An image processing-based method and system for forging forming of an aluminum alloy tube target

Abstract

The application provides a kind of aluminum alloy pipe target's forging forming method and system based on image processing, it is related to image processing technical field, including establishing three-dimensional digital grid model and topological relationship diagram;Multi-view image is obtained using dual-spectrum camera array and fusion;Temperature radiation intensity mapping image is reconstructed by neural radiation field network;Image enhancement processing is executed;Contour feature data is extracted using multi-scale texture decoupling network, and reflection, oxidation and forging components are separated;Deformation trend is predicted based on topological relationship diagram and contour feature data, and forging parameter adjustment instruction is generated;Forging equipment is controlled according to the instruction and real-time optimization until the forming quality is up to standard.The method realizes the accurate monitoring and intelligent control of aluminum alloy pipe target forging process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a forging forming method and system for an aluminum alloy pipe target based on image processing. BACKGROUND

[0002] As a key component in the field of advanced manufacturing, the aluminum alloy pipe target has long been plagued by technical bottlenecks such as insufficient precision and unstable quality in the forging forming process. Traditional forging methods mainly rely on empirical parameter setting and manual adjustment, making it difficult to meet the high-precision forming needs of complex-shaped aluminum alloy pipe targets. This method not only has low efficiency, but also relies heavily on the technical level of the operator, resulting in poor product consistency, high scrap rate, and inability to meet the demand for high-quality aluminum alloy pipe targets in modern industry.

[0003] With the development of industrial automation and intelligent manufacturing, image processing technology has been widely applied in the manufacturing field, but its application in the forging process of aluminum alloy pipe targets still has many limitations. Existing technologies mainly use single-spectrum image acquisition, which cannot effectively capture the complex optical properties of aluminum alloy during high-temperature forging. At the same time, the image data processing method is relatively simple, lacking a deep understanding of the material deformation mechanism, resulting in inaccurate forging parameter optimization and difficulty in predicting and controlling the deformation trend of aluminum alloy pipe targets during the forging process. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a forging forming method and system for an aluminum alloy pipe target based on image processing, which can solve the problems in the prior art.

[0005] In a first aspect of the embodiments of the present application, a forging forming method for an aluminum alloy pipe target based on image processing is provided, comprising:

[0006] A three-dimensional digital model of the aluminum alloy pipe target is established, the three-dimensional digital model is divided into grid elements, and a topological relationship diagram of the grid elements is established;

[0007] A dual-spectrum camera array is used to collect forging forming images of the aluminum alloy pipe target from multiple viewing angles, and adaptive exposure compensation is performed to obtain multi-view fusion images. Camera pose parameters in the multi-view fusion images are extracted to obtain corresponding viewing direction information. The three-dimensional space coordinates of the grid elements and the viewing direction information are respectively subjected to sinusoidal position encoding, and input into a neural radiance field network for volume rendering reconstruction to obtain a temperature radiation intensity mapping image;

[0008] Performing phase consistency registration and wavelet transform denoising on the temperature radiation intensity mapping image to obtain an enhanced aluminum alloy tube target image; inputting the enhanced aluminum alloy tube target image into a multi-scale texture decoupling network, obtaining local texture descriptors through multi-scale decomposition and texture feature extraction, inputting the local texture descriptors into a component separation unit to decouple and separate the reflection component, oxidation component and forging component, and outputting the profile feature data of the aluminum alloy tube target;

[0009] Based on the topological relationship graph of the grid unit and the profile feature data, predicting the deformation trend through graph structuring calculation to generate a forging parameter adjustment instruction; controlling the forging equipment according to the forging parameter adjustment instruction.

[0010] In an alternative embodiment:

[0011] Establishing a three-dimensional digital model of the aluminum alloy tube target, dividing the three-dimensional digital model into grid units, and establishing a topological relationship graph of the grid units includes:

[0012] Establishing a three-dimensional NURBS surface model of the aluminum alloy tube target, the three-dimensional NURBS surface model including a cylindrical surface and an end surface, the cylindrical surface and the end surface being constructed through control point coordinates, weight factors and B-spline basis functions;

[0013] Grid division is performed on the three-dimensional NURBS surface model, and an adaptive grid density control function is established according to a curvature function, the adaptive grid density control function calculating grid unit size based on a minimum grid size, a maximum grid size, a direction weight coefficient and a grid density control parameter to generate tetrahedral grid units;

[0014] A topological relationship graph is constructed based on the tetrahedral grid units, the topological relationship graph including a grid node set and a node connection edge set, the grid node set being taken as vertices and the node connection edge set being taken as edges.

[0015] In an alternative embodiment:

[0016] A dual-spectrum camera array is used to collect forging formation images of the aluminum alloy tube target from multiple viewing angles, adaptive exposure compensation is performed on the forging formation images to obtain a multi-view fusion image, camera pose parameters in the multi-view fusion image are extracted to obtain corresponding viewing direction information, and the three-dimensional space coordinates of the grid unit and the viewing direction information are respectively subjected to sinusoidal position encoding, and the encoded features are input into a neural radiance field network; the step of obtaining a temperature radiation intensity mapping image through numerical integration of the neural radiance field network by a volume rendering method includes:

[0017] A dual-spectrum camera array is used to collect visible light images and near-infrared images of an aluminum alloy tube target, an adaptive weight function is established based on scene brightness values to fuse the visible light images and the near-infrared images, and a multi-view fusion image is obtained;

[0018] An intrinsic matrix and an extrinsic matrix of the dual-spectrum camera array are obtained, a camera optical center position is calculated according to the intrinsic matrix and the extrinsic matrix, a view direction vector is constructed by combining the camera optical center position and an observation point position, and a correspondence between the multi-view fusion image and the view direction vector is established;

[0019] Three-dimensional space coordinates of the aluminum alloy tube target and the view direction vector are respectively subjected to multi-layer sinusoidal position encoding, and the encoded space coordinate features and view features are input into a neural radiance field network;

[0020] The neural radiance field network is trained based on the multi-view fusion image and the corresponding view direction vector; the neural radiance field network outputs a body density parameter and a direction-dependent radiance intensity parameter, numerical integration is performed on the body density parameter and the radiance intensity parameter along a light ray direction, cumulative transmittance is calculated, and a continuous radiance field expression is obtained;

[0021] Based on the continuous radiance field expression, a light ray cumulative radiation value is calculated, the light ray cumulative radiation value is projected and mapped according to a view direction, and a temperature radiance intensity mapping image is obtained.

[0022] In an alternative embodiment:

[0023] The step of training the neural radiance field network based on the multi-view fusion image and the corresponding view direction vector comprises:

[0024] Feature extraction and fusion are performed on the multi-view fusion image and the view direction vector to obtain view perception features;

[0025] An adaptive sampling density function is constructed in a three-dimensional space, the adaptive sampling density function determines a sampling point density according to an included angle between a surface normal vector of a sampling point and a view direction vector, and a sampling point sequence is obtained by performing layered sampling along a ray direction based on the sampling point density;

[0026] The view perception features are input into a feature encoder to construct a multi-layer feature pyramid structure, multi-scale feature representations are obtained through feature transformation and up-sampling fusion, view-related features in the multi-scale feature representations are extracted, and a feature-guided attention map is established;

[0027] The sampling point sequence and the attention map are input into the neural radiance field network, and sampling point features are adaptively weighted based on the attention map;

[0028] A composite loss function based on material characteristics is constructed, which includes a reconstruction loss term and a physical constraint loss term, the physical constraint loss term includes a thermal radiation continuity constraint and a surface oxide layer reflection characteristic constraint, the thermal radiation continuity constraint imposes a physical consistency constraint on the predicted radiation intensity distribution based on Planck's blackbody radiation law, and the surface oxide layer reflection characteristic constraint constrains the reflection behavior at different viewing angles based on the optical characteristics of the high-temperature oxidation film of the aluminum alloy.

[0029] In an optional implementation,

[0030] The enhanced aluminum alloy tube target image is input into a multi-scale texture decoupling network, local texture descriptors are obtained through multi-scale decomposition and texture feature extraction, the local texture descriptors are input into a component separation unit, and decoupling and separation of the reflection component, the oxidation component and the forging component are performed, and the step of outputting the profile feature data of the aluminum alloy tube target includes:

[0031] The enhanced aluminum alloy tube target image is generated into multiple scale levels of images through downsampling operation; local texture features are extracted for the multiple scale levels of images respectively, nonlinear feature transformation values are calculated in a local neighborhood of a preset size, and the nonlinear feature transformation values are weighted and combined with local weights to obtain local texture descriptors of each scale level;

[0032] The local texture descriptors are input into a component separation unit, the component separation unit realizes decoupling and separation of the reflection component, the oxidation component and the forging component through causal reasoning and hierarchical decoupling, and obtains decoupled features of each scale level;

[0033] The feature discriminability is calculated for the decoupled features of each scale level, the corresponding fusion weight coefficient is determined according to the feature discriminability, the decoupled features are weighted and combined by using the fusion weight coefficient, and the fusion features are obtained;

[0034] The aluminum alloy tube target image is reconstructed based on the fusion features, the difference value between the reconstructed image and the original image and the spatial gradient value of the decoupled features are calculated, the difference value and the spatial gradient value are used as constraint conditions to optimize the parameters of the component separation unit, and the optimized aluminum alloy tube target profile feature data is output.

[0035] In an optional implementation,

[0036] The component separation unit realizes decoupling and separation of the reflection component, the oxidation component and the forging component through causal reasoning and hierarchical decoupling, and the step of obtaining decoupled features of each scale level includes:

[0037] The component separation unit comprises a causal reasoning module and a hierarchical decoupling module; in the causal reasoning module, a causal graph structure among the reflectance component, the oxidation component and the forging component is constructed, conditional probability distribution among the components is calculated based on the causal graph structure, and mutual dependence relationship of component features is established; in the hierarchical decoupling module, a decoupling loss function comprising a reliability evaluation item is constructed, the decoupling confidence is calculated according to the feature distribution of the local area, and adaptive decoupling strength is adopted for different confidence areas;

[0038] Based on the mutual dependence relationship and the decoupling confidence, a progressive training strategy is adopted for feature decoupling, a decoupling benchmark is established to gradually expand the decoupling process, and reliable separation of the reflectance component, the oxidation component and the forging component is realized.

