to maintain consistent process parameters during the processing to obtain a block contour model. Construct a local coordinate system for each sub-region in the block contour model. Convert the global contour information into a local feature representation in each sub-region, thereby reducing the complexity and making the feature description more adaptable to the needs of local processing. Set the center position of a sub-region as the origin, and obtain a local feature set by reprojecting the features of the region into the local coordinate system, which contains the shape and spatial information in the sub-region. Through the representation of local coordinates, the processing parameters can be calculated more accurately and the errors caused by global differences can be reduced. The local feature set is input into the pre-trained deep reinforcement learning agent network to optimize the processing parameters. The pre-trained deep reinforcement learning agent network contains 4 fully connected layers, each of which uses the ReLU activation function to increase the nonlinear expression ability of the model so that it can better cope with complex feature combinations. The last layer is the output layer, which uses the Softmax activation function to map the action value of the processing parameter to a probability distribution, thereby selecting the optimal action value. Let the input local feature be , the output of the network is the initial processing parameter action value , then the calculation formula of the output layer is:

[0121] ;

[0122] in, Indicates that the network is in the output layer The input of a neuron, It is The probability of a processing parameter action, is the number of output actions. Through the Softmax function, a set of initial processing parameter action values ​​are obtained, including variables such as feed speed and cutting depth. Based on the initial processing parameter action values, a multi-objective reward function is constructed to evaluate the performance of these parameters. The reward function includes three objectives: processing accuracy, processing efficiency, and tool life. The objectives are comprehensively evaluated by weighted summation. Let the reward for accuracy be , the efficiency reward is , the tool life bonus is , then the calculation formula of the comprehensive reward value R is:

[0123] ;

[0124] in, are the weight coefficients of accuracy, efficiency and life, respectively, which are used to adjust the importance of different goals. Through the comprehensive reward function, the current processing parameter action is comprehensively evaluated to determine its superiority under multi-objective conditions. The comprehensive reward value is input into the advantage function estimator of the deep reinforcement learning agent network to improve the effect of parameter optimization. The advantage function estimator consists of two fully connected layers and uses the Tanh activation function to calculate the advantage of the action value in the current state. The purpose of the advantage function is to measure the superiority of the current action relative to the average strategy to help the network focus on those actions that perform better. Suppose the comprehensive reward value is , the action value in the current state is , then the calculation formula of the advantage function is:

[0125] ;

[0126] Among them, s represents the current state, Indicates the current action. is the discount factor, Represents the reward obtained by performing action a in state s. By calculating the advantage function value, the pros and cons of the current action are determined, providing guidance for the parameter update of the deep reinforcement learning agent network. Based on the advantage function value, the policy gradient algorithm is used to update the parameters of the deep reinforcement learning agent network to optimize the processing parameters. The policy gradient algorithm is an algorithm that optimizes the strategy by gradient ascent, aiming to maximize the cumulative reward. Through multiple iterative optimizations, the agent network gradually learns to select those processing parameters that can maximize the comprehensive reward, and finally obtains the optimized processing parameter set. The feasibility of the optimized processing parameter set is verified. The kinematic and dynamic constraint checks are used to verify whether these parameters meet the physical limitations of the machine tool. For example, some parameter combinations cause the machine tool to exceed its range of motion or bear excessive torque. Therefore, it is necessary to eliminate the parameter combinations that do not meet the physical constraints and obtain a feasible processing parameter set that meets the physical limitations of the machine tool. In order to ensure that the processing parameters between adjacent sub-areas have good continuity and smoothness, the feasible processing parameter set is interpolated and smoothed to avoid sudden changes in parameters during the processing process, reduce the impact on the machine tool, and improve the surface quality of the processing. Assume that the parameters of the two adjacent sub-areas are and , then these two parameters are smoothed by linear interpolation:

[0127] ;

[0128] in, represents the interpolated parameters, is the interpolation factor, which controls the transition between the two parameters. Through interpolation smoothing, the change of parameters during the processing is ensured to be smooth, thereby improving the stability and accuracy of the processing. Finally, a dynamic optimal processing parameter set is obtained, which can adapt to the characteristics of different sub-areas and maintain continuity and smoothness between adjacent areas.

