Artificial intelligence driven automated system for cold extrusion camera housing design

Abstract

This invention discloses an automated system and method for cold extrusion camera housing design driven by artificial intelligence, belonging to the field of intelligent manufacturing technology. The system includes a preprocessing module, a parameter generation module, a verification and prediction module, and an optimization module. The preprocessing module converts a 3D product model into signed distance field data; the parameter generation module, based on a conditional variational autoencoder model, samples from the latent space to generate SDF data for the cold extruded part and process parameters for the billet; the verification and prediction module performs geometric constraint verification and forming quality prediction; and the optimization module achieves closed-loop optimization of the design by adjusting the latent space sampling strategy. Through a self-learning closed loop of product SDF, latent space encoding, cold extrusion part generation, verification and prediction, and optimization feedback, end-to-end intelligent design from product geometry to cold extruded parts and billets is achieved, significantly reducing reliance on manual labor and simulation costs, and improving design efficiency and quality.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing technology, and more specifically, to an automated system for the design of cold-extruded camera housings based on artificial intelligence. Background Technology

[0002] As a key structural component in advanced driver assistance systems (ADAS) and autonomous driving systems, vehicle camera housings are typically made of aluminum alloy and formed through cold extrusion. These housings require high geometric precision and lightweight construction while maintaining material continuity and surface integrity during forming. Effective cold extrusion design ensures uniform material flow, minimizes defects such as overlaps and folds, optimizes material utilization, reduces cold extrusion load, and minimizes flash volume. Therefore, cold extrusion design and process optimization are typical high-complexity problems, making them highly suitable for incorporating artificial intelligence methods for pattern learning and automatic optimization. The complexity and dynamics of metal forming, along with its nonlinear multivariate interactions, make traditional optimization methods ineffective. Existing cold extrusion design and process parameter determination mainly rely on engineer experience and multiple rounds of finite element simulation verification, a time-consuming process that struggles to guarantee global optimality. While traditional finite element methods can accurately simulate metal flow and stress distribution, in practical design, even for the same model of vehicle camera housing, multiple rounds of geometric and process parameter adjustments are often required. Every minor modification (such as changes in wall thickness, rib position offsets, or mold angle adjustments) requires remodeling and recalculation of the simulation, resulting in high computational costs and long cycles, making rapid iteration and global optimization difficult during the design phase. When product models or local structures change further, the repetitive workload of the existing simulation process increases exponentially, severely restricting molding design efficiency. Manual judgment of molding quality is highly subjective and cannot systematically quantify the impact of geometric changes on molding behavior, leading to long design cycles, low efficiency, and unstable yield rates. With the rise of Industry 4.0, AI-driven smart manufacturing optimizes production plans; for example, machine learning algorithms can predict potential failures based on performance indicators such as temperature, vibration, and pressure.

[0003] Utilizing deep learning and machine learning methods to model 3D geometry and simulation data has become a common approach in engineering design. Artificial intelligence models can learn from finite element simulation results, establishing a statistical mapping relationship between geometric features and forming performance, enabling rapid prediction of new design schemes, thereby shortening design cycles and reducing costs. However, existing artificial intelligence methods often lack explicit physical constraints, making it difficult to completely replace finite element simulation under complex and variable process conditions. Their application effectiveness depends on the consistency and structural similarity of the samples.

[0004] Vehicle-mounted camera housings share high similarities in material systems, forming processes, and geometric features, providing a stable data distribution and pattern space for artificial intelligence (AI) models. Leveraging the structural commonalities of these products, models can be effectively trained under limited sample conditions, enabling reliable generalization and prediction accuracy. Introducing AI models into cold extrusion design and forming quality prediction can significantly reduce the computational burden of repetitive simulations, achieving data-driven rapid design and intelligent forming performance evaluation. However, a systematic AI-based method for generating and judging the quality of forgings for vehicle-mounted camera housings is currently lacking.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] The purpose of this application is to provide an automated system for cold extrusion camera housing design driven by artificial intelligence, which solves the problems of existing cold extrusion design for automotive camera housing products, such as reliance on manual experience, low efficiency of shape matching, large amount of finite element simulation calculation, and lag in forming performance prediction.

[0007] In the first aspect, this application provides an automated system for cold extrusion camera housing design based on artificial intelligence, the technical solution of which is as follows:

[0008] The preprocessing module receives the 3D product model of the vehicle-mounted camera housing, performs geometric repair, coordinate alignment and scale normalization, and converts the processed model into a signed distance field with a fixed resolution, outputting the product SDF data.

[0009] The parameter generation module is used to encode the product SDF data into the latent space based on a pre-trained conditional variational autoencoder model; and to reconstruct the corresponding cold extrusion part SDF data based on the latent vector sampled from the latent space by the decoder corresponding to the conditional variational autoencoder model and the product SDF data as conditions, and simultaneously output the associated set of billet process parameters through a parametric regression branch.

[0010] The verification and prediction module is used to analyze the SDF data of the cold extruded part to determine whether its volume, minimum wall thickness, and draft angle meet the preset process constraint thresholds, generate verification results, and receive the set of process parameters of the blank and output the forming quality prediction score.

[0011] Furthermore, the system also includes an optimization module; the optimization module is communicatively connected to the parameter generation module and the verification and prediction module, respectively.

[0012] The optimization module is configured to: receive the verification result and the forming quality prediction score; when the verification result does not meet the preset standard, generate a feedback signal based on the non-compliance item, and convert the feedback signal into an adjustment instruction for the latent space sampling strategy based on the conditional variational autoencoder model, so as to optimize and generate a new set of billet process parameters that simultaneously meet geometric constraints and high forming quality probability.

[0013] Furthermore, the verification and prediction module includes a graph neural network model;

[0014] The process of receiving the set of process parameters for the billet and outputting a forming quality prediction score includes:

[0015] The three-dimensional voxel mesh corresponding to the SDF data of the cold-extruded part is transformed into a graph structure; wherein each voxel is a node of the graph, and the node features include the SDF value, coordinate position and local gradient of the voxel; the set of billet process parameters is fused with the features of each node as global graph features;

[0016] After message passing and node feature updating through a multi-layer graph convolutional network in the graph neural network model, global information is aggregated through a graph pooling layer, and finally the forming quality prediction score is obtained through regression by a fully connected layer.

[0017] Furthermore, the decoder corresponding to the conditional variational autoencoder model adopts a multi-task learning architecture, which includes an SDF reconstruction branch and a parametric regression branch. The SDF reconstruction branch is used to predict signed distance values ​​on a voxel-by-voxel basis to form the SDF of the cold extruded part. The parametric regression branch shares hidden layer features with the SDF reconstruction branch in the intermediate layer of the decoding process, and uses an independent regression head network to map the shared features to the billet process parameter set. The billet process parameter set includes at least the billet diameter, length, volume compensation coefficient, and material identification code.

[0018] Furthermore, the optimization module is configured to perform the following specific optimization process:

[0019] If the verification result indicates a volume mismatch, the ratio of the current cold-extruded part volume to the target volume is calculated, and vector interpolation or perturbation sampling is performed in the potential space of the parameter generation module along the direction that makes the ratio approach 1 to generate a new set of billet process parameters.

[0020] If the forming quality prediction score is lower than a preset threshold, the gradient of the forming quality prediction score with respect to the potential vector is calculated using the differentiable property of the prediction model in the verification and prediction module. Then, an iterative search is performed in the potential space along the gradient ascent direction to find potential vectors that can generate higher prediction scores, thereby generating an optimized set of billet process parameters.

[0021] Furthermore, the scale normalization process includes: calculating the maximum outer envelope size of the three-dimensional product model in the three coordinate axes of the three-dimensional space, and using the length of the longest side in the maximum outer envelope size as the normalization reference, scaling the coordinates of all vertices of the three-dimensional product model proportionally so that the length of the longest side of the scaled model is equal to the preset reference length value.

[0022] Furthermore, the system also includes a human-computer interaction and decision visualization module, which is communicatively connected to the parameter generation module, the verification and prediction module, and the optimization module, and is configured to execute:

[0023] The latent space based on the conditional variational autoencoder model in the parameter generation module is dimensionality-reduced and mapped. In the visualization interface, the three-dimensional models of cold extruded parts generated by different latent points, the corresponding geometric constraint verification results output by the verification and prediction module, and the forming quality prediction score are displayed in association.