[0039] The feature reconstruction loss is calculated according to the decoupling benchmark, the parameters of the causal reasoning module are optimized based on the feature reconstruction loss, and the decoupled features of each scale level are obtained.

[0040] In an optional embodiment,

[0041] Based on the topological relationship graph of the grid unit and the contour feature data, the step of predicting the deformation trend by graph structuring and generating the forging parameter adjustment instruction comprises:

[0042] The contour feature data is mapped to the topological relationship graph of the grid unit, and a graph feature matrix is constructed, the graph feature matrix comprising geometric features and material state parameters of each grid node;

[0043] The graph neural network is used for feature propagation of the graph feature matrix, a message passing mechanism is established based on the connection relationship between the grid nodes, the stress distribution and the deformation prediction value of each node are calculated by message aggregation, a deformation state vector is constructed according to the stress distribution and the deformation prediction value, and the deformation state vector comprises node displacement field, stress field and material flow characteristics;

[0044] The deformation state vector is compared with a preset target shape, a shape deviation is calculated, and a subsequent deformation trend is predicted based on the shape deviation and the material deformation law;

[0045] According to the predicted deformation trend, a forging parameter mapping function is established, the forging parameter mapping function converts the prediction deviation into the adjustment amount of the forging force, the forging speed and the temperature control parameter, and generates the forging parameter adjustment instruction.

[0046] In a second aspect, an aluminum alloy pipe target forging forming system based on image processing is provided, comprising:

[0047] The first unit is configured to establish a three-dimensional digital model of the aluminum alloy pipe target, divide the three-dimensional digital model into grid cells, and establish a topological relationship diagram of the grid cells;

[0048] The second unit is configured to acquire forging forming images of the aluminum alloy pipe target from multiple perspectives by using a dual-spectrum camera array, perform adaptive exposure compensation to obtain a multi-perspective fusion image, extract camera pose parameters in the multi-perspective fusion image, acquire corresponding perspective direction information, respectively perform sinusoidal position encoding on three-dimensional space coordinates of the grid cells and the perspective direction information, input a neural radiance field network to perform volume rendering reconstruction, and acquire a temperature radiation intensity mapping image.

[0049] The third unit is configured to perform phase consistency registration and wavelet transform denoising on the temperature radiation intensity mapping image to obtain an enhanced aluminum alloy pipe target image, input the enhanced aluminum alloy pipe target image into a multi-scale texture decoupling network, acquire local texture descriptors by multi-scale decomposition and texture feature extraction, input the local texture descriptors into a component separation unit, perform decoupling and separation of reflection component, oxidation component and forging component, and output profile feature data of the aluminum alloy pipe target.

[0050] The fourth unit is configured to predict a deformation trend by graph structuring calculation based on the topological relationship diagram of the grid cells and the profile feature data, generate a forging parameter adjustment instruction, and control a forging equipment according to the forging parameter adjustment instruction.

[0051] In a third aspect, a computer readable storage medium is provided, and computer program instructions are stored on the computer readable storage medium, and the computer program instructions are executed by a processor to implement the method described above.

[0052] The present application combines the grid three-dimensional digital model of the aluminum alloy pipe target with the dual-spectrum image fusion technology, realizes the all-around monitoring and accurate modeling of the forging process, adopts the neural radiance field network to perform volume rendering reconstruction, can accurately capture the temperature radiation characteristics of the aluminum alloy in the high-temperature state, and provides a high-quality data basis for subsequent deformation analysis.

[0053] The present application is based on a deformation trend prediction method of graph structured computing, establishes a mapping relationship between material properties and forming parameters, and realizes intelligent optimization and real-time adjustment of forging parameters. The present application significantly improves the forging forming quality and consistency of aluminum alloy pipe targets, reduces the scrap rate, reduces manual intervention, and shortens the production cycle. At the same time, the method has strong adaptability and scalability, and can meet the forming needs of aluminum alloy pipe targets of different specifications and shapes, providing a new technical path and solution for the field of aluminum alloy precision forging. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A flowchart of a forging forming method of an aluminum alloy pipe target based on image processing according to an embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application. The following several specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in detail in some embodiments.

[0056] Figure 1 A flowchart of a forging forming method of an aluminum alloy pipe target based on image processing according to an embodiment of the present application is shown in Figure 1 The method comprises the following steps:

[0057] A three-dimensional digital model of the aluminum alloy pipe target is established, the three-dimensional digital model is divided into grid cells, and a topological relationship graph of the grid cells is established;

[0058] A dual-spectrum camera array is used to collect forging forming images of the aluminum alloy pipe target from multiple viewing angles, and adaptive exposure compensation is performed to obtain a multi-view fusion image. Camera pose parameters in the multi-view fusion image are extracted to obtain corresponding viewing direction information. The three-dimensional space coordinates of the grid cells and the viewing direction information are respectively subjected to sinusoidal position encoding, input into a neural radiance field network for volume rendering reconstruction, and a temperature radiation intensity mapping image is obtained;

[0059] Phase consistency registration and wavelet transform denoising are performed on the temperature radiation intensity mapping image to obtain an enhanced aluminum alloy pipe target image. A phase consistency registration method is used to align multiple images, phase information of the images is extracted through Fourier transform, displacement relationships between the images are determined by maximizing phase correlation, and sub-pixel level accurate registration is realized. Wavelet transform denoising technology is applied to decompose the image into sub-bands of different frequencies, soft threshold or hard threshold shrinkage methods are used to suppress noise components of each sub-band coefficient, and the enhanced aluminum alloy pipe target image is reconstructed through inverse wavelet transform to retain edge details while effectively suppressing noise interference generated in the thermal imaging process;

[0060] The enhanced aluminum alloy tube target image is input into a multi-scale texture decoupling network, local texture descriptors are obtained through multi-scale decomposition and texture feature extraction, the local texture descriptors are input into a component separation unit, decoupling and separation of reflection component, oxidation component and forging component are performed, and profile feature data of the aluminum alloy tube target are output;

[0061] Based on the topological relationship graph of the grid unit and the profile feature data, a deformation trend is predicted through graph structuring calculation, a forging parameter adjustment instruction is generated, the forging equipment is controlled according to the forging parameter adjustment instruction, and optimization adjustment is performed based on real-time feedback data until the forming quality of the aluminum alloy tube target reaches a preset standard.

[0062] In an alternative embodiment:

[0063] A three-dimensional digital model of the aluminum alloy tube target is established, the three-dimensional digital model is divided into grid units, and the step of establishing a topological relationship graph of the grid unit includes:

[0064] A three-dimensional NURBS surface model of the aluminum alloy tube target is established, the three-dimensional NURBS surface model includes a cylindrical surface and an end surface, and the cylindrical surface and the end surface are constructed through control point coordinates, weight factors and B-spline basis functions;

[0065] The three-dimensional NURBS surface model is meshed, an adaptive grid density control function is established according to a curvature function, the adaptive grid density control function calculates the grid unit size based on a minimum grid size, a maximum grid size, a direction weight coefficient and a grid density control parameter, and tetrahedral grid units are generated;

[0066] A topological relationship graph is constructed based on the tetrahedral grid units, the topological relationship graph includes a grid node set and a node connection edge set, the grid node set is taken as a vertex, and the node connection edge set is taken as an edge.

[0067] For example, a three-dimensional NURBS surface model of an aluminum alloy tube target is established. The aluminum alloy tube target is mainly composed of a cylindrical surface and an end surface. In constructing the cylindrical surface, the size parameters of the radius and length are selected as 15 mm and 120 mm, respectively. Twelve control points are arranged along the circumferential direction, and ten control points are arranged along the axial direction to form a control point grid. Each control point is assigned a three-dimensional coordinate value and a corresponding weight factor. For example, the weight factors of the first row of control points on the cylindrical surface are all set to 0.7, and the weight factors of the remaining rows of control points are set to 1.0 to ensure accurate expression of the surface shape. In constructing the end surface, eight control points are used to form a closed circular end surface. The control points are distributed on a circumference with a radius of 15 mm, and the weight factors are all set to 0.9. For the order selection of the NURBS basis function, a 3-order B-spline basis function is used in the circumferential direction to accurately express the circular cross-section, and a 2-order B-spline basis function is used in the axial direction to meet the shape change requirements of the tube target along the axial direction. The node vectors are uniformly distributed, with the node vector in the circumferential direction being [0, 0, 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1, 1, 1] and the node vector in the axial direction being [0, 0, 0.125, 0.25, 0.375, 0.5, 0.625, 0.75, 0.875, 1, 1]. Through the combination of control point coordinates, weight factors, basis function orders, and node vectors, the accurate construction of the three-dimensional NURBS surface model of the aluminum alloy tube target is realized.

[0068] During the forging process of the aluminum alloy tube target, different parts may experience different degrees of deformation, and an adaptive grid density control function needs to be established according to the geometric characteristics. First, the curvature values of each point on the surface of the three-dimensional NURBS surface model are calculated. For the cylindrical surface of the tube target, the principal curvature radius is taken as 15 mm, and the corresponding curvature value is 0.067. For the edge transition area of the end surface, the curvature value increases significantly, with a maximum of 0.2. Based on the curvature distribution, an adaptive grid density control function is established. The minimum grid size is set to 0.5 mm, and the maximum grid size is set to 3 mm. In areas with large curvature, such as the transition area between the end surface and the cylindrical surface, the grid size is close to the minimum value of 0.5 mm. In areas with small curvature, such as the middle area of the cylindrical surface, the grid size is close to the maximum value of 3 mm. At the same time, a directional weight coefficient is introduced to make the grid have different distribution densities in the axial and circumferential directions. The axial weight coefficient is set to 1.2, and the circumferential weight coefficient is set to 0.9 to adapt to the deformation characteristics of the tube target in different directions. The grid density control parameter is set to 0.8 to adjust the mapping relationship between the curvature and the grid size. The control function is inversely proportional to the target size of the grid element and the curvature, while considering the influence of the directional weight. For example, in an area with a curvature value of 0.1, considering the weight coefficient, the axial grid size is about 1.5 mm, and the circumferential grid size is about 2 mm.