[0129] In a specific embodiment, the process of executing step S600 may specifically include the following steps:

[0130] Inputting the dynamic optimal machining parameter set into the self-correcting finite element model, constructing the initial geometric model and material property model of the workpiece, and obtaining the initial finite element model;

[0131] The initial finite element model is meshed and boundary conditions are set. The adaptive mesh refinement algorithm is used to perform local mesh encryption in the predicted high stress area to obtain the optimized finite element mesh model.

[0132] The multi-source processing data collected in real time, including cutting force, temperature and vibration data, are input into the optimized finite element mesh model for static and dynamic analysis to obtain stress distribution and deformation field;

[0133] The stress distribution and deformation field are subjected to feature extraction and feature dimensionality reduction to obtain low-dimensional feature vectors, which are then input into a recursive neural network predictor. The recursive neural network predictor includes an LSTM layer and two fully connected layers. The tanh activation function is used to predict the deformation of the workpiece at future moments to obtain a predicted deformation field.

[0134] Based on the predicted deformation field, the back propagation algorithm is used to update the parameters of the self-correcting finite element model, including adjusting the material properties and boundary conditions, to obtain a corrected finite element model;

[0135] The output of the corrected finite element model is compared with the actual measured deformation data, the error is calculated and a compensation function is constructed. The compensation function maps the predicted error to the correction amount of the tool path to obtain the compensation strategy;

[0136] Based on the compensation strategy, real-time adjustment instructions for the tool path are generated and converted into motion instructions for each axis in combination with the machine tool kinematics inverse algorithm to obtain the target compensation instruction sequence.

[0137] Specifically, the dynamic optimal processing parameter set is input into the self-correcting finite element model to construct the initial geometric model and material property model of the workpiece and obtain the initial finite element model. The initial finite element model is meshed and boundary conditions are set for numerical solution. Meshing is to divide the entire geometric model into several small units so that finite element solution can be performed. An adaptive mesh refinement algorithm is used to improve the accuracy of high stress areas. By locally encrypting the mesh in the predicted high stress areas, a denser mesh structure is established in these areas to improve the accuracy of stress and deformation analysis. Suppose the initial mesh model is , the local stress is , if the stress exceeds a certain threshold , then the mesh is encrypted in this area to obtain the optimized finite element mesh model The process of local mesh refinement is described as:

[0138] ;

[0139] Among them, Refine Represents the grid Perform refinement operations. is the stress threshold. Through adaptive meshing, the density of the mesh is higher in key high-stress areas, which can more accurately capture the stress and deformation characteristics in these areas. The multi-source processing data collected in real time is input into the optimized finite element mesh model for static and dynamic analysis. These processing data include cutting force, temperature, vibration, etc. These parameters reflect the stress and environmental changes of the workpiece during the actual processing process. Static analysis is used to solve the stress distribution and deformation field of the workpiece under the current stress and temperature conditions, while dynamic analysis is used to solve the stress and deformation changes of the workpiece in the time dimension. By inputting these multi-source data, the stress distribution and deformation field of the workpiece during the processing process are obtained. In order to simplify the calculation and extract effective information, feature extraction and feature dimensionality reduction are performed on the stress distribution and deformation field to obtain a low-dimensional feature vector. It is assumed that the characteristics of the deformation field are high-dimensional vectors , through dimensionality reduction methods (such as principal component analysis), we can obtain a low-dimensional feature vector :

[0140] ;

[0141] in, represents the low-dimensional feature vector after dimensionality reduction, is a dimensionality reduction matrix, and its column vector is the retained principal component. In this way, the high-dimensional feature representation is converted into a compact low-dimensional representation, reducing the computational complexity and retaining the most important information. The low-dimensional feature vector is input into the recurrent neural network predictor to predict the deformation of the workpiece at future moments. Recurrent neural networks are suitable for processing input data with time series characteristics and can capture the temporal correlation of features. The predictor contains a long short-term memory (LSTM) layer and two fully connected layers, where the LSTM layer is used to capture the dependencies in the time series, and the two fully connected layers use the tanh activation function to predict the future deformation of the workpiece. Suppose the low-dimensional feature of the input is , the output of LSTM is , the prediction value after the fully connected layer is

[0142] ;

[0143] in, is the output state of LSTM, is the weight matrix of the fully connected layer, is the tanh activation function. Through the prediction of the recurrent neural network, the deformation field of the workpiece at the future moment is obtained to provide data support for compensation control. Based on the predicted deformation field, the back propagation algorithm is used to update the parameters of the self-correcting finite element model. The updated parameters include material properties and boundary conditions, so that the finite element model can better fit the actual workpiece deformation. Through back propagation, the error between the model prediction result and the actual result is calculated, and the parameters of the finite element model are adjusted by minimizing the error. Assume that the actual deformation data is , the predicted deformation is , the error function is:

[0144] ;

[0145] in, is the error function, and Respectively The predicted value and actual value of each node are minimized , update the parameters of the model and obtain the corrected finite element model. Compare the output of the corrected finite element model with the actual measured deformation data, calculate the error and construct a compensation function. The role of the compensation function is to map the predicted error to the correction amount of the tool path and obtain the compensation strategy. Let the error be , then the compensation function is expressed as:

[0146] ;

[0147] in, Indicates the correction amount of the tool path, is a compensation function used to convert deformation errors into corresponding tool position adjustments. Through the compensation function, the tool path is dynamically corrected according to the predicted error to ensure machining accuracy. Based on the compensation strategy, real-time adjustment instructions for the tool path are generated. In order to adapt these adjustment instructions to the physical operation of the machine tool, the correction amount of the tool path is converted into motion instructions for each axis in combination with the inverse solution algorithm of the machine tool kinematics. The inverse solution algorithm of the machine tool kinematics is used to decompose the global tool position and posture requirements into displacement instructions for each axis to ensure that the machine tool can execute the compensation strategy. After this process, the target compensation instruction sequence is obtained, which guides the various axes of the machine tool to perform real-time compensation actions, so as to always maintain the expected workpiece contour and accuracy during the machining process.

[0148] See also Figure 2 , Figure 2 A schematic block diagram of the structure of a contour processing device 200 for a multi-axis linkage machine tool provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, a contour processing device 200 for a multi-axis linkage machine tool comprises:

[0149] The acquisition module 210 is used to perform multi-angle image acquisition and three-dimensional convolution feature extraction on the rough workpiece to obtain a three-dimensional rough workpiece feature map;

[0150] An enhancement module 220 is used to input the three-dimensional rough workpiece feature map into a multi-scale feature enhancement network to perform contour feature hierarchical enhancement processing to obtain a multi-layer enhanced feature map;

[0151] The correction module 230 is used to perform spatial adaptive transformation and feature correction on the multi-layer enhanced feature map to obtain a target corrected feature map;

[0152] The reconstruction module 240 is used to input the target correction feature map into the cascaded deep transposed convolutional network and the graph convolutional network for contour iterative reconstruction and refinement processing to obtain a target contour model;

[0153] A calculation module 250 is used to perform multi-objective optimization calculation on the processing parameters based on the target contour model to obtain a dynamic optimal processing parameter set;

[0154] The control module 260 is used to perform real-time workpiece deformation prediction and adaptive compensation control on the dynamic optimal processing parameter set and the multi-source processing data collected in real time to obtain a target compensation instruction sequence.

[0155] Through the collaboration of the above components, multi-angle image acquisition and three-dimensional convolution feature extraction, combined with a multi-scale feature enhancement network, high-precision recognition of complex surface contours is achieved. The spatial adaptive transformation and feature correction technology can dynamically adjust the feature processing strategy according to the characteristics of different workpieces, improving the robustness and adaptability of contour recognition. The cascaded deep transposed convolutional network and graph convolutional network are used for iterative contour reconstruction and refinement, which can capture tiny contour details and improve the accuracy of the contour model. The multi-objective optimization algorithm based on deep reinforcement learning can simultaneously consider multiple goals such as machining accuracy, efficiency and tool life, and realize intelligent dynamic optimization of machining parameters. Combined with the self-correcting finite element model and the recursive neural network predictor, the real-time prediction of the deformation of the workpiece during the machining process is realized. By constructing a compensation function and adjusting the tool path in real time, dynamic compensation for errors in the machining process is realized, significantly improving the machining accuracy.

[0156] See also Figure 3 , Figure 3 A schematic block diagram of the structure of a contour processing device 300 for a multi-axis linkage machine tool provided in an embodiment of the present application, wherein the contour processing device 300 for a multi-axis linkage machine tool includes a processor 301 and a memory 302, wherein the processor 301 and the memory 302 are connected via a device bus 303, wherein the memory 302 may include a non-volatile storage medium and an internal memory.

[0157] The non-volatile storage medium can store a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor 301, the processor 301 can execute any of the above-mentioned contour processing methods for a multi-axis linkage machine tool.