[0024] The system receives user selection and annotation operations for specific regions or directions in the potential space from the visualization interface, generates corresponding interactive instructions, and provides these interactive instructions as initial sampling preferences or constraints to the optimization module.

[0025] Secondly, this application provides an automated design method for cold-extruded camera housings based on artificial intelligence, comprising the following steps:

[0026] The system receives a 3D product model of the vehicle-mounted camera housing, performs geometric repair, coordinate alignment, and scale normalization on it, and converts the processed model into a signed distance field with a fixed resolution to obtain the product SDF data.

[0027] The product SDF data is input into a pre-trained conditional variational autoencoder model, which encodes the product SDF data into a latent space. Based on the decoder corresponding to the encoder, latent vectors are sampled from the latent space, and the product SDF data is used as a generation condition input into the decoder to reconstruct the corresponding cold extrusion part SDF data. Simultaneously, through the parameter regression branch in the decoder, a set of billet process parameters associated with the cold extrusion part SDF data is output synchronously.

[0028] The SDF data of the cold-extruded part is analyzed to calculate its volume, minimum wall thickness, and draft angle characteristics, and compared with the preset process constraint thresholds to generate manufacturability verification results. At the same time, the set of process parameters of the blank is input into the performance prediction model to obtain the forming quality prediction score.

[0029] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the method described in the second aspect.

[0030] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described in the second aspect.

[0031] As shown above, the preprocessing module converts the 3D product model into signed distance field data; the parameter generation module, based on a conditional variational autoencoder model, samples from the latent space to generate SDF data for cold extrusion parts and process parameters for billets; the verification and prediction module performs geometric constraint verification and forming quality prediction; and the optimization module achieves closed-loop optimization of the design by adjusting the latent space sampling strategy. Through a self-learning closed loop of product SDF, latent space encoding, cold extrusion part generation, verification and prediction, and optimization feedback, end-to-end intelligent design from product geometry to cold extrusion parts and billets is achieved, significantly reducing reliance on manual labor and simulation costs, and improving design efficiency and quality. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of an automated system structure for cold extrusion camera housing design based on artificial intelligence, as disclosed in an embodiment of the present invention.

[0034] Figure 2 This is a flowchart illustrating the automated design method steps for cold extrusion camera housing design based on artificial intelligence, as disclosed in an embodiment of the present invention. Detailed Implementation

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments belong; the terminology used herein and in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit these embodiments; the terms in the specification of these embodiments and the foregoing description of the accompanying drawings include and have, and any variations thereof, and are intended to cover non-exclusive inclusion. The terms first, second, etc., in the specification of these embodiments and the foregoing drawings are used to distinguish different objects and not to describe a particular order.

[0036] The implementation details of the technical solution in this embodiment are described in detail below:

[0037] Firstly, this application provides an automated system for cold extrusion camera housing design based on artificial intelligence, such as... Figure 1 As shown. The system includes the following:

[0038] The preprocessing module 101 is used to receive the three-dimensional product model of the vehicle-mounted camera housing, perform geometric repair, coordinate alignment and scale normalization processing, and convert the processed model into a signed distance field with fixed resolution, and output it as product SDF data.

[0039] Specifically, in this embodiment, the original 3D product model of the vehicle-mounted camera housing is received and subjected to geometric repair, coordinate alignment, and scale normalization. Geometric repair aims to detect and correct potential topological and geometric defects in the original model, such as non-manifold edges, self-intersecting surfaces, holes, isolated vertices, or faces, ensuring the model is a closed and watertight entity, providing an unambiguous geometric basis for subsequent voxelization and distance field calculations. Coordinate alignment refers to transforming the repaired model from its original, possibly arbitrary, local coordinate system to a standardized world coordinate system by calculating its principal axis of inertia or based on a preset assembly / processing reference plane (e.g., aligning the model's key assembly planes with the coordinate plane, or placing the model's centroid at the origin), ensuring consistency and comparability of all input models in spatial orientation. Scale normalization involves scaling the model without changing its shape proportions, so that its maximum spatial bounding dimension (such as the diagonal length of the bounding box) matches the preset normalized size unit (such as a unit cube). This eliminates the scale effect caused by differences in the original design size of different models, creating a uniform scale condition for subsequent fixed-resolution processing.

[0040] The process of voxelizing the processed model and converting it into a fixed-resolution 3D signed distance field, outputting normalized product SDF data, involves defining a 3D regular mesh with a fixed resolution (e.g., 256x256x256) within the normalized model space after the aforementioned processing. The voxelization process involves calculating the shortest distance from the center point of each voxel to the model surface on this mesh and assigning a sign—typically, points inside the model are designated as negative distances, external points as positive distances, and points exactly on the model surface are designated as zero distances. This generates a discrete scalar field representing the model's shape information, i.e., a 3D signed distance field. This conversion process not only performs regular digital sampling of the model at a fixed resolution, but more importantly, the SDF data precisely and continuously encodes the model's spatial structure, internal and external regions, and surface position information in the form of distances and signs. The output product SDF data is this normalized and structured 3D data representation.

[0041] In practical applications, this solution ensures the geometric integrity of the input data through systematic geometric repair, a prerequisite for the correct execution of all subsequent processing. Coordinate alignment and scale normalization eliminate the randomness of the model in spatial pose and physical dimensions, enabling camera housing models from different design sources or models to be compared, analyzed, or subjected to subsequent intelligent processing (such as defect detection and intelligent manufacturing planning) within a unified and unbiased reference framework. The final conversion to fixed-resolution SDF data not only transforms the continuous geometric model into a discrete data form suitable for computation, storage, and deep learning model processing, but more importantly, SDF, as a fundamental and powerful geometric representation, inherently possesses continuity, symbolic meaning, and distance information. This provides a direct, efficient, and geometrically semantically rich data foundation for subsequent tasks such as shape analysis, collision detection, simulation, or deep learning-based feature extraction and pattern recognition.

[0042] Specifically, in some of the above embodiments, the functions of the preprocessing module 101 can be further refined. The preprocessing module 101 performs geometric repair, coordinate alignment, and scale normalization, and converts the processed model into a signed distance field with a fixed resolution, including:

[0043] The geometric repair steps are as follows: First, traverse all edges and faces of the model, identify and stitch non-manifold edges, and eliminate self-intersecting faces; then, detect and fill all boundary loops to repair holes on the model surface; finally, remove all redundant isolated vertices and faces to ensure that the output model is a single, closed, manifold triangular mesh entity. The coordinate alignment steps are as follows: Calculate the model's inertia tensor and solve for its eigenvectors. Align the eigenvectors corresponding to the largest and second-largest eigenvalues ​​to the positive X and Y axes of the world coordinate system, respectively, to achieve principal axis alignment based on the model's geometric distribution; or, identify three predefined mutually perpendicular reference planes on the model (such as mounting surfaces, symmetry planes, etc.) and align them to the XY, XZ, and YZ planes of the world coordinate system using least squares fitting. The scale normalization steps are as follows: Calculate the axial bounding box of the model in the space after coordinate alignment and obtain its maximum side length L_max; then, uniformly scale the coordinates of all vertices of the model using a scaling factor S = preset normalized size / L_max. The specific steps for converting to SDF are as follows: Within the normalized model space, construct a unit cube space with a side length of 1, and divide it evenly into N parts along the XYZ directions (N is a fixed resolution, such as 256), forming an N³ voxel grid; for the center of each voxel, use a spatial acceleration structure (such as an AABB tree) to calculate its signed distance to the nearest triangle; finally, store the calculated distance value matrix as the product SDF data file.

[0044] Specifically, this series of preprocessing operations aims to transform raw, potentially messy CAD models into high-quality, standardized digital geometric assets. Geometric repair is crucial for data cleaning, resolving low-level errors caused by design software export, format conversion, or modeling oversights, ensuring the mathematical rigor of the model. Coordinate alignment, through principal axis alignment or datum plane alignment, assigns a definite spatial orientation to the model, which is essential for subsequent applications that rely on spatial relationships (such as assembly analysis and perspective rendering). Scale normalization places all models under a uniform metric scale, making shell models of different sizes comparable in numerical characteristics. Conversion to fixed-resolution SDF is a core step bridging continuous geometry and discrete computation. Fixed resolution ensures uniform data size, facilitating batch processing and neural network input. The signed distance field itself not only records surface positions (zero isosurfaces) but also contains rich local and global shape information (gradients, curvature, etc., can be approximated through distance field calculations), making it an intermediate representation extremely suitable for geometric reasoning and analysis. The entire preprocessing process essentially translates diverse raw inputs into a standardized, digital, and information-rich common language (SDF), laying a solid and consistent data foundation for subsequent intelligent analysis and processing.