[0069] In the mesh generation process, first generate a triangular mesh on the model surface, control the number of units between 10000 and 15000, then expand to generate tetrahedral mesh units to the interior, the total number of units is about 30000 to 40000. The generated mesh needs to meet the quality standard: the minimum dihedral angle is greater than 15 degrees, the maximum dihedral angle is less than 160 degrees, and the volume ratio is greater than 0.2, to ensure the stability of numerical calculation. Based on the tetrahedral mesh unit, a topological relationship graph is constructed. Each mesh node is taken as a vertex of the graph, and the connection edge between the nodes is taken as an edge of the graph. For a typical aluminum alloy pipe target model, the number of mesh nodes is about 8000 to 10000, and the number of connection edges is about 45000 to 55000. Each mesh node stores its spatial coordinate information (x, y, z), and each connection edge records the index of the two nodes connected.

[0070] In the topological relationship graph, each node is assigned an attribute label, including node type (surface node or internal node), part to which it belongs (cylindrical surface area or end surface area), and other information. For surface nodes, additional surface normal vector information is stored for subsequent deformation analysis. At the same time, each edge is assigned a weight value, which is proportional to the edge length, for weight calculation in subsequent mesh deformation analysis.

[0071] In addition, an adjacency matrix or adjacency table is constructed in the topological relationship graph to record the connection relationship between nodes. For a mesh with n nodes, the adjacency matrix is an n x n sparse matrix, where the non-zero elements represent the connection relationship between nodes. Considering the sparsity characteristics of the aluminum alloy pipe target mesh, the adjacency table representation method is generally used, and each node has an average connectivity of about 5 to 7.

[0072] This method fully considers the geometric characteristics and deformation characteristics of the aluminum alloy pipe target, accurately expresses the complex geometric shape through NURBS surface, ensures the balance between calculation accuracy and efficiency through adaptive mesh division, and captures the spatial relationship between mesh units through the topological relationship graph.

[0073] In an alternative embodiment:

[0074] A dual-spectrum camera array is used to capture forging forming images of the aluminum alloy pipe target from multiple viewing angles, and adaptive exposure compensation is performed on the forging forming images to obtain a multi-view fusion image. Camera pose parameters in the multi-view fusion image are extracted to obtain corresponding viewing direction information. The three-dimensional spatial coordinates of the mesh units and the viewing direction information are respectively subjected to sinusoidal position encoding, and the encoded features are input into a neural radiance field network. The steps of numerically integrating the neural radiance field network by volume rendering method to obtain a temperature radiation intensity mapping image include:

[0075] A plurality of observation positions are uniformly arranged in a spherical coordinate system, a dual-spectrum camera array is used to collect visible light images and near-infrared images of an aluminum alloy pipe target at the observation positions, an adaptive weight function is established based on scene brightness values, and the visible light images and the near-infrared images are fused according to the adaptive weight function to obtain a multi-view fusion image;

[0076] An inner parameter matrix and an outer parameter matrix of the dual-spectrum camera array are obtained by using a calibration board, a camera optical center position is calculated according to the inner parameter matrix and the outer parameter matrix, a view direction vector is constructed by the camera optical center position and an observation point position, and a corresponding relationship between the multi-view fusion image and the view direction vector is established;

[0077] Three-dimensional space coordinates of the aluminum alloy pipe target and the view direction vector are respectively subjected to multi-layer sinusoidal position encoding, the multi-layer sinusoidal position encoding maps input features to a high-dimensional space through different frequency sinusoidal and cosine functions, and encoded space coordinate features and view features are input into a neural radiance field network;

[0078] The neural radiance field network is trained based on the multi-view fusion image and the corresponding view direction vector; the neural radiance field network outputs a body density parameter and a direction-dependent radiance intensity parameter, numerical integration is performed on the body density parameter and the radiance intensity parameter along a light ray direction, cumulative transmittance is calculated, and a continuous radiance field expression is obtained;

[0079] A light ray cumulative radiation value is calculated based on the continuous radiance field expression, the light ray cumulative radiation value is projected and mapped according to a view direction, and a temperature radiance intensity mapping image is obtained.

[0080] To ensure full coverage of the aluminum alloy pipe target, 12 observation points are arranged on a hemispherical surface with the pipe target as the center and a radius of 1.5 meters. These observation points are divided into three layers in the zenith angle direction, at 30 degrees, 60 degrees and 75 degrees respectively, and are uniformly distributed in the azimuth angle direction, with four points in each layer and an azimuth angle interval of 90 degrees between adjacent points. This arrangement ensures sufficient observation of the pipe target surface while avoiding observation dead angles. A dual-spectrum camera is installed at each observation position, which consists of two sensors: a visible light sensor (400-700 nm) and a near-infrared sensor (700-1100 nm). The two sensors are separated by a dichroic prism to ensure simultaneous acquisition of the same scene. For each observation position, visible light images and near-infrared images are simultaneously collected. During the forging process, the temperature of the aluminum alloy pipe target is usually in the range of 400-600℃, which mainly exhibits red-orange radiation in the visible light spectrum, while there is a strong thermal radiation signal in the near-infrared spectrum. Since the visible light image is easily overexposed at high temperatures, and the near-infrared image is more sensitive to temperature changes, it is necessary to fuse the two images to obtain more comprehensive information.

[0081] Before image fusion, adaptive exposure compensation is first performed. For the visible light image, the proportion of overexposed regions is calculated through histogram analysis, and when the overexposure proportion exceeds 10%, the exposure time is adjusted to 70% of the original; for the near-infrared image, when the average gray value of the image is lower than 30% of the total gray range, the exposure time is increased to 120% of the original. In this way, the best image quality can be ensured under different temperature conditions.

[0082] An adaptive weight function is established based on the scene brightness value to perform image fusion. For each pixel position, the normalized brightness values of the visible light image and the near-infrared image are calculated, and when the visible light pixel value exceeds 240 (8-bit image, maximum value 255), the visible light weight of this position is set to 0.3 and the near-infrared weight is set to 0.7; when the visible light pixel value is lower than 50, the visible light weight of this position is set to 0.6 and the near-infrared weight is set to 0.4; for the intermediate brightness region, the weight changes linearly with the brightness. According to this weight distribution strategy, the two images are weighted and fused to obtain a multi-view fusion image.

[0083] A standard chessboard calibration plate is used for camera calibration. The calibration plate is in a 9x7 format, with each grid size being 25mmx25mm. During the calibration process, the calibration plate is placed at 12 different positions and angles to ensure coverage of the entire field of view of the camera. By identifying the corner points of the calibration plate, the intrinsic matrix and distortion coefficients of the camera are calculated. The intrinsic matrix includes the focal length and principal point coordinates, for example, in a typical intrinsic matrix, the focal length value is about 1200 pixels, and the principal point coordinates are located near the center of the image. The distortion coefficients include radial distortion and tangential distortion parameters, and the radial distortion parameters k1 are usually about -0.1 and k2 are about 0.02. The extrinsic matrix is calculated by setting the origin of the world coordinate system at the fixed reference point of the forging equipment and using the known camera installation position and orientation. For example, for a camera located at an altitude angle of 30 degrees and an azimuth angle of 0 degrees, the rotation part of its extrinsic matrix represents the rotation angle of the camera coordinate system relative to the world coordinate system, and the translation part represents the position coordinates of the camera optical center in the world coordinate system, which is about (1.3, 0, 0.75) meters. According to the intrinsic matrix and the extrinsic matrix, the optical center position and the principal axis direction of each camera are calculated. The optical center position is directly obtained from the translation part of the extrinsic matrix, and the principal axis direction is calculated from the rotation part of the extrinsic matrix. The optical center position is connected with the observation point on the aluminum alloy pipe target to construct a view direction vector. For a typical setting, the view direction vector covers various angles from vertical observation to nearly horizontal observation, ensuring comprehensive capture of the pipe target surface.

[0084] For three-dimensional spatial coordinates, 10 frequency levels are selected, starting from the base frequency 20, with an increment of 1.5 for each level, forming a frequency sequence: 20, 30, 45, 67.5... and so on. For each frequency, the sine and cosine values of the coordinate in three directions are calculated, resulting in a 60-dimensional encoding vector (10 frequencies x 3 directions x 2 functions). Similarly, the viewing angle direction vector is also encoded using 8 frequency levels, resulting in a 48-dimensional direction encoding vector (8 frequencies x 3 directions x 2 functions). This high-dimensional encoding method effectively solves the difficulty of neural network learning low-frequency functions. The encoded spatial coordinate features and viewing angle features are input into the neural radiance field network. The network uses a multi-layer perceptron structure, containing 8 hidden layers, each with 256 neurons, using ReLU activation function. The network is divided into two parts: the first 5 layers process the spatial coordinate encoding, outputting the body density value; the last 3 layers combine the spatial encoding and direction encoding, outputting the RGB color value and radiance intensity value. The training data set contains tens of thousands of ray samples, with 64 points sampled on each ray. The Adam optimizer is used for training, with an initial learning rate of 0.0005, and the learning rate is reduced to half every 50,000 steps. After training is complete, the neural radiance field network can predict the body density and radiance intensity parameters of any three-dimensional point. To generate temperature radiance intensity mapping images, the predicted body density and radiance intensity are numerically integrated along each ray direction. Specifically, for each ray, 128 points are uniformly sampled from 0.5 meters near the end to 2.0 meters far away, the body density and radiance intensity of each sampling point are calculated, and then the cumulative transmittance is calculated. The cumulative transmittance represents the proportion of light remaining that has not been absorbed as the light propagates to a certain point, decreasing from far to near. Finally, the cumulative radiation value of the light is calculated based on the continuous radiance field expression. Multiply the body density value by the radiance intensity value, then multiply by the cumulative transmittance, and integrate along the ray to get the final radiation value. These radiation values are projected and mapped according to the viewing angle direction to form the temperature radiance intensity mapping image. For each point on the tube target surface, multiple radiance intensity values can be obtained from different viewing angles, and the average value is taken as the final radiance intensity.

[0085] The method combines dual-spectrum imaging and neural radiance field technology, overcoming the measurement deviation caused by the reflection and emissivity change of high-temperature metal surfaces in traditional thermal imaging, significantly improving the accuracy and spatial resolution of temperature field reconstruction. The reconstructed temperature radiance intensity mapping image intuitively displays the temperature distribution on the tube target surface, providing a reliable basis for subsequent forging process parameter adjustment, effectively improving the forging quality and consistency of aluminum alloy tube targets.