[0158] The processor 301 is used to provide computing and control capabilities to support the operation of the entire contour processing device 300 for a multi-axis linkage machine tool.

[0159] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned contour processing methods for a multi-axis linkage machine tool.

[0160] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present application, and does not constitute a limitation on the contour processing equipment 300 for multi-axis linkage machine tools involved in the scheme of the present application. The specific contour processing equipment 300 for multi-axis linkage machine tools may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0161] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0162] It should be noted that technical personnel in the relevant field can clearly understand that, for the convenience and conciseness of description, the specific working process of the contour processing equipment 300 for multi-axis linkage machine tools described above can refer to the corresponding process of the aforementioned contour processing method for multi-axis linkage machine tools, and will not be repeated here.

[0163] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by one or more processors, the one or more processors implement the contour processing method for a multi-axis linkage machine tool as provided in the embodiment of the present application.

[0164] The computer-readable storage medium may be an internal storage unit of the contour processing device 300 for a multi-axis linkage machine tool in the aforementioned embodiment, such as a hard disk or memory of the contour processing device 300 for a multi-axis linkage machine tool. The computer-readable storage medium may also be an external storage device of the contour processing device 300 for a multi-axis linkage machine tool, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc., equipped with the contour processing device 300 for a multi-axis linkage machine tool.

[0165] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0166] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0167] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A contour processing method for a multi-axis linkage machine tool, characterized in that: include: The blank workpiece is subjected to multi-angle image acquisition and three-dimensional convolution feature extraction to obtain a three-dimensional blank workpiece feature map; specifically comprising: rotating the blank workpiece 360 ​​degrees at equal intervals, acquiring high-resolution images at each angle, and obtaining a multi-angle original image set containing at least 24 images; performing Gaussian filtering denoising, histogram equalization contrast enhancement and perspective transformation correction on each image in the multi-angle original image set to obtain a pre-processed multi-angle image set; performing three-dimensional reconstruction based on structured light on the pre-processed multi-angle image set, calculating the depth information of each pixel through the triangulation principle, and obtaining an initial three-dimensional point cloud model; performing voxelization processing on the initial three-dimensional point cloud model, setting the voxel size to 0.5 mm, converting continuous point cloud data into discrete voxel data, and obtaining a 128x128x128 large A small voxelized three-dimensional model; inputting the voxelized three-dimensional model into a three-dimensional convolutional neural network, the three-dimensional convolutional neural network comprising five three-dimensional convolutional layers, the convolution kernel size of each layer being 3x3x3, the step size being 1, the padding being 1, the activation function being ReLU, adding a maximum pooling layer after the second and fourth layers, the pooling kernel size being 2x2x2, and obtaining a preliminary three-dimensional feature map; inputting the preliminary three-dimensional feature map into three residual blocks, each residual block comprising two three-dimensional convolutional layers and a skip connection, fusing low-level and high-level features through element-by-element addition operations, and obtaining a multi-scale fused three-dimensional feature map; performing global average pooling on the multi-scale fused three-dimensional feature map, compressing the feature map to a size of 1x1x1, and then adjusting the feature dimension to 256 through a fully connected layer, and finally performing L2 normalization to obtain a three-dimensional rough workpiece feature map; Inputting the three-dimensional rough workpiece feature map into a multi-scale feature enhancement network to perform contour feature hierarchical enhancement processing to obtain a multi-layer enhanced feature map; Performing spatial adaptive transformation and feature correction on the multi-layer enhanced feature map to obtain a target corrected feature map; Inputting the target correction feature map into the cascaded deep transposed convolutional network and the graph convolutional network for contour iterative reconstruction and refinement processing to obtain a target contour model; Based on the target contour model, multi-objective optimization calculation is performed on the processing parameters to obtain a dynamic optimal processing parameter set; The dynamic optimal processing parameter set and the multi-source processing data collected in real time are used to predict the deformation of the workpiece in real time and to adaptively control the compensation, so as to obtain a target compensation instruction sequence; specifically, the dynamic optimal processing parameter set is input into a self-correcting finite element model, an initial geometric model and a material property model of the workpiece are constructed, and an initial finite element model is obtained; meshing and boundary condition setting are performed on the initial finite element model, and an adaptive mesh refinement algorithm is used to perform local mesh encryption in the predicted high stress area to obtain an optimized finite element mesh model; the multi-source processing data collected in real time, including cutting force, temperature and vibration data, is input into the optimized finite element mesh model, and static and dynamic analysis is performed to obtain stress distribution and deformation field; feature extraction and feature dimensionality reduction are performed on the stress distribution and the deformation field to obtain a low-dimensional feature vector. The low-dimensional feature vector is input into a recursive neural network predictor, which includes an LSTM layer and two fully connected layers, and uses a tanh activation function to predict the deformation of the workpiece at a future moment to obtain a predicted deformation field; based on the predicted deformation field, the parameters of the self-correcting finite element model are updated using a back propagation algorithm, including adjusting material properties and boundary conditions to obtain a corrected finite element model; the output of the corrected finite element model is compared with the actual measured deformation data, the error is calculated and a compensation function is constructed, the compensation function maps the predicted error to a correction amount of the tool path to obtain a compensation strategy; based on the compensation strategy, a real-time adjustment instruction of the tool path is generated, and combined with the machine tool kinematics inverse solution algorithm, it is converted into motion instructions of each axis to obtain a target compensation instruction sequence.