[0045] The parameter generation module 102 is used to encode the product SDF data into the latent space based on a pre-trained conditional variational autoencoder model; and to reconstruct the corresponding cold extrusion part SDF data based on the latent vector sampled from the latent space by the decoder corresponding to the conditional variational autoencoder model and the product SDF data as conditions, and simultaneously output the associated billet process parameter set through a parametric regression branch.

[0046] Specifically, in this embodiment, the encoder that inputs the product's SDF data into the conditional variational autoencoder model refers to inputting a fixed-resolution 3D SDF data tensor representing the final geometry of the camera housing into an encoder composed of a deep convolutional neural network. This encoder, through multiple layers of convolution and downsampling operations, progressively extracts multi-scale geometric features (such as overall contours, local grooves, hole locations, wall thickness distribution, and other high-level abstract features) contained in the SDF data. Ultimately, it maps the high-dimensional voxel data to a relatively low-dimensional, continuous latent space and outputs the mean vector and variance vector (i.e., probability distribution parameters) corresponding to this latent vector.

[0047] The process of sampling a latent vector based on probability distribution parameters and using this latent vector, along with the product's SDF data, as conditional input to the decoder involves defining a multivariate Gaussian distribution using the mean and variance vectors output by the encoder. A specific latent vector z is sampled from this distribution using reparameterization techniques. This latent vector z is a continuous and somewhat random implicit representation of the product's geometry after compression, capturing key variability factors of the shape. During the decoding stage, not only is this sampled latent vector z input to the decoder, but the original product SDF data (or its main features) is also input as additional conditional information. This conditional input method ensures that the decoding process is explicitly guided and constrained by the given target product geometry, ensuring that the decoder's generation target is not an arbitrary cold-extruded part shape, but rather an intermediate blank shape specifically designed for manufacturing that particular product.

[0048] In practical applications, the decoder reconstructs the SDF data of the cold-extruded part and simultaneously outputs a set of billet process parameters through a parametric regression branch. This means that the decoder is a deep deconvolutional neural network that receives the sampled latent vector z and the product's SDF conditional information. Through multiple layers of upsampling and convolution operations, it progressively decompresses and synthesizes the 3D SDF data of the cold-extruded part, matching the geometry of the input product. This cold-extruded part SDF data represents the ideal geometry of the initial metal billet (after the cold extrusion process) required to form the camera housing. Simultaneously, a parametric regression branch (typically one or more fully connected layers) sharing weights with some layers of the decoder network is activated. This branch receives the same latent vector z and / or intermediate layer features from the decoding process, mapping them to a specific set of billet process parameters. These parameters typically include, but are not limited to, the billet's initial diameter, initial height, blank weight, extrusion ratio, or machining allowances for specific areas to meet subsequent machining requirements. The key is that, since the parameter regression branch shares some encoding features with the decoder that generates the geometry of the cold extruded part and is jointly trained, the process parameters it outputs are inherently consistent and highly coordinated with the geometry of the cold extruded part generated by the decoder in terms of both physical and process logic. Together, they constitute a complete and feasible reverse process design scheme from the final product to the manufacturing medium.

[0049] This application's solution constructs an intelligent mapping bridge from product geometry to blank geometry and process parameters by introducing a pre-trained conditional variational autoencoder model. The encoding process abstracts product geometry into a probability distribution in the latent space, which achieves data compression and feature extraction, and introduces reasonable design variability through probabilistic sampling, providing flexibility for exploring feasible process solutions. Decoding is performed conditionally on the product's SDF, ensuring that the generation process always revolves around the target product and avoiding the generation of irrelevant shapes. More importantly, through a parametric regression branch that shares underlying features and executes synchronously, this solution achieves tight coupling and integrated output between geometric design and process parameter formulation. This design makes the generated cold-extruded part SDF not just a geometric shape, but also an executable process instruction blank with key manufacturing parameters, greatly improving the automation and intelligence level from design to manufacturing preparation.

[0050] Specifically, in some of the above embodiments, the function of the parameter generation module 102 can be further refined. The parameter generation module 102 processes data based on a conditional variational autoencoder model, including the following specific steps:

[0051] Encoding step: The product SDF data (with dimensions [D, H, W]) is input into the encoder network E. After passing through L convolutional layers (including downsampling), a high-level feature map is obtained. Then, after flattening and fully connected layers, the mean vector μ and the log variance vector log(σ²) are output, that is, μ, log(σ²) = E(SDF_product).

[0052] Sampling step: Calculate the latent vector z = μ + ε ⊙ exp(0.5 * log(σ²)), where ε is a random vector sampled from the standard normal distribution N(0, I), and ⊙ represents element-wise multiplication. This step realizes differentiable sampling from the posterior distribution q(z|SDF_product) defined by (μ, σ²).

[0053] Conditional decoding and reconstruction step: Concatenate or perform weighted fusion on the sampled vector z and the product SDF conditional information c that has been appropriately processed (such as extracting features through an independent lightweight conditional encoder or directly downsampling) to form the conditional latent code [z; c]. Input this conditional latent code into the decoder network G. After passing through L transposed convolutional layers (including upsampling), the SDF data of the cold extruded part is reconstructed, that is, SDF_blank = G([z; c]).

[0054] Parameter regression synchronous output step: The parameter regression branch R shares the first K layers (K < L) of feature extraction layers with the decoder G. During the decoding process, when calculating to the Kth layer, the feature map output by this layer is fed to the subsequent layers of the decoder and the parameter regression branch R simultaneously. The branch R usually consists of a global average pooling layer and several fully connected layers, which map the feature map to the blank process parameter vector P, that is, P = R(Feature_K). The process parameter set P can specifically include: the initial diameter D0 of the blank, the initial height H0, the volume V (or mass M), and the unilateral machining allowance δ at key parts (such as the lens mounting boss).

[0055] The verification and prediction module 103 is used to analyze the SDF data of the cold extruded part to determine whether its volume, minimum wall thickness, and draft angle meet the preset process constraint thresholds, generate a verification result; and receive the set of blank process parameters and output a forming quality prediction score.

[0056] Specifically, in this embodiment, based on the SDF data of the cold-extruded part (a three-dimensional scalar field, where each voxel stores the signed distance to the model surface), its geometric and physical properties are extracted and analyzed using numerical methods. The volume is obtained by summing the physical space volume occupied by all voxel units located inside the model (SDF value ≤ 0). The volume of each voxel unit is the cube of its physical dimension (determined by the model's normalized scale and SDF resolution). The calculated volume V_blank must satisfy: V_min ≤ V_blank ≤ V_max, where V_min ensures sufficient material to fill the cavity, and V_max avoids material waste and excessive flash. Utilizing the characteristic that the SDF field can easily calculate the sum of distances from any point within the model along a specific direction to both sides of the surface, the thickness value T_min of the thinnest region in the solid part of the model is found through scanning calculations using algorithms (such as ray projection along the normal direction or principal axis of points inside the model). This value must be greater than or equal to the preset minimum allowable wall thickness threshold T_th to avoid defects such as tearing or incomplete filling due to poor material flow or excessively rapid cooling during extrusion. For all surfaces in the model that could potentially serve as the demolding direction (usually the main extrusion direction), their tilt angle relative to that direction is analyzed. By calculating the angle between the gradient direction of the SDF field (i.e., the surface normal) and the demolding direction, and statistically analyzing all surface regions, the minimum draft angle θ_min is obtained. This value must be greater than the preset minimum draft angle threshold θ_th to ensure that the workpiece can smoothly exit the mold cavity after cold extrusion forming, preventing workpiece damage and mold wear caused by sticking or scratching.