[0086] In an alternative embodiment:

[0087] The step of training the neural radiance field network based on the multi-view fusion image and the corresponding viewing angle direction vector includes:

[0088] perform feature extraction and fusion on the multi-view fusion image and the view direction vector to obtain view perception features;

[0089] An adaptive sampling density function is constructed in the three-dimensional space, the adaptive sampling density function determines the sampling point density according to the included angle between the sampling point surface normal vector and the view direction vector, and layered sampling is performed along the ray direction based on the sampling point density to obtain a sampling point sequence;

[0090] The view perception features are input into a feature encoder to construct a multi-layer feature pyramid structure, and multi-scale feature representations are obtained through feature transformation and up-sampling fusion; view-related features in the multi-scale feature representations are extracted to establish a feature-guided attention map for modulating the feature aggregation process of the sampling points;

[0091] The sampling point sequence and the attention map are input into the neural radiance field network, the sampling point features are adaptively weighted based on the attention map, and the neural radiance field network outputs the volume density parameters and the radiance intensity parameters of each sampling point;

[0092] A composite loss function based on material characteristics is constructed, the composite loss function includes a reconstruction loss term and a physical constraint loss term, the physical constraint loss term includes a thermal radiation continuity constraint and a surface oxide layer reflection characteristic constraint, the thermal radiation continuity constraint imposes a physical consistency constraint on the predicted radiance intensity distribution based on the Planck blackbody radiation law, and the surface oxide layer reflection characteristic constraint constrains the reflection behavior under different viewing angles based on the optical characteristics of the aluminum alloy high-temperature oxide film; the parameters of the neural radiance field network are optimized and trained based on the composite loss function.

[0093] For example, for each multi-view fusion image, ResNet-34 is used as the backbone network for feature extraction, and the feature maps of conv1 to conv5 are retained, with feature dimensions of 64, 128, 256, 512 and 512 respectively. At the same time, the view direction vector is encoded, a 6-layer fully connected network is used, the number of neurons in each layer is 64, 128, 256, 256, 128 and 64, the activation function is LeakyReLU, and the tilt coefficient is 0.2. In the feature fusion stage, an attention mechanism is used to convert the view direction encoding features into channel attention weights. In specific implementation, the 64-dimensional view encoding features are mapped to a weight vector matching the number of image feature channels through a two-layer fully connected network, and then each channel of the image features is weighted. For example, for the 256-channel feature map output by conv3, a 256-dimensional weight vector is generated, each weight value ranges from 0.5 to 1.5, and is used to modulate the importance of each channel. In this way, the network can adaptively adjust the feature extraction strategy according to different viewing directions to form view perception features.

[0094] Next, an adaptive sampling density function is constructed in three-dimensional space. To improve sampling efficiency, areas with obvious surface features are focused on. The cosine value of the angle between the surface normal vector of the sampling point and the view direction vector is defined as the weight factor. When the angle is close to 90 degrees (the cosine value is close to 0), it indicates that the line-of-sight direction is nearly parallel to the surface, and these areas often contain more surface detail information, so the sampling density should be increased; when the angle is close to 0 degrees or 180 degrees (the absolute value of the cosine value is close to 1), it indicates that the line-of-sight direction is nearly perpendicular to the surface, and these areas have relatively smooth surface features, so the sampling density can be reduced.

[0095] Based on the above weight factor, the sampling density function is defined as the product of the base density and the weight modulation term. The base density is set to 80 sampling points per meter, and the weight modulation term ranges from 0.5 to 2.0. In actual implementation, when the absolute value of the weight factor is less than 0.2, the sampling density is increased to twice the base density; when the absolute value of the weight factor is greater than 0.8, the sampling density is reduced to 0.5 times the base density; the sampling density in the intermediate region varies linearly. In this way, in critical areas such as the edges of the tube target and the transition of the curved surface, the sampling point spacing can be as small as 0.625 centimeters, while in smooth areas such as the cylindrical surface of the main body of the tube target, the sampling point spacing can be as large as 2.5 centimeters. When layering sampling is performed along the ray direction, the proximal end is set to 0.5 meters and the distal end is set to 2.0 meters, and the total number of sampling points is controlled between 64 and 192, which is dynamically adjusted according to the adaptive sampling density function. For rays with a view direction nearly parallel to the surface of the tube target, the number of sampling points is close to 192; for rays with a view direction nearly perpendicular to the surface of the tube target, the number of sampling points is close to 64. Through this adaptive sampling strategy, the sampling accuracy of critical areas is ensured, and the computational load is controlled.

[0096] The view angle perception features are input into a feature encoder to construct a multi-layer feature pyramid structure. The feature pyramid contains 5 scale levels, and the resolution gradually up-samples from 1 / 32 of the original image to the original resolution. At each scale level, features are extracted through a 3x3 convolution layer, and the output channel number of the convolution layer is 512, 256, 128, 64, and 32, respectively. The up-sampling process uses a bilinear interpolation method with an up-sampling factor of 2. Feature fusion is performed between adjacent levels through a skip connection, which specifically involves up-sampling the features of the upper layer and concatenating them with the features of the current layer in the channel dimension, and then adjusting the channel number through a 1x1 convolution.

[0097] In the multi-scale feature representation, view-dependent features are extracted, and a feature-guided attention map is established. The attention map generation process includes two steps: first, the encoded features of the view direction vector are mapped to a weight vector with the same number of channels as the feature map through a fully connected layer; then, for each spatial position of the feature map, the dot product of the feature vector and the weight vector is calculated, and normalized to the range of 0 to 1 through the Sigmoid function to form a spatial attention map. In practical applications, for different regions of the aluminum alloy tube target, such as the middle of the cylindrical surface, the end surface, and the transition region, the attention value distribution has obvious differences. Typically, the attention values of the edge and transition regions are between 0.7 and 0.9, while the attention values of the smooth regions are between 0.3 and 0.5.

[0098] The sample point sequence and the attention map are input into the neural radiance field network, and the sample point features are adaptively weighted based on the attention map. The neural radiance field network uses an 8-layer fully connected network with 256 neurons per layer and a ReLU activation function. The input of the network is the spatial coordinates of the sample points, the view direction vector, and the attention value at the corresponding position; the output is the volume density parameter and the radiance intensity parameter. The volume density parameter is a scalar value representing the "existence probability" of a spatial point, with a value range of 0 to 1, and a value closer to 1 indicates that the point is more likely to be on the surface of the object; the radiance intensity parameter is a three-dimensional vector corresponding to the three channel values of the RGB color space, representing the radiance intensity of the point when observed from a specific view angle.

[0099] In the feature aggregation process, the sample point position is projected onto a two-dimensional feature map, and the feature vector at the corresponding position is extracted, then multiplied by the attention value for weighting. In this way, the network can adaptively adjust the contribution of features according to the importance of different regions. For regions with high attention values, the sample point feature weight is larger, and the influence on the final prediction result is greater; for regions with low attention values, the sample point feature weight is smaller, and the influence on the final prediction result is smaller.

[0100] This paragraph does not fully describe the composite loss function, and lacks specific calculation methods and implementation details. The following is the supplemented and improved content:

[0101] A composite loss function based on material characteristics is constructed, including a reconstruction loss term and a physical constraint loss term. The reconstruction loss term adopts mean square error to calculate the difference between the predicted RGB color value and the true image RGB color value. In the specific implementation, the difference square of each pixel position is calculated for the three channels of RGB respectively, and then the difference square of all pixel positions and three channels is summed and averaged. In the early stage of training, the weight of the reconstruction loss is set to 1.0, and gradually decreases to 0.7 according to the exponential decay strategy, with a decay rate of 10% per 50,000 steps, so that the network pays more attention to the physical constraints. The physical constraint loss term includes thermal radiation continuity constraint and surface oxide layer reflection characteristic constraint. The thermal radiation continuity constraint is based on Planck's blackbody radiation law, which requires the predicted radiation intensity distribution to meet the continuity characteristics with temperature changes. In the specific implementation, adjacent sampling point pairs are taken along the ray direction, the difference in radiation intensity between each pair of points is calculated, and then L1 norm (absolute value) is applied as the gradient metric. When the gradient absolute value exceeds the preset threshold 0.05, a quadratic penalty is applied to the exceeding part, and the penalty coefficient is set to 0.1. The surface oxide layer reflection characteristic constraint considers the optical properties of the aluminum alloy surface forming an oxide film at high temperature, and requires the reflection behavior at different viewing angles to comply with the physical law. The implementation is as follows: for the same three-dimensional space point, 4-6 different viewing angle observation results are randomly selected from the data set, and the Pearson correlation coefficient of the predicted radiation intensity between each pair of viewing angles is calculated. Under normal circumstances, the correlation coefficient should not be less than 0.7; when the correlation coefficient is less than this threshold, the penalty term is calculated as the square of the difference between the threshold and the actual correlation coefficient, multiplied by the penalty coefficient 0.2. The final composite loss function is the weighted sum of the reconstruction loss, the thermal radiation continuity constraint loss and the surface oxide layer reflection characteristic constraint loss, and the weight ratio is 0.7:0.1:0.2 in the later stage of training.

[0102] The parameters of the neural radiation field network are optimized and trained based on the composite loss function. The Adam optimizer is adopted, and the initial learning rate is set to 0.0001, which is reduced to 50% of the original every 50,000 steps. The batch size is set to 4096 rays, and the number of sampling points on each ray is dynamically adjusted according to the adaptive sampling strategy. The training process is divided into two stages: the first stage performs 200,000 iterations, mainly optimizing the reconstruction loss; the second stage performs 100,000 iterations, optimizing the reconstruction loss and the physical constraint loss. During the training process, a model checkpoint is saved every 10,000 steps for subsequent evaluation and selection of the optimal model.

[0103] The method combines the advanced feature extraction technology of computer vision and the physical constraints of material science, effectively solves the problem of high temperature metal surface reflection variability and strong view angle dependence. The trained network not only can reconstruct visually realistic temperature radiation field, but also can ensure that the reconstruction result meets the physical law, provides a reliable tool for accurate monitoring and control of aluminum alloy tube target forging process, and significantly improves the accuracy of forging quality evaluation and the efficiency of forging parameter optimization.