2. The contour processing method for a multi-axis linkage machine tool according to claim 1, characterized in that: The step of inputting the three-dimensional rough workpiece feature map into a multi-scale feature enhancement network to perform contour feature hierarchical enhancement processing to obtain a multi-layer enhanced feature map includes: The three-dimensional rough workpiece feature map is decomposed at multiple scales, and three feature maps with different resolutions are obtained by downsampling the Gaussian pyramid, which are original resolution, 1 / 2 resolution and 1 / 4 resolution respectively; Perform edge detection on the feature map of each resolution to obtain three edge feature maps of different scales, and perform feature fusion on the three edge feature maps of different scales with the original feature map of the corresponding resolution to obtain three enhanced feature maps by element-by-element addition; Performing nonlinear transformation on the three enhanced feature maps respectively to obtain three nonlinear enhanced feature maps, and inputting the three nonlinear enhanced feature maps into a feature pyramid network to perform multi-scale feature fusion to obtain a preliminary fused feature map; The preliminary fusion feature map is subjected to residual feature enhancement to obtain a residual enhanced feature map, and the residual enhanced feature map is subjected to feature channel rearrangement to obtain a multi-layer enhanced feature map.

3. The contour processing method for a multi-axis linkage machine tool according to claim 1, characterized in that: The step of performing spatial adaptive transformation and feature correction on the multi-layer enhanced feature map to obtain a target corrected feature map includes: Performing feature statistical analysis on the multi-layer enhanced feature map, calculating the feature mean and variance of each spatial position to obtain a feature statistical map, and based on the feature statistical map, dividing the feature space into a high-frequency region and a low-frequency region to obtain a region segmentation map; Performing morphological processing on the region segmentation map to obtain an optimized region segmentation map, and weighted fusion of the optimized region segmentation map and the multi-layer enhanced feature map, assigning a higher weight to the high-frequency region and a lower weight to the low-frequency region, to obtain a preliminary adaptive guidance map; Applying an edge-preserving filter to the preliminary adaptive guided map to perform edge information processing and region internal smoothing to obtain a refined adaptive guided map, and constructing a spatial transformation parameter set of a spatial transformation network based on the refined adaptive guided map; According to the spatial transformation parameter set, the multi-layer enhanced feature map is spatially adaptively transformed to obtain a feature map after spatial transformation, and the feature map after spatial transformation is adaptively normalized to obtain a target correction feature map.