[0057] The process of receiving a set of billet process parameters and inputting them into a pre-trained forming quality prediction model to output a score involves using the set of billet process parameters (such as initial diameter D0, height H0, extrusion ratio λ, and allowance δ for key parts) synchronously output by the parameter generation module 102 as a feature vector, and inputting it into a trained machine learning model (such as a gradient boosting decision tree, deep neural network, or support vector regression model). This prediction model is trained on historical production data (including billet parameters, process settings, and final workpiece quality labels, such as no cracks, no folds, dimensional accuracy meeting standards, and good surface finish), and learns a complex nonlinear mapping from billet parameters to potential forming quality risks. The model outputs a comprehensive forming quality prediction score Q_score (e.g., a continuous value from 0 to 100, or a rating of excellent, good, average, or poor). This score quantifies the probability or quality level of obtaining a qualified product when using the currently generated billet scheme for production. In practical applications, generating structured verification results and forming quality prediction scores means that the above analysis process will produce a structured verification report. The report not only includes calculated values ​​for various indicators such as volume, minimum wall thickness, and draft angle, but also clearly records the comparison results (pass or fail) with the corresponding process constraint thresholds, and can further point out specific areas or numerical deviations that fail to meet the standards. The forming quality prediction score Q_score provides a data-driven process feasibility assessment independent of geometric constraints. The combined use of the two constitutes a dual-check mechanism for the generated cold extrusion part scheme, consisting of hard constraint verification and flexible quality prediction. Geometric verification ensures that the scheme is feasible within physical rules and basic manufacturing boundaries; while forming quality prediction, based on historical experience data, assesses the tendency of the scheme to lead to high-quality results in actual production environments. For example, a scheme may pass all hard geometric constraint verifications (volume, wall thickness, and draft angle all meet the standards), but its blank parameter combination (such as a specific high extrusion ratio and a small initial diameter combination) may correspond to a higher risk of cracking in the prediction model, resulting in a lower forming quality prediction score. This comprehensive evaluation enables the system not only to screen out manufacturable schemes, but also to further optimize schemes that are easy to manufacture and of high quality.

[0058] This application's solution achieves real-time, objective, and multi-dimensional engineering feasibility assessment of AI-generated process solutions by integrating automated geometric constraint verification with data-driven quality prediction. Based on volume, wall thickness, and draft angle analysis using SDF data, the solution's compliance with core manufacturing constraints is directly and accurately verified—something traditional methods relying on manual interpretation and subsequent CAE verification struggle to accomplish in real time. Simultaneously, a forming quality prediction model trained on historical data is introduced, encoding and quantifying implicit expert experience and nonlinear process knowledge. This allows for early warning of solutions that, while meeting basic geometric constraints, may have potential process risks. This dual-track evaluation system of geometric verification and process prediction significantly enhances the reliability and practicality of AI-driven process design, elevating solution selection from simple rule-based filtering to experience-based intelligent decision-making, effectively shortening the iteration cycle from design to reliable manufacturing.

[0059] Specifically, in some of the above embodiments, the functions of the verification and prediction module 103 can be further refined. The verification and prediction module 103 performs analysis and prediction, including the following specific steps:

[0060] Volume calculation: Traverse the SDF data volume. For each voxel, if the SDF value at its center point is ≤ 0, then the voxel is determined to belong to the billet entity. The total volume of the billet V_blank = N_inner * (L / N)^3, where N_inner is the total number of internal voxels, L is the normalized spatial physical size, and N is the SDF resolution.

[0061] Minimum wall thickness analysis: First, the triangular mesh of the billet surface is indirectly extracted from the SDF using algorithms such as the moving cube. Then, a large number of points are uniformly sampled inside the model. For each internal sampling point, a ray is projected along the opposite direction of its surface normal (or a predefined main force / flow direction). The SDF value is used to quickly locate the two side surfaces, and the total length of this path is calculated as the local wall thickness. The minimum wall thickness of all sampling points is T_min.

[0062] Draft angle evaluation: Determine the principal demolding direction (e.g., Z-axis). For each vertex of the extracted surface mesh, calculate its unit normal vector n. Then the local draft angle θ at that vertex is 90° - arccos(|n·d|), where d is the unit demolding direction vector. Count the θ values ​​at all vertices; the minimum value is θ_min.

[0063] Verification result generation: The calculated (V_blank, T_min, θ_min) is compared with the preset process constraint threshold list. Output results are as follows: {Volume verification: {Calculated value: V1, Threshold range: [V_min, V_max], Status: Pass / Fail}, Minimum wall thickness verification: {Calculated value: T1, Threshold: T_th, Status: Pass / Fail}, Draft angle verification: {Calculated value: θ1, Threshold: θ_th, Status: Pass / Fail}}. Forming quality prediction steps:

[0064] Feature preparation: The received set of billet process parameters (such as [D0, H0, λ, δ1, δ2, …]) is standardized to form a feature vector X.

[0065] Model prediction: The feature vector X is input into a pre-trained forming quality prediction model M. The model M outputs a predicted value y_hat.

[0066] Scoring Mapping: The predicted value y_hat (which may be a probability value, regression value, or category score) output by the model is converted into an intuitive shaping quality prediction score Q_score through a mapping function f. For example, if y_hat is the predicted pass probability, then Q_score = y_hat * 100; if it is a classification model, then excellent, good, average, and poor can be mapped to 90, 70, 50, and 30 points, respectively.

[0067] Furthermore, the system also includes an optimization module; the optimization module is communicatively connected to the parameter generation module and the verification and prediction module, respectively.

[0068] The optimization module is configured to: receive the verification result and the forming quality prediction score; when the verification result does not meet the preset standard, generate a feedback signal based on the non-compliance item, and convert the feedback signal into an adjustment instruction for the latent space sampling strategy based on the conditional variational autoencoder model, so as to optimize and generate a new set of billet process parameters that simultaneously meet geometric constraints and high forming quality probability.

[0069] Specifically, receiving the verification results and predicted scores and generating feedback signals means that the optimization module continuously monitors the output of the verification and prediction module 103. For structured verification results, it analyzes each item in the report (such as volume, minimum wall thickness, draft angle). If an item fails, a first type of feedback signal is generated. This signal is structured, for example: {Type: Geometric constraints not met, Item: Minimum wall thickness, Calculated value: T1, Required threshold: T_th, Deviation direction: Too thin}. For the forming quality prediction score Q_score, the system presets an acceptable minimum quality score threshold Q_th (e.g., 70 points). If Q_score < Q_th, a second type of feedback signal is generated, for example: {Type: Forming quality prediction too low, Current score: Q_score, Target threshold: Q_th}. These two types of signals may appear individually or simultaneously, together constituting a comprehensive evaluation of the current generated scheme.

[0070] The transformation of feedback signals into adjustment instructions for the latent space sampling strategy involves the optimization module's core function: translating upper-level, understandable engineering feedback (insufficient wall thickness, low quality score) into lower-level, operable generative model control instructions. This is achieved by manipulating the latent space of the conditional variational autoencoder model: For the first type of feedback signal (geometric constraints not met), the deviation direction is analyzed. For example, for excessively thin walls, the optimization module infers the need to generate a geometrically thicker billet variant. This is achieved by directionally adjusting the sampling points in the latent space. Specifically, the optimization module accesses the latent distribution parameters (mean vector μ and variance σ²) output by the conditional variational autoencoder for the current product's SDF data. To increase wall thickness, it might generate an adjustment vector Δz pointing in the latent space in a direction associated with larger volume / thicker wall thickness (this direction can be learned through model analysis or historical data). Then, the new sampling point z_new = z_original + α * Δz, where α is the adjustment step size. This is equivalent to taking a step in the latent space from the original sampling point towards a direction satisfying specific geometric constraints. For the second type of feedback signal (low forming quality prediction score): This usually means that the current combination of [latent vector z, product condition c] maps to a region with poor process feasibility. The optimization module's strategy is to perform exploratory sampling in the latent space while keeping the product condition c constant. It may: (a) increase the randomness of sampling, i.e., resample in the original distribution using variance σ² to obtain a different z'; or, combine some form of gradient information (if differentiable) provided by the graph neural network model to calculate how to fine-tune z to improve the Q_score and move the sampling points in that direction. The goal is to find a different z such that, under the same product condition c, the decoded billet scheme can obtain a higher quality prediction score.

[0071] In practical applications, generating a new scheme based on adjustment instructions means that the optimization module encapsulates the above adjustment logic into specific instructions and sends them to the conditional variational autoencoder model in the parameter generation module 102. For example, the instructions might be: {Operation: Directional adjustment, Original latent vector: z_orig, Adjustment direction: Δz_direction, Adjustment coefficient: α} or {Operation: Exploratory resampling, Distribution parameters: (μ, σ²), Exploration intensity: β}. The parameter generation module 102, based on the received instructions, performs sampling of the adjusted latent vector, and then inputs the new latent vector z_new and the original product SDF condition c back into the decoder and parameter regression branch, thereby generating a new set of directionally optimized cold-extruded part SDF data and billet process parameters. Subsequently, this new scheme is again sent to the verification and prediction module 103 for evaluation. This process can be iterated to form an automated closed-loop optimization cycle of generation-verification / prediction-feedback-optimization-regeneration until a satisfactory solution is generated that simultaneously satisfies all geometric hard constraints (verification passed) and the forming quality prediction score is higher than the threshold (e.g., >85 points), or the preset maximum number of iterations is reached.