[0104] In an alternative embodiment,

[0105] The enhanced aluminum alloy tube target image is input into a multi-scale texture decoupling network, local texture descriptors are obtained through multi-scale decomposition and texture feature extraction, the local texture descriptors are input into a component separation unit, and decoupling and separation of the reflection component, the oxidation component and the forging component are performed, and the step of outputting the profile feature data of the aluminum alloy tube target includes:

[0106] The enhanced aluminum alloy tube target image is generated into multiple scale levels of images through downsampling operation, the downsampling operation performs scale reduction processing on the aluminum alloy tube target image layer by layer; local texture feature extraction is performed on the multiple scale levels of images respectively, nonlinear feature transformation values are calculated in a local neighborhood of a preset size, and the nonlinear feature transformation values and local weights are weighted and combined to obtain local texture descriptors of each scale level;

[0107] The local texture descriptors are input into a component separation unit, the component separation unit realizes decoupling and separation of the reflection component, the oxidation component and the forging component through causal reasoning and hierarchical decoupling, and obtains decoupled features of each scale level;

[0108] The feature discriminability of each scale level of decoupled features is calculated, the corresponding fusion weight coefficient is determined according to the feature discriminability, the decoupled features are weighted and combined by using the fusion weight coefficient, and the fusion features are obtained;

[0109] The aluminum alloy tube target image is reconstructed based on the fusion features, the difference value between the reconstructed image and the original image and the spatial gradient value of the decoupled features are calculated, the difference value and the spatial gradient value are used as constraint conditions to optimize the parameters of the component separation unit, and the optimized aluminum alloy tube target profile feature data is output.

[0110] Exemplarily, the enhanced aluminum alloy tube target image is generated into multiple scale levels of images through downsampling operation, adopts a Gaussian pyramid structure for multi-scale decomposition, and a total of 5 scale levels are arranged. Taking the original image resolution of 1024x768 pixels as an example, after downsampling processing, images of 512x384, 256x192, 128x96 and 64x48 pixels are generated in turn. A 5x5 Gaussian kernel is used for convolution in the downsampling process, and the standard deviation is set to 1.2, and then 2 times downsampling is performed. This pyramid structure can capture image features at different scales and adapt to different size texture features on the surface of the aluminum alloy tube target. Local texture feature extraction is performed on the images of multiple scale levels. At each scale level, a local descriptor method is used to extract texture features. Specifically, taking each pixel as the center, a nonlinear feature transformation value is calculated in a preset size local neighborhood. For the first level (original resolution), the local neighborhood size is set to 9x9 pixels; for the second level, it is set to 7x7 pixels; for the third level and below, it is set to 5x5 pixels. In each local neighborhood, the following features are calculated: local direction gradient histogram (8 direction intervals), local binary pattern (using a circular neighborhood with a radius of 2, a total of 8 sampling points) and local contrast feature (calculating the difference between the maximum and minimum values in the neighborhood). These features are processed through nonlinear transformation. For the direction gradient histogram, normalization processing is performed to make the cumulative value of each direction equal to 1; for the local binary pattern, the binary code is converted to a decimal value; for the local contrast feature, the sigmoid function is used for normalization to make the value range between 0 and 1. Then, these features are mapped to a 64-dimensional feature space through a fully connected layer to obtain the preliminary texture description.

[0111] Next, local weights are calculated and combined. The local weights are determined based on texture complexity, which is evaluated by local variance and edge strength. The local variance is obtained by calculating the standard deviation of the pixel values in the neighborhood; the edge strength is obtained by calculating the gradient amplitude with the Sobel operator. After normalization, the two indicators are combined with weights of 0.4 and 0.6, respectively, to obtain the texture complexity score. The complexity score is converted to weight by the softmax function, making the weight sum equal to 1, and then multiplied element by element with the aforementioned 64-dimensional texture description to obtain the weighted local texture descriptor.

[0112] The local texture descriptor input component separation unit separates the decoupling of the highlight component, the oxidation component, and the forging component. The component separation unit adopts a multi-branch encoder-decoder structure and includes three parallel branches corresponding to the highlight component, the oxidation component, and the forging component. Each branch includes 3 convolution layers, the convolution kernel size is 3*3, the channel number is 64, 128, and 64 respectively, and the activation function is LeakyReLU with a negative slope of 0.2. During the component separation process, the input features are first mapped to the latent space through a shared encoder, and then the three components are decoded through three dedicated decoders. To achieve causal reasoning and hierarchical decoupling, two key mechanisms are introduced: first, the information flow is controlled through an attention gating mechanism to maintain appropriate independence between components; second, prior knowledge is introduced to guide the decoupling process. The highlight component is mainly related to surface smoothness and incident light direction, so in the decoding process, the spatial attention mechanism is used to highlight the highlight area; the oxidation component is related to temperature distribution and oxidation time, and in the decoding process, the channel attention mechanism is used to enhance the features related to color change; the forging component is related to material deformation and internal structure, and in the decoding process, the multi-scale fusion mechanism is used to retain detailed texture information.

[0113] Through this separation mechanism, the component decoupling is performed on the features of each scale level, and 15 decoupled feature maps are obtained (5 scale levels * 3 components). The dimension of each decoupled feature map is the corresponding scale image size * 64 (feature dimension).

[0114] The feature discriminability is calculated for the decoupled features of each scale level to determine the fusion weight coefficients. The feature discriminability is evaluated by calculating two indicators: feature variance and feature entropy. The feature variance reflects the dispersion degree of the feature distribution, and the feature entropy reflects the uncertainty of the feature distribution. For each decoupled feature, the variance of its 64-dimensional feature vector is calculated to obtain the variance value; the feature histogram is calculated (the feature values are divided into 10 uniform intervals), and then the entropy value of the histogram is calculated. After normalization, the variance value and the entropy value are weighted and summed with a ratio of 0.5:0.5 to obtain the feature discriminability score.

[0115] Based on the feature discriminability, the fusion weight coefficients are calculated. The discriminability score is converted to a weight using a softmax function, and the sum of the weights of all scale levels is 1. In practical applications, the weight of a higher scale level (such as the original resolution and the second resolution) is usually between 0.25-0.35, and the weight of a lower scale level is between 0.1-0.2, reflecting the importance of high-resolution features for contour extraction.

[0116] The decoupled features are weighted and combined using a fusion weight coefficient. The feature maps of each scale level are upsampled to the original resolution, then weighted and summed according to the weight coefficient to obtain the fusion features of the three components. In practical implementation, the bilinear interpolation method is used for upsampling to ensure smooth transition of the feature maps.

[0117] The aluminum alloy tube target image is reconstructed based on the fusion features. The reconstruction process uses a convolutional neural network with 3 convolutional layers, a convolution kernel size of 3x3, channel numbers of 64, 32, and 3 (corresponding to RGB three channels), and a ReLU activation function. The last layer uses a Sigmoid function to normalize the output value to the range of 0-1. The difference between the reconstructed image and the original image is calculated by mean square error as the reconstruction loss.

[0118] At the same time, the spatial gradient value of the decoupled features is calculated as a regularization constraint. The spatial gradient is calculated by the Sobel operator to calculate the gradient of the feature map in the horizontal and vertical directions, and then the L1 norm of the gradient amplitude is calculated as the gradient loss. In the training process, the total loss function is the weighted sum of the reconstruction loss and the gradient loss, with a weight ratio of 1:0.2.

[0119] The parameters of the component separation unit are optimized based on the above loss function. The Adam optimizer is used with an initial learning rate of 0.001, which is reduced to 80% of the original every 50 periods. The batch size is set to 16 and the training is performed for 200 periods. The training data includes 500 enhanced aluminum alloy tube target images, of which 400 are used for training and 100 are used for verification. During the training process, the performance of the verification set is evaluated every 10 periods, and the model parameters with the best performance are saved.

[0120] After optimization, the profile feature data of the aluminum alloy tube target is output from the component separation unit. Specifically, the feature map of the forging component is extracted, threshold segmentation (threshold value set to 1.5 times the feature mean) is applied to extract the high response area, and then morphological operations (first erosion and then dilation, structure element is 3x3 rectangle) are applied to remove noise, obtaining the profile feature data of the aluminum alloy tube target. The profile feature data includes edge position coordinates and edge intensity values, which can be used for subsequent deformation analysis and quality evaluation.

[0121] This method effectively separates the reflection, oxidation, and forging features on the surface of the aluminum alloy tube target through multi-scale decomposition and texture feature decoupling, overcoming the limitations of traditional methods in dealing with complex surface characteristics. Especially in the high-temperature forging environment, the aluminum alloy surface has multiple phenomena such as light reflection, oxidation discoloration, and deformation texture. This method can accurately extract the surface features caused by forging and filter out interference factors such as reflection and oxidation, greatly improving the accuracy and stability of profile feature extraction, providing a reliable data foundation for subsequent quality evaluation and process parameter optimization.

[0122] In an alternative embodiment,

[0123] The component separation unit decouples the specular component, the oxidation component, and the forging component by causal reasoning and hierarchical decoupling, and the step of obtaining decoupled features of each scale level includes:

[0124] The component separation unit includes a causal reasoning module and a hierarchical decoupling module; in the causal reasoning module, a causal graph structure among the specular component, the oxidation component, and the forging component is constructed, conditional probability distribution among the components is calculated based on the causal graph structure, and mutual dependency relationship of component features is established; in the hierarchical decoupling module, a decoupling loss function including a reliability evaluation term is constructed, the decoupling confidence is calculated according to the feature distribution of the local region, and adaptive decoupling strength is adopted for different confidence regions;

[0125] Based on the mutual dependency relationship and the decoupling confidence, a progressive training strategy is adopted for feature decoupling, first a decoupling benchmark is established in a high confidence region, then the decoupling process is gradually expanded to a low confidence region, and reliable separation of the specular component, the oxidation component, and the forging component is realized;

[0126] The feature reconstruction loss is calculated according to the decoupling benchmark, the parameters of the causal reasoning module are optimized based on the feature reconstruction loss, and decoupled features of each scale level are obtained.

[0127] For example, the component separation unit first receives multi-scale texture descriptors as input. For the characteristics of aluminum alloy tube targets, the texture descriptor dimension of each scale level is 64, corresponding to 1024x768x64 at the original resolution, 512x384x64 at the second level, and so on. These descriptors contain rich information of surface texture, but the specular, oxidation, and forging features are mixed with each other and need to be further separated.