4. The contour processing method for a multi-axis linkage machine tool according to claim 1, characterized in that: The target correction feature map is input into the cascaded deep transposed convolutional network and the graph convolutional network for contour iterative reconstruction and refinement processing to obtain a target contour model, including: Inputting the target correction feature map into the first layer of a cascaded deep transposed convolutional network, wherein the first layer of the cascaded deep transposed convolutional network comprises 128 3x3 transposed convolution kernels, a step size of 2, a padding of 1, and uses a ReLU activation function and batch normalization to obtain a preliminary up-sampled feature map; Performing a residual connection on the preliminary up-sampled feature map, and adding it element by element with the target corrected feature map to obtain a fused feature map; Inputting the fused feature map into the second layer of the cascaded deep transposed convolutional network, the second layer of the cascaded deep transposed convolutional network contains 64 3x3 transposed convolution kernels, a step size of 2, a padding of 1, and uses a ReLU activation function and batch normalization to obtain a secondary up-sampled feature map; Convert the secondary up-sampled feature map into graph structure data, with each pixel as a node of the graph and edges connected between adjacent pixels to obtain an initial graph structure; Inputting the initial graph structure into the first layer of a graph convolutional network, the first layer of the graph convolutional network comprising 32 graph convolution kernels, each of which performs a convolution operation on a node and its first-order neighbors, and uses a ReLU activation function to obtain a local feature enhancement graph; Inputting the local feature enhancement map into the second layer of the graph convolutional network, the second layer of the graph convolutional network includes 16 graph convolution kernels, each of which performs a convolution operation on a node and its second-order neighbors, and uses a ReLU activation function to obtain a global feature enhancement map; The global feature enhancement graph is subjected to edge refinement processing, and the nodes in the graph are labeled and optimized using a conditional random field to obtain a target contour model.

5. The contour processing method for a multi-axis linkage machine tool according to claim 1, characterized in that: The multi-objective optimization calculation of the processing parameters is performed based on the target contour model to obtain a dynamic optimal processing parameter set, including: Adaptively performing multi-layer block division on the target contour model, dividing the contour into multiple sub-regions based on the curvature change rate, wherein the curvature change rate of each sub-region does not exceed a preset threshold, and obtaining a block contour model; Constructing a local coordinate system for each sub-region in the block contour model, converting global contour information into a local feature representation, and obtaining a local feature set; Inputting the local feature set into a pre-trained deep reinforcement learning agent network, wherein the pre-trained deep reinforcement learning agent network comprises four fully connected layers, each layer uses a ReLU activation function, and the output layer uses a Softmax function to obtain an initial processing parameter action value; Based on the initial machining parameter action value, a multi-objective reward function is constructed, including three objectives: machining accuracy, machining efficiency and tool life, and the performance of the current action value is comprehensively evaluated by weighted summation to obtain a comprehensive reward value; Input the comprehensive reward value into the advantage function estimator of the deep reinforcement learning agent network, wherein the advantage function estimator is composed of two fully connected layers, and uses the Tanh activation function to calculate the advantage of the action value in the current state to obtain the advantage function value; Based on the advantage function value, the policy gradient algorithm is used to update the parameters of the deep reinforcement learning agent network, and the processing parameters are iteratively optimized to obtain an optimized processing parameter set; Performing feasibility verification on the optimized machining parameter set, eliminating parameter combinations that do not meet the physical limitations of the machine tool through kinematic and dynamic constraint checks, and obtaining a feasible machining parameter set; The parameters in the feasible processing parameter set are interpolated and smoothed to ensure the continuity and smoothness of the parameters between adjacent sub-regions, so as to obtain a dynamic optimal processing parameter set.

6. A contour processing device for a multi-axis linkage machine tool, characterized in that: A method for performing contour machining for a multi-axis linkage machine tool as claimed in any one of claims 1 to 5, comprising: The acquisition module is used to perform multi-angle image acquisition and three-dimensional convolution feature extraction on the rough workpiece to obtain a three-dimensional rough workpiece feature map; An enhancement module is used to input the three-dimensional rough workpiece feature map into a multi-scale feature enhancement network to perform contour feature hierarchical enhancement processing to obtain a multi-layer enhanced feature map; A correction module, used for performing spatial adaptive transformation and feature correction on the multi-layer enhanced feature map to obtain a target corrected feature map; A reconstruction module is used to input the target correction feature map into a cascaded deep transposed convolutional network and a graph convolutional network for contour iterative reconstruction and refinement processing to obtain a target contour model; A calculation module, used for performing multi-objective optimization calculation on the processing parameters based on the target contour model to obtain a dynamic optimal processing parameter set; The control module is used to perform real-time workpiece deformation prediction and adaptive compensation control on the dynamic optimal processing parameter set and multi-source processing data collected in real time to obtain a target compensation instruction sequence.

7. A contour processing device for a multi-axis linkage machine tool, characterized in that: The contour processing device for a multi-axis linkage machine tool comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the contour processing device for a multi-axis linkage machine tool to execute the contour processing method for a multi-axis linkage machine tool according to any one of claims 1 to 5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, a contour processing method for a multi-axis linkage machine tool as described in any one of claims 1 to 5 is implemented.

Images