[0072] Furthermore, the verification and prediction module includes a graph neural network model;

[0073] The process of receiving the set of process parameters for the billet and outputting a forming quality prediction score includes:

[0074] The three-dimensional voxel mesh corresponding to the SDF data of the cold-extruded part is transformed into a graph structure; wherein each voxel is a node of the graph, and the node features include the SDF value, coordinate position and local gradient of the voxel; the set of billet process parameters is fused with the features of each node as global graph features;

[0075] After message passing and node feature updating through a multi-layer graph convolutional network in the graph neural network model, global information is aggregated through a graph pooling layer, and finally the forming quality prediction score is obtained through regression by a fully connected layer.

[0076] Specifically, converting the SDF data of cold-extruded parts into a graph structure and integrating the billet process parameters means: based on a three-dimensional voxel mesh with a fixed resolution (e.g., 256x256x256), each non-empty voxel (usually referring to voxels inside or near the surface to reduce computational cost) is defined as an independent graph node. A feature vector h_i^(0) is constructed for each node i. This initial feature includes at least: SDF value: the signed distance function value at the center point of the voxel, directly representing the position of the point relative to the billet surface (inside, surface, or outside). Coordinate position: the (x, y, z) coordinates of the voxel in normalized three-dimensional space, providing absolute spatial information. Local gradient: the gradient vector (gx, gy, gz) or its magnitude and direction, approximated by calculating the difference in SDF values ​​of the voxel's surrounding neighborhood. This gradient information characterizes the degree of drastic change in the orientation and shape of the local surface, and is a key clue reflecting stress concentration and the ease or difficulty of material flow. Meanwhile, the set of billet process parameters (such as initial diameter D0, height H0, extrusion ratio λ, etc.) is regarded as a global feature vector g related to the entire billet geometry. When constructing the graph, this global feature vector g is fused with the initial feature h_i^(0) of each node through broadcasting or concatenation (for example, g is concatenated as an additional feature dimension after each h_i^(0)), thereby ensuring that each node knows the overall process parameter conditions at the beginning of feature update. The edges of the graph are defined according to the spatial adjacency relationship between voxels (such as 6-adjacency, 18-adjacency, or 26-adjacency) to capture the local connectivity of the three-dimensional geometry.

[0077] The process of message passing, aggregation, and final regression scoring using a graph neural network involves inputting the constructed graph data, which incorporates global process parameters, into a multi-layer graph convolutional network. In each layer of graph convolution, each node receives feature information (messages) from its neighboring nodes and updates it based on its own features. This process can be formalized as: h_i^(l+1) = UPDATE( h_i^(l), AGGREGATE({h_j^(l), ∀j ∈ N(i)}) ), where h_i^(l) represents the feature of node i in layer l, and N(i) is the set of neighboring nodes of node i. Through multiple iterations (e.g., 3-5 layers), the node features not only contain their own local geometric information but also gradually aggregate broader contextual geometric information from multi-hop neighbors (such as the relationship between a thin-walled region and its surrounding thick regions). After L layers of graph convolution, a node feature set {h_i^(L)} containing rich local and semi-global geometric-process correlation information is obtained. Subsequently, a graph pooling layer (such as global average pooling, global max pooling, or attention pooling) is used to aggregate the final features h_i^(L) of all nodes, generating a fixed-length graph-level feature vector H_global that represents the global state of the entire cold-extruded part. This vector integrates all the key features of the entire billet geometry under specific process parameters. Finally, H_global is input into a regression network consisting of several fully connected layers. This network decodes the high-dimensional global feature vector into a single, continuous-valued forming quality prediction score Q_score through learned nonlinear mappings. This score directly reflects the expected potential of the combination of billet geometry and process parameters to achieve high-quality forming in subsequent cold extrusion processes.

[0078] In practical applications, using graph neural network models for quality prediction offers significant advantages over traditional methods or simple fully connected networks. The graph structure perfectly matches the spatial discreteness and adjacency relationships of 3D voxel data, enabling the model to explicitly utilize and infer the topological structure and spatial configuration of 3D geometry. The message passing mechanism between nodes allows the model to capture the complex physical processes by which local defects (such as an overly sharp corner) affect the surrounding area and even the overall forming quality through stress transmission. Integrating billet process parameters as global features with each node ensures that the model analyzes geometric features within a specific process condition context, thereby enabling the evaluation of different quality performances of the same geometry under different process parameters (such as different extrusion ratios).

[0079] This application's solution introduces a graph neural network as the forming quality prediction model, achieving a deep comprehensive analysis of cold-extruded parts, from overall parameters to micro-geometry. It transforms the SDF voxel data of the billet into a graph, enabling the algorithm to examine both local details (such as fillet size and wall thickness transition) and the overall structural layout, much like a process expert reviewing 3D drawings. It also understands the interplay of these geometric elements under given process parameters. The message-passing mechanism of graph convolution essentially simulates the transmission and interaction of stress and strain information during material deformation, giving the prediction model a preliminary physical perception capability. Therefore, the final forming quality prediction score is not merely based on statistical correlations of historical data, but also incorporates a deep structural understanding of the current 3D geometry, significantly improving prediction accuracy, robustness, and generalization ability for unknown solutions. This provides a powerful technical tool for quickly and reliably selecting the optimal solution from a massive pool of AI-generated solutions.

[0080] Specifically, in some of the above embodiments, the steps regarding prediction by the graph neural network model can be further refined.

[0081] The specific steps for outputting the forming quality prediction score through the graph neural network model include: Graph construction step: Node generation and feature initialization: Traversing the SDF voxel mesh of the cold extruded part. For voxels with SDF values ​​less than or equal to a certain threshold τ (e.g., τ = 1 voxel unit), they are considered valid voxels and nodes are created. An initial feature vector is created for node i: h_i^(0) = [sdf_i, x_i, y_i, z_i, |∇sdf|_i, nx_i, ny_i, nz_i]. Wherein, sdf_i is the SDF value; (x_i, y_i, z_i) are the normalized coordinates; |∇sdf|_i is the magnitude of the SDF gradient; (nx_i, ny_i, nz_i) is the normalized gradient direction (approximately the surface normal). Edge definition: For each node, find other valid voxel nodes in its 26 neighborhoods (up, down, left, right, front, back, and all diagonal directions) and establish undirected edges. Global feature fusion: Normalize the set of billet process parameters (e.g., [D0, H0, λ, δ]) to obtain the global feature vector g. Concatenate g with the initial feature h_i^(0) of each node to obtain the fused node features: h_i^(0)' = CONCAT(h_i^(0), g). Graph convolution and update steps (perform L layers, e.g., L=4): For the l-th layer (l from 0 to L-1): Message generation: For each edge from node j to node i, generate the message m_{j->i} = MESSAGE(h_j^(l)), where MESSAGE is usually a learnable linear transformation or neural network. Neighborhood Aggregation: For node i, aggregate the messages sent by all its neighboring nodes j ∈ N(i): a_i^(l) = AGGREGATE({m_{j->i}, ∀j ∈ N(i)}), where AGGREGATE can be a summation, mean pooling, or max pooling. Node Update: Combining the node's own features and the aggregated neighborhood information, update the node features: h_i^(l+1) = UPDATE(h_i^(l), a_i^(l)), where UPDATE is usually a learnable neural network (such as a fully connected layer followed by an activation function). Graph Pooling and Regression Steps: Global Pooling: After L layers of graph convolution, the final node features {h_i^(L)} are obtained. Global attention pooling is used: the attention weight α_i = softmax(MLP_att(h_i^(L))) for each node is calculated, and then a weighted sum is obtained to obtain the graph-level feature vector: H_global = Σ (α_i * MLP_feat(h_i^(L))). Score regression: H_global is input into a multilayer perceptron regressor: Q_score = MLP_regressor(H_global). The MLP_regressor outputs a scalar value, which is the forming quality prediction score.