[0128] In the causal reasoning module, a causal graph structure is constructed among the reflectance component, the oxidation component, and the forging component. Based on physical prior knowledge, there is a clear causal relationship among the three components: the reflectance component is mainly determined by the surface geometry and the incident light direction; the oxidation component is affected by the temperature distribution and the oxidation time; and the forging component is directly related to the material flow and the degree of deformation. Therefore, in the constructed causal graph, the reflectance component → oxidation component → forging component forms the main causal link, while the direct influence of the reflectance component → forging component is also considered. The causal graph structure is represented by a directed graph network, with nodes corresponding to the three components and edges representing the causal relationship. Each node contains a feature extractor composed of a 3-layer convolutional network with a kernel size of 3x3 and channel numbers of 64, 128, and 64, respectively, followed by batch normalization and LeakyReLU activation (negative slope of 0.2) after each layer. The causal strength of the edge is realized through an attention gate mechanism, for example, the edge from the reflectance component to the oxidation component generates an attention map (with the same dimension as the feature map) to control the strength of information transmission.

[0129] The conditional probability distribution between the components is calculated based on the causal graph structure. For each component, its feature representation is conditionally dependent on its parent node component. For example, the oxidation component feature is conditionally dependent on the reflectance component feature; the forging component feature is conditionally dependent on the reflectance component and oxidation component features. The conditional probability distribution is estimated by a variational inference method, modeling the feature representation of each component as a Gaussian distribution, with the distribution parameters (mean and variance) output by the feature extractor. In actual implementation, for the oxidation component, its conditional distribution parameters are mapped from the reflectance component feature through a fully connected layer; for the forging component, its conditional distribution parameters are mapped from the concatenation of the reflectance component and oxidation component features through a fully connected layer. The mutual dependence relationship of component features is established through two ways: feature reconstruction and mutual information maximization. Feature reconstruction requires that the child node component feature can be predicted from the parent node component feature. For example, the oxidation component feature is reconstructed from the reflectance component feature, with a 2-layer fully connected network as the reconstruction network and a hidden layer dimension of 128. Mutual information maximization requires that the information is shared between related components and minimized between unrelated components. The specific implementation uses a contrastive learning method, with the positive sample pair being the feature pair of related components and the negative sample pair being the feature pair of unrelated components, optimized by the InfoNCE loss function with a temperature parameter of 0.07.

[0130] In the hierarchical decoupling module, a decoupling loss function containing reliability evaluation terms is constructed. The reliability evaluation is based on the feature distribution of local regions to calculate the decoupling confidence. For each position in the image, the consistency of features in its 9x9 neighborhood is calculated. The consistency is evaluated by the cosine of the angle between feature vectors, and the closer the cosine value is to 1, the more consistent the features are, and the higher the decoupling confidence is. In practical applications, the decoupling confidence of smooth regions such as the cylindrical surface of the tube target body is usually between 0.8-0.95; the decoupling confidence of edge and texture complex regions such as the forging wrinkles is usually between 0.4-0.7.

[0131] An adaptive decoupling strength is adopted for different confidence regions. The decoupling strength is controlled by a regularization coefficient. For high confidence regions (e.g. confidence > 0.8), a higher decoupling strength is adopted, and the regularization coefficient is set to 1.0; for medium confidence regions (e.g. confidence between 0.5-0.8), the regularization coefficient is set to 0.6; for low confidence regions (e.g. confidence < 0.5), the regularization coefficient is set to 0.3. This adaptive strategy can strengthen decoupling in regions with clear feature distribution and maintain appropriate flexibility in regions with ambiguous feature distribution.

[0132] Based on the interdependence relationship and the decoupling confidence, a progressive training strategy is adopted for feature decoupling. The progressive training is divided into three stages: the first stage (first 50 cycles) only trains in high confidence regions (confidence > 0.8) to establish a decoupling benchmark; the second stage (51-100 cycles) extends to medium confidence regions (confidence > 0.5); the third stage (101-200 cycles) extends to all regions. This strategy can first establish a reliable decoupling pattern in regions with clear feature distribution, and then migrate the decoupling knowledge to more complex regions. In each training stage, the network weight update method is also different. In the first stage, only the parameters related to high confidence regions are updated, and other parameters are frozen; in the second stage, the parameters related to high and medium confidence regions are updated; in the third stage, all parameters are updated. The learning rate is also adjusted in stages: 0.001 in the first stage, 0.0005 in the second stage, and 0.0001 in the third stage. The optimizer uses Adam, with momentum parameters beta1 set to 0.9 and beta2 set to 0.999.

[0133] After the reliable separation of the reflection component, the oxidation component, and the forging component, the feature maps of each component clearly show the surface characteristics of different physical causes. The reflection component feature map highlights the highlight areas, such as the smooth parts of the tube target surface; the oxidation component feature map highlights the color change areas, such as the oxidation film formed in the high temperature area; the forging component feature map highlights the texture and deformation areas, such as the streamlines and wrinkles formed in the forging process.

[0134] The feature reconstruction loss is calculated according to the decoupling reference. The feature reconstruction adopts an encoder-decoder structure, the encoder maps the mixed features to three latent spaces, and the decoder reconstructs three component features from the three latent spaces respectively, and then compares the reconstructed features with the decoupling reference features to calculate the mean square error as the reconstruction loss. For high confidence areas, the reconstruction loss weight is set to 1.0; for medium confidence areas, the weight is set to 0.7; for low confidence areas, the weight is set to 0.4.

[0135] In addition to the reconstruction loss, an adversarial loss and a mutual information loss are also introduced. The adversarial loss is realized through a discriminator network, and the discriminator structure is a 4-layer convolutional network to judge the true and false of the reconstructed features and the decoupling reference features, so as to make the generated feature distribution closer to the real distribution. The mutual information loss is realized by minimizing the mutual information between unrelated components, for example, the mutual information between the reflection component and the forging component should be minimized, while the mutual information between related components should be within a certain range.

[0136] Based on the above loss function combination, the parameters of the causal inference module are optimized. The total loss function is the weighted sum of the reconstruction loss, the adversarial loss and the mutual information loss, and the weight ratio is 1:0.5:0.3. Training adopts small batch gradient descent with a batch size of 16, and each cycle contains 25 batches. During the training process, the model performance is evaluated on the validation set every 10 cycles, and when the performance does not improve for 5 consecutive evaluations, the training is terminated in advance.

[0137] The optimized causal inference module can accurately separate the reflection component, the oxidation component and the forging component from the multi-scale texture descriptors, and obtain decoupled features at each scale level. For the original resolution image of the aluminum alloy tube target, the decoupled feature dimension is 1024x768x64x3 (widthxheightxchannel numberxcomponent number); for the second level, the feature dimension is 512x384x64x3, and so on. These decoupled features retain the spatial structure and semantic information of the original features, while realizing the separation of physical causes, providing a basis for subsequent contour feature extraction and quality evaluation.

[0138] The method effectively solves the problem of mixed features on the surface of the aluminum alloy tube target through causal inference and hierarchical decoupling, and realizes accurate separation of features with different physical causes. Compared with traditional methods, the present scheme fully considers the causal relationship and confidence distribution between features, adopts a progressive training strategy to expand the decoupling process from high confidence areas to low confidence areas, and greatly improves the accuracy and robustness of decoupling. Especially for complex scenes where surface reflection, oxidation and deformation exist simultaneously in the forging process, pure forging components can be extracted.

[0139] In an alternative embodiment:

[0140] The step of predicting a deformation trend by graph structure calculation and generating a forging parameter adjustment instruction based on the topological relationship graph of the grid unit and the profile feature data comprises:

[0141] Mapping the profile feature data of the aluminum alloy pipe target to the topological relationship graph of the grid unit, and constructing a graph feature matrix, wherein the graph feature matrix comprises geometric features and material state parameters of each grid node;

[0142] Performing feature propagation on the graph feature matrix by using a graph neural network, establishing a message passing mechanism based on the connection relationship between the grid nodes, and calculating stress distribution and deformation prediction values of each node by message aggregation; constructing a deformation state vector according to the stress distribution and the deformation prediction values, wherein the deformation state vector comprises a node displacement field, a stress field and a material flow feature, and is used to represent the overall deformation state of the aluminum alloy pipe target;

[0143] Comparing the deformation state vector with a preset target shape, calculating a shape deviation, and predicting a subsequent deformation trend based on the shape deviation and a material deformation law;

[0144] According to the predicted deformation trend, a forging parameter mapping function is established, which converts the prediction deviation into an adjustment amount of the forging force, the forging speed and the temperature control parameter, and generates a forging parameter adjustment instruction.

[0145] Exemplarily, mapping the profile feature data of the aluminum alloy pipe target to the topological relationship graph of the grid unit and constructing a graph feature matrix is a continuous process. First, the aluminum alloy pipe target is discretized by using a hexahedral grid, the standard pipe target model is divided into 12 units in the radial direction, 24 units in the circumferential direction and 36 units in the axial direction, and a total of 10368 grid units are formed. These grid units form the basic framework of the topological relationship graph, each grid unit corresponds to a node in the graph, and the shared surface between adjacent units corresponds to an edge in the graph. The grid topological relationship is represented by an adjacency matrix, the matrix dimension is 10368x10368, and the element value 1 represents adjacency and 0 represents non-adjacency.

[0146] The profile feature data provides key observation information of the surface nodes. By using a feature matching algorithm, these profile features are accurately positioned to the corresponding nodes of the topological graph. The matching process is based on the principle of spatial coordinate nearest neighbor. For each profile feature point, the Euclidean distance between the feature point and the grid surface node is calculated, and the feature is assigned to the node with the smallest distance. When a grid node corresponds to multiple feature points, a weighted average method is used to fuse the feature values, and the weight is inversely proportional to the distance.

[0147] For each surface node, the mapped features include three-dimensional coordinate values (X, Y, Z in millimeter), normal vector three components (normalized to [-1, 1] range), principal curvature values (in 1 / millimeter) and forging component intensity values (normalized to [0, 1] range). For example, the feature data of a surface node in the middle of the tube target can be: coordinate (125.36, 0.00, 87.42) millimeter, normal vector (1.00, 0.00, 0.00), principal curvature 0.008 / millimeter, and forging component intensity 0.65. To ensure the accuracy of the mapping, a bilinear interpolation method is used to handle the case where the contour feature points and the grid nodes do not correspond completely, and the interpolation radius is set to 1.5 times the size of the grid element.

[0148] After the surface feature mapping is completed, the internal node features cannot be directly observed and need to be filled in through feature propagation from the surface to the interior. The propagation adopts a physics-based interpolation method to establish a feature transmission channel from the surface node to the internal node. Specifically, using the basic principles of elastoplastic mechanics, based on the surface deformation data and boundary conditions, the initial state estimation of the internal node is calculated using Laplace interpolation. The form of Laplace interpolation is that for each internal node, its feature value is equal to the weighted average of the feature values of all adjacent nodes, and the weight is inversely proportional to the distance. This ensures smooth transition of the physical field, high calculation efficiency and results consistent with physical meaning. For different types of physical fields, special propagation strategies are adopted: the steady-state solution form of the heat conduction equation is used for interpolation for the temperature field, and the temperature distribution satisfies the Laplace equation; the interpolation method constrained by the elastic equilibrium equation is used for the stress field to ensure that the internal stress field satisfies the balance condition; the deformation field is constrained by the displacement continuity condition to ensure the continuity and smoothness of the deformation field.

[0149] Through contour feature mapping and internal feature propagation, complete feature information is assigned to each node in the topology relationship graph, thereby constructing a graph feature matrix. The matrix includes the geometric features and material state parameters of each grid node. The geometric features include node coordinates (3D), displacement values (3D) and local deformation gradients (9D); the material state parameters include equivalent stress (1D), equivalent strain (1D), principal stress components (3D), temperature value (1D), material flow velocity (3D) and flow direction (3D), hardening parameter (3D) and damage index (2D). The finally constructed graph feature matrix has a dimension of node number x feature dimension, about 11200 x 32 (including surface and internal nodes) for a standard model, which completely records the geometric shape and material state distribution of the aluminum alloy tube target in the current forging state.

[0150] The graph neural network is used for feature propagation of the graph feature matrix. The message passing neural network (MPNN) structure is adopted, including four key steps of feature transformation, message generation, message aggregation, and state update. The feature transformation is realized by a multi-layer perceptron, which maps the original node features to a 128-dimensional hidden feature space. The multi-layer perceptron includes two fully connected layers, and the activation function is ReLU. The specific layer structure is 32-64-128 (input dimension-hidden dimension-output dimension).

[0151] The message generation process considers the interaction between the node's own features and the adjacent node features. For each pair of connected nodes, a feature difference vector is calculated, and a message vector is generated by combining the edge attributes (such as node distance, connection direction, etc.). The message generation function adopts a three-layer fully connected network, and the layer structure is 256-128-128 (connected input dimension-hidden dimension-output dimension). For example, for surface adjacent nodes, the generated message contains information such as relative displacement, stress gradient, and temperature gradient, which describes the local deformation state.

[0152] The message aggregation is realized by a weighted summation mechanism. The weight is determined based on the distance between nodes and the material flow direction. The closer the distance and the higher the consistency of the flow direction, the greater the weight of the node pair. The typical weight calculation adopts an attention mechanism, and the attention score is obtained after point multiplication and softmax normalization. The temperature parameter is set to 0.1. The dimension of the aggregated message vector is kept at 128, which contains information from all adjacent nodes.

[0153] The state update combines the current state of the node and the aggregated message, and generates a new node state through a gated update mechanism. The gated update is similar to a GRU unit, which includes a reset gate and an update gate to control the proportion of information retention and update. The gating parameters are calculated based on node features, and the typical values are: for high stress area nodes, the update proportion is about 0.7-0.9; for low stress area nodes, the update proportion is about 0.3-0.5. Through repeated execution of message passing and state update (usually iterating 3-5 times), the graph neural network can capture the global deformation pattern.

[0154] After feature propagation, the stress distribution and deformation prediction value of each node are calculated based on the connection relationship between the grid nodes. The stress distribution is obtained by decoding the node features, including three principal stress components and equivalent stress values. The deformation prediction value includes the displacement increment of the node at the future time step, which is predicted by a decoding network. The decoding network is a three-layer fully connected network with a layer structure of 128-64-16 (feature dimension-hidden dimension-output dimension), and the activation function is Tanh. The output of the 16-dimensional vector contains information such as principal stress components, equivalent stress, displacement increment, and flow velocity.

[0155] A deformation state vector is constructed according to the stress distribution and deformation prediction. The deformation state vector is a compact representation of the overall deformation state of the aluminum alloy tube target, and the key components are extracted from the node features by the principal component analysis (PCA) method. In specific implementation, first, the feature data of all nodes are standardized, and then PCA is applied for dimension reduction, retaining the principal components required to explain 95% of the variance (usually 20-30 components). The deformation state vector contains three parts of information: node displacement field, stress field, and material flow characteristics. The displacement field describes the displacement direction and amplitude of each node; the stress field describes the stress state of each node; and the material flow characteristics describe the flow velocity and direction of the material. For example, part of the deformation state vector value under a typical forging state is: the maximum displacement is 0.85 mm (occurring at the end of the tube target), the maximum equivalent stress is 278 MPa (occurring at the center of the deformation zone), and the maximum material flow velocity is 3.2 mm / s (in the radial direction).

[0156] The deformation state vector is compared with the preset target shape to calculate the shape deviation. The target shape is defined by a three-dimensional CAD model and contains the precise geometric dimensions of the target tube target. The shape deviation calculation uses the vertex-to-surface minimum distance method, which calculates the distance from each surface node to the target surface. Typical shape deviation indicators include: maximum deviation (usually required to be controlled within 0.5 mm), average deviation (usually required to be controlled within 0.2 mm), and deviation standard deviation (usually required to be controlled within 0.15 mm).

[0157] Based on the shape deviation and material deformation law, the subsequent deformation trend is predicted. The deformation trend prediction uses a combination of extrapolation and physical constraints. First, based on the current deformation rate, the position at the future time point is predicted by linear extrapolation; then, the extrapolation results are corrected by applying the volume conservation and material constitutive relationship. The correction process considers the effects of material strengthening and temperature softening, and the prediction results are made to conform to the physical law through an iterative solution process. For quantitative evaluation of the deformation trend, two methods are used: deviation change rate and critical region identification. The deviation change rate is calculated by the deviation change of two consecutive time steps, with a positive value indicating an increase in deviation and a negative value indicating a decrease in deviation. The critical region refers to the region where the shape deviation or stress concentration occurs, which is identified by setting a threshold value (such as deviation exceeding 2 times the average value or stress exceeding 90% of the material yield strength). According to the predicted deformation trend, a forging parameter mapping function is established. The forging parameter mapping uses a combination of rule-based methods and lookup table methods to map the shape deviation and deformation trend to specific forging parameter adjustments. The rule base is constructed based on forging expert experience and experimental data, and contains processing strategies for different shape deviation types. For example:

[0158] When the radial expansion of the tube target end is too large (deviation > 0.3 mm), reduce the radial forging force by 5-10 kN, and lower the temperature in this region by 10-15°C;

[0159] When the radial contraction of the middle part of the tube target is too large (deviation > 0.25 mm), increase the axial forging speed by 1-2 mm / s, and increase the temperature of this area by 5-10°C at the same time;

[0160] When the surface of the tube target appears uneven deformation (deviation standard deviation > 0.2 mm), adjust the forging force distribution, and increase the forging force in the area of insufficient deformation by 3-8 kN;

[0161] When it is predicted that local stress concentration may occur (stress > 300 MPa), extend the current forging stage time by 1-2 seconds, and reduce the forging speed by 2-3 mm / s.

[0162] The table lookup method realizes fast parameter query through a pre-calculated parameter mapping table. The mapping table has dimensions of deviation type x deviation degree x current forging stage, and the table cells contain corresponding parameter adjustment suggestions. For example, for the "end radial expansion" type, with a deviation of 0.35 mm and in the second forging stage, the adjustment suggestion obtained by looking up the table is: reduce the radial forging force by 8 kN, reduce the end temperature by 12°C, and reduce the forging speed by 1.5 mm / s.

[0163] The generated forging parameter adjustment instructions include the upper die forging force adjustment amount (range ± 50 kN), the lower die forging force adjustment amount (range ± 50 kN), the radial forging speed adjustment amount (range ± 5 mm / s), the axial forging speed adjustment amount (range ± 3 mm / s), the temperature control parameter adjustment amount of the four regions (each region range ± 15°C), and the duration adjustment amount of the three forging stages (each stage range ± 2s). The parameter adjustment instructions are transmitted to the forging equipment control system in a structured data format, and the equipment implements the instructions step by step after receiving the instructions, and verifies the effect after each adjustment through sensor feedback.

[0164] The method combines material forming basic theory and engineering practice experience, realizes accurate prediction of deformation trend and real-time optimization of parameters in the forging process of aluminum alloy tube target. For aluminum alloy tube targets with complex shapes, by establishing a structured representation of material deformation and a rule-based parameter mapping, potential deformation defects can be predicted and process parameters can be adjusted in advance.

[0165] In a second aspect, an aluminum alloy tube target forging system based on image processing is provided, comprising:

[0166] A first unit for establishing a three-dimensional digital model of an aluminum alloy tube target, dividing the three-dimensional digital model into grid cells, and establishing a topological relationship diagram of the grid cells;

[0167] The second unit is configured to acquire forging forming images of the aluminum alloy pipe target from multiple perspectives by using a dual-spectrum camera array, perform adaptive exposure compensation to obtain a multi-perspective fusion image, extract camera pose parameters in the multi-perspective fusion image, and obtain corresponding perspective direction information; and perform sinusoidal position coding on three-dimensional space coordinates of the grid unit and the perspective direction information respectively, input a neural radiance field network to perform volume rendering reconstruction, and obtain a temperature radiation intensity mapping image.

[0168] The third unit is configured to perform phase consistency registration and wavelet transform denoising on the temperature radiation intensity mapping image to obtain an enhanced aluminum alloy pipe target image; input the enhanced aluminum alloy pipe target image into a multi-scale texture decoupling network, obtain local texture descriptors by multi-scale decomposition and texture feature extraction, input the local texture descriptors into a component separation unit, perform decoupling separation of reflection component, oxidation component and forging component, and output profile feature data of the aluminum alloy pipe target.

[0169] The fourth unit is configured to predict a deformation trend by graph structuring calculation based on a topological relationship graph of the grid unit and the profile feature data, generate a forging parameter adjustment instruction, and control a forging equipment according to the forging parameter adjustment instruction.

[0170] In a third aspect, a computer readable storage medium is provided, and computer program instructions are stored on the computer readable storage medium. When the computer program instructions are executed by a processor, the foregoing method is implemented.