[0082] Furthermore, the decoder corresponding to the conditional variational autoencoder model adopts a multi-task learning architecture, which includes an SDF reconstruction branch and a parametric regression branch. The SDF reconstruction branch is used to predict signed distance values ​​on a voxel-by-voxel basis to form the SDF of the cold extruded part. The parametric regression branch shares hidden layer features with the SDF reconstruction branch in the intermediate layer of the decoding process, and uses an independent regression head network to map the shared features to the billet process parameter set. The billet process parameter set includes at least the billet diameter, length, volume compensation coefficient, and material identification code.

[0083] Specifically, in this embodiment, the decoder employs a multi-task learning architecture with a shared backbone. This means that the core of the decoder is a deep deconvolutional neural network (backbone network), whose input is a fused representation of the sampled latent vector z and the product SDF condition c. This backbone network progressively expands the low-dimensional fused representation into a high-resolution feature map through a series of deconvolutional and upsampling layers. In the intermediate layers of the backbone network (e.g., when decoding to 1 / 4 or 1 / 8 resolution), the feature map is simultaneously fed to two independent branches:

[0084] SDF Reconstruction Branch: This branch inherits intermediate features from the backbone network and continues processing through subsequent deconvolutional layers, upsampling layers, and a final 1x1 convolutional layer. Finally, at the output, this branch outputs a scalar value for each target voxel grid location (e.g., 256x256x256), which is the predicted signed distance from that location to the surface of the cold-extruded part. The predicted values ​​of all voxels collectively constitute the 3D SDF data of the cold-extruded part. The learning objective of this branch is to accurately reconstruct the geometry.

[0085] Parametric Regression Branch: This branch branches off from the backbone network at the same intermediate layer. Instead of participating in subsequent spatial upsampling and detail reconstruction, it focuses on extracting global information related to the overall billet properties from the current intermediate feature map. To this end, the parametric regression branch first compresses the spatially dimensional feature map into a one-dimensional global feature vector through a global average pooling (GAP) or global max pooling layer. This global feature vector condenses the geometric and contextual information of the entire billet at the current stage (corresponding to a specific decoding depth).

[0086] The parameter regression branch outputs a set of billet process parameters through an independent regression head network. This means that after obtaining the global feature vector, the parameter regression branch inputs it into a fully connected layer network (regression head) independent of the SDF reconstruction branch. This regression head network typically consists of 2 to 3 fully connected layers, with the final layer having a number of neurons equal to the number of billet process parameters to be output. The activation value of each neuron directly corresponds to the predicted value of a process parameter.

[0087] Specifically, the set of billet process parameters includes at least:

[0088] Blank diameter (D_blank) and length (L_blank): These are the two most basic dimensions that define the initial shape of a cylindrical blank. They directly determine the volume and shape of the blank and are the basis for blanking and mold design.

[0089] Volume compensation coefficient (α_vol): This is a crucial process compensation parameter. Due to factors such as incomplete material filling, die elastic deformation, workpiece cooling shrinkage, and subsequent machining allowances during cold extrusion, the theoretical calculated volume of the billet often differs from the net volume of the final product. The volume compensation coefficient α_vol is used to adjust the theoretical volume and is typically expressed as: Actual billet volume = Theoretical product net volume × α_vol (where α_vol > 1). This parameter is automatically predicted by the model based on factors such as product complexity and material properties.

[0090] Material ID: This is a classification or embedding vector used to identify the type of billet material to be used (e.g., aluminum alloy A6061, steel 20Cr, etc.). Different materials have different flow stresses, hardening behaviors, and forming limits, directly affecting the formulation of process plans. The model learns to associate material selection with product geometry.

[0091] In practical applications, this multi-task learning architecture offers significant advantages. The shared backbone network ensures that the two branches share the same understanding of the input conditions (z and c) and extract common features of the generated target during the early and mid-stages of decoding. This forces the model to learn a unified and compact feature representation that can effectively reconstruct geometry and accurately infer global parameters. The bifurcation design allows the two branches to focus on their respective tasks: the SDF reconstruction branch continues to refine spatial details, while the parametric regression branch directly reads global parameters from relatively abstract but information-rich intermediate features, avoiding the difficulty of inferring parameters from ultra-high-dimensional final voxel data.

[0092] More importantly, this architecture performs end-to-end optimization during training using a joint loss function (such as L1 / L2 loss for SDF reconstruction + mean squared error loss for parametric regression), enabling the model to spontaneously establish a deep and quantitative intrinsic relationship between geometric features and process parameters. For example, the model learns that when intermediate feature maps indicate the presence of deep cavities or thin-walled regions in the product, the parametric regression branch should tend to predict a slightly larger volume compensation coefficient (α_vol) to avoid incomplete filling; or, when feature maps show a complex product shape, materials with better forming performance (such as a specific Material_ID) should be recommended. This relationship is data-driven, implicit, yet highly synergistic.

[0093] This application's solution innovatively achieves integrated and collaborative generation from a single conditional input to dual geometric and process outputs by employing a multi-task learning decoder architecture. The SDF reconstruction branch ensures the geometric evolution logic of the generated cold-extruded part shape and the target product; the parameter regression branch ensures the feasibility of the manufacturing parameters associated with this shape in terms of process logic. Both share underlying features, guaranteeing the self-consistency of the output results—that is, the generated billet shape (SDF) can be efficiently formed through the cold extrusion process using the predicted billet dimensions (D, L), after volume compensation (α_vol), and with the suggested material (Material_ID). This completely changes the problem in traditional processes where geometric design and process parameter determination are often performed separately and may be disconnected, merging the two-step decision-making into a single intelligent reasoning step, greatly improving the completeness, consistency, and practicality of the generated solution, and laying a high-quality data foundation for subsequent automated verification and optimization.

[0094] Specifically, in some of the above embodiments, the specific implementation of the decoder multi-task learning architecture can be further refined. The specific implementation steps of the decoder multi-task learning architecture include:

[0095] (1) Forward propagation steps of the shared backbone network: Input the fused conditional latent code [z; c] (dimension [B, C_latent], where B is the batch size). Project it to the initial spatial size (e.g., [B, 512, 4, 4, 4]) through the first fully connected layer. Upsample through L (e.g., L=5) deconvolution blocks (each block contains deconvolution, batch normalization, and activation function) to gradually increase the spatial resolution. Let the feature map output by the l-th deconvolution block be F_l. Branch point operation steps: After a specific k-th deconvolution block (e.g., when the spatial resolution of feature map F_k reaches [B, C_k, 16, 16, 16]): SDF reconstruction branch path: Input F_k into the subsequent (Lk) deconvolution blocks and continue upsampling to the target resolution (e.g., 256). Finally, by passing a 1x1x1 convolutional layer (without an activation function or by using Tanh or other methods to limit the output range), the number of channels is compressed to 1, resulting in the final SDF prediction tensor SDF_pred (dimension [B, 1, 256, 256, 256]).

[0096] Parametric regression branch path:

[0097] a. Feature extraction and pooling: For the current feature map F_k, first, channel adjustment and feature refinement are performed through an additional 1x1x1 convolutional layer to obtain F_k'.

[0098] b. Global information aggregation: Apply global average pooling (GAP) to F_k' to aggregate the feature maps of [B, C_k', 16,16, 16] into a global feature vector g_global of [B, C_k'].

[0099] c. Parameter Regression: Input g_global into the parameter regression head network. An example network is: FC1: Linear(C_k', 256) -> ReLU -> Dropout -> FC2: Linear(256, 128) -> ReLU -> FC3: Linear(128, N_params). Here, N_params is 4 (corresponding to the encoding of D_blank, L_blank, α_vol, and Material_ID). For Material_ID, if it's for classification, the last layer can use Softmax; if it's for regressing its embedding vector, it's a continuous value. The final output is the parameter vector Params_pred.

[0100] (2) Training and Loss Calculation Steps: The model is trained by minimizing the joint loss function: Total_Loss = λ_sdf * L_sdf(SDF_pred, SDF_gt) + λ_param * L_param(Params_pred, Params_gt). Here, L_sdf can be a smoothed L1 loss, and L_param uses mean squared error loss for continuous parameters (D, L, α) and cross-entropy loss for Material_ID. λ_sdf and λ_param are hyperparameters that balance the weights of the two tasks.

[0101] Furthermore, the optimization module is configured to perform the following specific optimization process:

[0102] If the verification result indicates a volume mismatch, the ratio of the current cold-extruded part volume to the target volume is calculated, and vector interpolation or perturbation sampling is performed in the potential space of the parameter generation module along the direction that makes the ratio approach 1 to generate a new set of billet process parameters.

[0103] If the forming quality prediction score is lower than a preset threshold, the gradient of the forming quality prediction score with respect to the potential vector is calculated using the differentiable property of the prediction model in the verification and prediction module. Then, an iterative search is performed in the potential space along the gradient ascent direction to find potential vectors that can generate higher prediction scores, thereby generating an optimized set of billet process parameters.

[0104] Specifically, in this embodiment, the optimization process for volume mismatch refers to the following: The optimization module first parses the verification results to obtain the calculated volume V_current of the current cold-extruded part and the preset target volume range [V_min, V_max]. The volume deviation is calculated: if V_current < V_min, the volume is insufficient; if V_current > V_max, the volume is excessive. The optimization module calculates a target volume adjustment factor β. For example, for insufficient volume, β can be set to sqrt(V_min / V_current) (slightly greater than 1) to increase the volume to near the target lower limit; for excessive volume, β can be set to sqrt(V_max / V_current) (less than 1).

[0105] Next, the optimization module performs operations in the latent space. It knows the latent vector z_orig corresponding to the current scheme, and the latent distribution mean μ of the conditional variational autoencoder's output of the SDF data for the same product. μ can be considered as the anchor point or most likely representation of the product in the latent space. The optimization strategy is as follows:

[0106] Vector interpolation: Generates a new latent vector z_new = μ + (z_orig - μ) * β. The principle is to scale along the ray from the distribution center μ to the current sampling point z_orig according to β. If the volume needs to be increased (β > 1), it is extrapolated along this direction; if the volume needs to be decreased (β < 1), it is contracted back towards the center μ. This method assumes that volume changes are roughly monotonically related to movement in the latent space along a specific direction (from μ to z_orig).

[0107] Perturbation sampling method: Based on the original latent vector z_orig, a perturbation related to the volume deviation is applied. For example, z_new = z_orig + η * (β - 1) * d, where η is the step size coefficient, and d is a preset or learned latent space direction vector related to volume increase or decrease (which can be associated with the aforementioned d_geom direction). The strength and direction of the perturbation are controlled by adjusting β.

[0108] Regardless of the method used, the new latent vector z_new is then input into the decoder to generate a new round of SDF data and parameters for cold extrusion parts. The goal is to systematically correct volume deviations while keeping other geometric features roughly unchanged.

[0109] The optimization process for addressing the low predicted forming quality score involves initiating a more refined, gradient-based search when Q_score < Q_th. This process leverages a key property: the latent vector z leading to the predicted forming quality score Q_score is differentiable throughout the generation-evaluation chain. The specific path is: z → Conditional Variational Autoencoder Decoder → Cold Extruded Part SDF Data → Graph Neural Network (GNN) → Q_score.

[0110] Gradient Calculation: The optimization module fixes all network parameters (decoder, GNN) and sets the current latent vector z_current as a variable. Using automatic differentiation, backpropagation calculates the gradient ∇_zQ of the forming quality prediction score Q_score with respect to the latent vector z_current. This gradient ∇_zQ is a vector whose direction indicates the direction in which fine-tuning z_current in the latent space results in the fastest increase in Q_score.

[0111] Iterative search: Based on this gradient, the optimization module performs one (or more) gradient ascent updates:

[0112] z_new = z_current + γ * (∇_z Q / ||∇_z Q||). Here, γ is the learning rate or step size, used to control the magnitude of the update. Gradient normalization is often used for stability and to avoid oscillations.

[0113] Evaluation and Iteration: Generate a new solution using the updated z_new and recalculate Q_score_new. If Q_score_new is still below the threshold and there is still room for improvement, repeat steps 1 and 2 for multiple iterations until the score meets the target or the maximum number of iterations is reached. This process is similar to hill climbing in the latent space, with the goal of finding regions that result in higher prediction quality scores.

[0114] In practical applications, these two optimization mechanisms complement each other, forming a hierarchical and refined optimization system. Volume optimization targets explicit and quantifiable violations of physical constraints, employing deterministic / semi-deterministic adjustments based on proportions or rules to quickly and directly correct obvious flaws. Quality scoring optimization, on the other hand, targets more complex and nonlinear process feasibility problems, using gradient-based numerical optimization methods. It does not directly specify how to modify (e.g., where to thicken), but rather uses the knowledge within the model (gradients) to guide the search direction, aiming to explore variants with higher success potential.

[0115] This design embodies a combination of constraint-driven and goal-driven optimization. When a solution violates hard constraints (such as volume), the system prioritizes and quickly satisfies these constraints; when a solution satisfies all hard constraints but has poor soft quality metrics, the system focuses on improving its overall quality. Both are achieved by operating on the source of the generative model—the latent vectors—ensuring that optimization actions directly affect the core of the generation process, thereby efficiently guiding solution iteration.

[0116] This application's solution, through refining the latent space optimization strategy, endows the system with powerful adaptive and self-improving capabilities. For the interpolation / perturbation method addressing volume mismatch, it cleverly maps an external physical quantity (volume) to a geometric transformation operation in the latent space, enabling volume correction without compromising the overall shape semantics. Gradient-based quality scoring optimization deeply couples the data-driven predictive model (GNN) with the generative model (CVAE), forming a differentiable, end-to-end generation-evaluation loop. This allows the system not only to evaluate solutions but also to use evaluation signals to accurately infer how to improve them. This gradient-based optimization essentially utilizes the complex knowledge encoded by a trained neural network about which geometric and process combinations are more likely to lead to high quality to guide the generation process, representing an efficient and intelligent design space exploration. The combined effect of these two methods enables the system to automatically and efficiently iterate from initial, potentially imperfect generation results to arrive at excellent process solutions that simultaneously meet geometric constraints and high-quality requirements.

[0117] Specifically, in some of the above embodiments, the specific steps of the optimization process can be further refined. The specific execution steps of the optimization process include: Scenario 1: Optimization of volume mismatch.

[0118] Input and Judgment: Receive the verification result and extract {Volume Verification: {Calculated Value: V_c, Status: Failed, Deviation Direction: s}}, where s represents insufficient or excessive volume. The preset target volume is V_target (which can be the median of V_min or V_max, or a boundary value selected based on the deviation direction).

[0119] Calculate the adjustment factor: If s is insufficient, β = 1 + λ_v * (V_target - V_c) / V_c; if s is excessive, β = 1 - λ_v * (V_c - V_target) / V_c. λ_v is the volume adjustment intensity coefficient (0 < λ_v <= 1).

[0120] Latent space operation: Scheme A (linear interpolation): z_new = (1 - β) * μ + β * z_current. When β > 1 (volume needs to be increased), move further in the z_current direction; when β < 1 (volume needs to be decreased), shrink back in the μ direction.

[0121] Option B (perturbation along a specific direction): z_new = z_current + (β - 1) * d_vol. d_vol is a predefined latent space unit direction vector, either defined or learned from an auxiliary model, that is strongly correlated with the overall size scaling.

[0122] Generate and re-validate: Generate a new scheme using z_new and perform volume validation again until it passes or reaches the maximum number of attempts for this single optimization.

[0123] Scenario 2: Optimization of low forming quality score

[0124] Input and Decision: Receive the predicted score Q_current and the threshold Q_th. If Q_current < Q_th, enter the iterative optimization loop and set the maximum number of iterations N_iter.

[0125] Single-step iterative process:

[0126] a. Forward Propagation and Gradient Calculation: Input the current z_current into a fixed decoder to obtain SDF_blank. Construct a graph using SDF_blank and the corresponding process parameters, input it into a fixed GNN, and perform forward propagation to obtain Q_current. Keep z_current set to require_grad=True, perform backpropagation, and calculate ∇_zQ.

[0127] b. Gradient processing and update: Prune ∇_zQ (e.g., limit the maximum norm) and normalize it: g_norm = ∇_zQ / (||∇_zQ|| + ε). Update the latent vector: z_current := z_current + γ * g_norm. Where γ is the adaptive step size, which can be dynamically adjusted according to the score improvement.

[0128] c. Evaluation: Generate a new scheme using the updated z_current and compute the new Q_new.

[0129] d. Judgment: If Q_new >= Q_th or the number of iterations reaches N_iter, then exit the loop; otherwise, return to step a.

[0130] Output: The solution corresponding to z_final, which is obtained from the last iteration and has the highest score, is used as the optimized output.