Claims

1. A forging method for an aluminum alloy tube target based on image processing, characterized in that, include: A three-dimensional digital model of the aluminum alloy tube target is established, the three-dimensional digital model is divided into mesh cells, and a topological relationship diagram of the mesh cells is established. A dual-spectral camera array is used to acquire forging images of an aluminum alloy tube target from multiple perspectives, and adaptive exposure compensation is performed to obtain a multi-view fused image. The camera pose parameters in the multi-view fused image are extracted to obtain the corresponding view direction information. The three-dimensional spatial coordinates of the mesh unit and the view direction information are sinusoidally encoded and input into a neural radiation field network for volume rendering and reconstruction to obtain a temperature radiation intensity mapping image. Phase-consistent registration and wavelet transform denoising are performed on the temperature radiation intensity mapping image to obtain an enhanced aluminum alloy tube target image. The enhanced aluminum alloy tube target image is then input into a multi-scale texture decoupling network. Local texture descriptors are obtained through multi-scale decomposition and texture feature extraction. These local texture descriptors are then input into a component separation unit for decoupling and separating the reflective, oxidation, and forging components, outputting the contour feature data of the aluminum alloy tube target. Specifically, this includes: generating multiple scale-level images from the enhanced aluminum alloy tube target image through downsampling operations; extracting local texture features from each of the multiple scale-level images; calculating nonlinear feature transformation values ​​within a preset local neighborhood; weighting and combining the nonlinear feature transformation values ​​with local weights to obtain a local texture descriptor for each scale level; and inputting the local texture descriptors into a component separation unit, which performs causal inference... Hierarchical decoupling is used to separate the reflective component, oxidation component, and forging component, resulting in decoupling features at each scale level. The reflective component refers to the optical feature component related to the surface smoothness of the aluminum alloy tube target and the incident light direction; the oxidation component refers to the color change feature component related to temperature distribution and oxidation time; and the forging component refers to the texture feature component related to material deformation and internal structure. For each scale level of decoupling features, a feature discriminant is calculated. Based on the feature discriminant, a corresponding fusion weight coefficient is determined. The decoupling features are then weighted and combined using the fusion weight coefficient to obtain fused features. Based on the fused features, an image of the aluminum alloy tube target is reconstructed. The difference between the reconstructed image and the original image, as well as the spatial gradient value of the decoupling features, are calculated. The difference and spatial gradient values ​​are used as constraints to optimize the parameters of the component separation unit, outputting optimized aluminum alloy tube target contour feature data. Based on the topological relationship diagram of the grid cells and the contour feature data, the deformation trend is predicted through graph structured calculation, and forging parameter adjustment instructions are generated; the forging equipment is controlled according to the forging parameter adjustment instructions.

2. The method according to claim 1, characterized in that, The steps of establishing a three-dimensional digital model of the aluminum alloy tube target, dividing the three-dimensional digital model into mesh cells, and establishing a topological relationship diagram of the mesh cells include: A three-dimensional NURBS surface model of an aluminum alloy tube target is established. The three-dimensional NURBS surface model includes a cylindrical surface and an end face. The cylindrical surface and the end face are constructed by control point coordinates, weighting factors and B-spline basis functions. The three-dimensional NURBS surface model is meshed, and an adaptive mesh density control function is established based on the curvature function. The adaptive mesh density control function calculates the mesh cell size based on the minimum mesh size, maximum mesh size, direction weight coefficient and mesh density control parameters, and generates tetrahedral mesh cells. A topological graph is constructed based on the tetrahedral mesh unit. The topological graph includes a set of mesh nodes and a set of node connecting edges, with the set of mesh nodes as vertices and the set of node connecting edges as edges.

3. The method according to claim 1, characterized in that, A dual-spectral camera array is used to acquire forging images of an aluminum alloy tube target from multiple perspectives. Adaptive exposure compensation is applied to the forging images to obtain multi-view fused images. Camera pose parameters are extracted from the multi-view fused images to obtain the corresponding view direction information. The three-dimensional spatial coordinates of the grid cells and the view direction information are sinusoidally encoded, and the encoded features are input into a neural radiation field network. The steps of obtaining a temperature radiation intensity mapping image by numerically integrating the neural radiation field network using a volume rendering method include: Visible light and near-infrared images of an aluminum alloy tube target are acquired using a dual-spectrum camera array. An adaptive weighting function is established based on the scene brightness value to fuse the visible light and near-infrared images, resulting in a multi-view fused image. Obtain the intrinsic and extrinsic parameter matrices of the dual-spectrum camera array, calculate the camera optical center position based on the intrinsic and extrinsic parameter matrices, construct a view direction vector by combining the camera optical center position with the observation point position, and establish the correspondence between the multi-view fused image and the view direction vector. The three-dimensional spatial coordinates of the aluminum alloy tube target and the viewpoint direction vector are respectively subjected to multi-layer sinusoidal position encoding, and the encoded spatial coordinate features and viewpoint features are input into the neural radiation field network. The neural radiation field network is trained based on the multi-view fused image and the corresponding view direction vector; the neural radiation field network outputs volume density parameters and direction-related radiation intensity parameters, and the volume density parameters and radiation intensity parameters are numerically integrated along the light direction to calculate the cumulative transmittance and obtain the continuous radiation field expression. The cumulative radiation value of light is calculated based on the continuous radiation field expression, and the cumulative radiation value of light is projected and mapped according to the viewing direction to obtain a temperature radiation intensity mapping image.

4. The method according to claim 3, characterized in that, The steps for training the neural radiation field network based on the multi-view fused image and the corresponding view direction vector include: Feature extraction and fusion are performed on the multi-view fused image and the view direction vector to obtain view perception features; An adaptive sampling density function is constructed in three-dimensional space. The adaptive sampling density function determines the sampling point density based on the angle between the surface normal vector of the sampling point and the view direction vector. Based on the sampling point density, layered sampling is performed along the ray direction to obtain a sampling point sequence. The view-aware features are input into a feature encoder to construct a multi-layer feature pyramid structure. Multi-scale feature representations are obtained through feature transformation and upsampling fusion. View-related features are extracted from the multi-scale feature representations to establish a feature-guided attention map. The sampling point sequence and the attention map are input into the neural radiation field network, and the sampling point features are adaptively weighted based on the attention map. A composite loss function based on material properties is constructed, comprising a reconstruction loss term and a physical constraint loss term. The physical constraint loss term includes thermal radiation continuity constraints and surface oxide layer reflectivity constraints. The thermal radiation continuity constraints apply physical consistency constraints to the predicted radiation intensity distribution based on Planck's blackbody radiation law, while the surface oxide layer reflectivity constraints constrain reflection behavior under different viewing angles based on the optical properties of the high-temperature oxide film of the aluminum alloy. The parameters of the neural radiation field network are then optimized and trained based on the composite loss function.

5. The method according to claim 1, characterized in that, The component separation unit achieves decoupling and separation of the reflective component, oxidation component, and forging component through causal reasoning and hierarchical decoupling. The steps for obtaining the decoupling features at each scale level include: The component separation unit includes a causal reasoning module and a hierarchical decoupling module. In the causal reasoning module, a causal graph structure is constructed between the reflective component, the oxidation component, and the forging component. Based on the causal graph structure, the conditional probability distribution between the components is calculated, and the interdependence of component features is established. In the hierarchical decoupling module, a decoupling loss function containing a reliability evaluation term is constructed. The reliability evaluation term calculates the decoupling confidence based on the feature distribution of the local region, and an adaptive decoupling strength is adopted for regions with different confidence levels. Based on the aforementioned interdependencies and the aforementioned decoupling confidence, a progressive training strategy is adopted to decouple features, establish a decoupling benchmark, and gradually extend the decoupling process to achieve reliable separation of reflective components, oxidation components, and forging components. The feature reconstruction loss is calculated based on the decoupling benchmark, and the parameters of the causal inference module are optimized based on the feature reconstruction loss to obtain the decoupling features at each scale level.

6. The method according to claim 1, characterized in that, Based on the topological relationship diagram of the mesh cells and the contour feature data, the steps of predicting deformation trends and generating forging parameter adjustment instructions through graph-structured calculations include: The contour feature data is mapped to the topological relationship graph of the grid cells to construct a graph feature matrix; A graph neural network is used to propagate features from the graph feature matrix. A message passing mechanism is established based on the connection relationship between grid nodes. The stress distribution and deformation prediction value of each node are calculated through message aggregation. A deformation state vector is constructed based on the stress distribution and deformation prediction value. The deformation state vector includes the node displacement field, stress field and material flow characteristics. The deformation state vector is compared with the preset target shape, the shape deviation is calculated, and the subsequent deformation trend is predicted based on the shape deviation and the material deformation law. Based on the predicted deformation trend, a forging parameter mapping function is established. The forging parameter mapping function converts the predicted deviation into the adjustment amount of forging force, forging speed and temperature control parameters, and generates forging parameter adjustment instructions.

7. A forging system for an aluminum alloy tube target based on image processing, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to establish a three-dimensional digital model of the aluminum alloy tube target, divide the three-dimensional digital model into grid units, and establish a topological relationship diagram of the grid units; The second unit is used to acquire forging images of the aluminum alloy tube target from multiple perspectives using a dual-spectrum camera array, perform adaptive exposure compensation to obtain a multi-view fused image; extract the camera pose parameters from the multi-view fused image to obtain the corresponding view direction information; perform sinusoidal position encoding on the three-dimensional spatial coordinates of the mesh unit and the view direction information respectively, input them into the neural radiation field network for volume rendering reconstruction, and obtain a temperature radiation intensity mapping image. The third unit is used to perform phase consistency registration and wavelet transform denoising on the temperature radiation intensity mapping image to obtain an enhanced aluminum alloy tube target image; the enhanced aluminum alloy tube target image is input into a multi-scale texture decoupling network, and local texture descriptors are obtained through multi-scale decomposition and texture feature extraction; the local texture descriptors are input into a component separation unit to decouple and separate the reflective component, oxidation component and forging component, and output the contour feature data of the aluminum alloy tube target; The fourth unit is used to predict deformation trends through graph-structured calculation based on the topological relationship diagram of the grid cells and the contour feature data, generate forging parameter adjustment instructions, and control the forging equipment according to the forging parameter adjustment instructions.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images