[0131] Furthermore, the scale normalization process includes: calculating the maximum outer envelope size of the three-dimensional product model in the three coordinate axes of the three-dimensional space, and using the length of the longest side in the maximum outer envelope size as the normalization reference, scaling the coordinates of all vertices of the three-dimensional product model proportionally so that the length of the longest side of the scaled model is equal to the preset reference length value.

[0132] Specifically, in this embodiment, calculating the axial outer envelope dimension of the 3D product model and determining its maximum value means that after completing geometric repair (ensuring the model is a closed entity) and coordinate alignment (the model is in a known standard coordinate system), the system traverses all vertices of the model (each vertex of the triangular mesh). For the model vertex set V = {v_i =(x_i, y_i, z_i)}, its extreme values ​​in the X, Y, and Z coordinate axes are calculated respectively.

[0133] The maximum value in the X direction is X_max = max({x_i}), and the minimum value is X_min = min({x_i}).

[0134] The maximum value in the Y direction is Y_max = max({y_i}), and the minimum value is Y_min = min({y_i}).

[0135] The maximum value in the Z direction is Z_max = max({z_i}), and the minimum value is Z_min = min({z_i}).

[0136] Therefore, the outer envelope dimensions (i.e., the side lengths of the axial bounding box) of the model in the three principal axis directions are: L_x = X_max - X_min; L_y = Y_max - Y_min; L_z = Z_max - Z_min.

[0137] Next, select the maximum value from these three side lengths: L_max = max(L_x, L_y, L_z). L_max is the longest external dimension of the product model in the current spatial orientation, which represents the model's main extension scale in space.

[0138] The process of calculating the scaling ratio based on the baseline scale and a preset baseline length value and then uniformly scaling means that the system presets a standard baseline length value L_base (e.g., L_base = 1.0 unit length, or L_base = 100.0 corresponding to a standardized 100-unit system). This L_base serves as a common reference for all subsequent model comparisons and processing at a uniform scale. The scaling factor s is calculated as: s = L_base / L_max. The significance of this scaling factor s is that if all dimensions of the model (including all vertex coordinates) are multiplied by s, the longest side dimension of the model in space will change from the original L_max to L_max * s = L_base, which is exactly equal to the preset baseline length value.

[0139] Subsequently, a uniform scaling transformation is performed on all vertices v_i = (x_i, y_i, z_i) of the model: v_i' = (x_i * s, y_i * s, z_i * s), for all i. After this transformation, all geometric features of the model (shape, scale, angles) are preserved because the scaling is uniform. The only change is the absolute size of the model, with its longest side normalized to a uniform L_base. The processed model will then exist in a scale-normalized coordinate system.

[0140] In practical applications, this longest-side normalization method has clear engineering significance and computational advantages. From an engineering perspective, cold extrusion process design focuses on the geometric similarity of shapes and the relative proportions of dimensions (such as wall thickness ratio and depth-to-width ratio), rather than absolute dimensions. Unifying the longest side means that camera housings of different sizes (such as a larger housing for forward viewing and a smaller housing for surround viewing) are aligned to the same "scale" for comparison and processing. This allows subsequent neural network models (such as conditional variational autoencoders and graph neural networks) to focus on the topological and structural information of the shape itself when extracting features and learning patterns, without having to learn or adapt to the huge scale differences between different models. This greatly reduces the complexity of model learning and improves generalization ability.

[0141] From a computational perspective, normalization to a fixed range (such as [0,1] or [-1,1] around the origin) is a standard requirement for many numerical algorithms and deep learning models. Normalization based on the longest side ensures that the entire model is contained within a cubic space centered at the origin with a side length not exceeding L_base. This is crucial for subsequent conversion to a fixed resolution (such as 256^3) SDF, as it guarantees that regardless of the size of the original model, the normalized model can be effectively and untruncatedly sampled into the voxel grid, thus avoiding information loss due to parts of the model exceeding the sampling range.

[0142] This application's solution establishes a stable and consistent geometric data benchmark for the entire intelligent process design flow by employing a scale normalization method based on the longest side. It solves the problem of data scale chaos caused by the diverse product series and varying sizes of automotive camera housings, providing a unified scale input for subsequent modules such as model encoding, shape generation, geometric verification, and quality prediction. This normalization is not simply data preprocessing, but a crucial step in mapping the continuous size spectrum in the physical world to a standardized discrete representation space suitable for computation and machine learning models. It is precisely because all models undergo this rigorous and consistent scale normalization that the learning-based parameter generation module can extract common, scale-independent "shape-process" mapping rules from "large" and "small" housings of different sizes. It also ensures that thresholds (such as minimum wall thickness) in the verification module have a unified physical meaning, thereby enabling the entire system to robustly handle product design tasks of various sizes and specifications.

[0143] Specifically, in some of the above embodiments, the specific steps regarding scale normalization processing can be further refined. The specific execution steps of the scale normalization processing include:

[0144] Input and preparation steps: The input object is a 3D product model that has been geometrically repaired and aligned, represented as a triangular mesh M = (V, F), where V is an Nx3 vertex coordinate matrix and F is an Mx3 face index matrix.

[0145] Bounding box calculation steps: Calculate vertex coordinate extrema: Parallelize or vectorize the calculation of each column of the vertex matrix V (corresponding to x, y, z coordinates respectively): min_xyz = [min(V[:,0]), min(V[:,1]), min(V[:,2])]; max_xyz = [max(V[:,0]), max(V[:,1]), max(V[:,2])]; Calculate the axial bounding box dimensions: L_x = max_xyz[0] - min_xyz[0]; L_y = max_xyz[1] - min_xyz[1]; L_z = max_xyz[2] - min_xyz[2]. Create the dimension vector L_vec = [L_x, L_y, L_z].

[0146] Determine the baseline and calculate the scaling steps: Determine the longest side: L_max = max(L_vec). Record the axial index idx_max (0, 1, or 2) corresponding to the longest side. Obtain the preset baseline: Read the preset baseline length value L_base from the system configuration (e.g., L_base = 1.0). Calculate the scaling factor: scale_factor = L_base / L_max. To ensure numerical stability, a small protection factor can be added: scale_factor = L_base / (L_max + epsilon), where epsilon is a very small positive number (e.g., 1e-8).

[0147] Model vertex coordinate transformation steps: Calculate the translation (optional but recommended): To ensure the normalized model is approximately centered, first calculate the bounding box center: center_xyz = (min_xyz + max_xyz) / 2.0. However, note that if the previous coordinate alignment step has already placed the model in the ideal position (e.g., the centroid is at the origin), this step can be omitted or adjusted. Perform scaling transformation: If performing translation centering: First, V_centered = V - center_xyz (broadcast subtraction). Then, V_normalized = V_centered * scale_factor. If not performing translation centering: directly V_normalized = V * scale_factor.

[0148] Output: Output the scaled vertex matrix V_normalized and the invariant patch matrix F, forming the scale-normalized 3D model M_normalized. Simultaneously, record the scaling factor (scale_factor) used in this operation, which can be used in subsequent steps requiring restoration of the original scale (such as when outputting final process parameters).

[0149] Specifically, this series of steps achieves a deterministic, reversible geometric transformation. The core idea is to maintain shape similarity. By scaling the longest side to a fixed value, all other dimensions of the model (such as the lengths of the other two sides, wall thickness, aperture, etc.) are scaled by the same scale_factor. This is equivalent to defining a unit "1" for each model based on its own maximum size. All models use their own "1" to correspond to the system's preset "1" (L_base), thus achieving cross-model scale alignment. This step is crucial in the context of actual process design. For example, a shell with a length of 200mm and another with a length of 100mm, after normalization, both are represented in the system as geometric objects with a length of "1 standard unit". When the subsequent AI model generates cold-extruded parts for these objects, it learns "what kind of 'unit blank' shape and process parameters should correspond to this aspect ratio and this feature distribution of 'unit shape'". When a real, physically dimensional process design is required, the system combines the original scale_factor or product model information to scale the "unit blank" dimension back to the physical world. Essentially, scale normalization removes the diversity of physical dimensions from the fundamental characteristics of geometric shapes, allowing AI to focus on learning the more fundamental mapping relationship between shape and process, thereby significantly improving the versatility and intelligence of the design system.

[0150] Furthermore, the system also includes a human-computer interaction and decision visualization module, which is communicatively connected to the parameter generation module, the verification and prediction module, and the optimization module, and is configured to execute:

[0151] The latent space based on the conditional variational autoencoder model in the parameter generation module is dimensionality-reduced and mapped. In the visualization interface, the three-dimensional models of cold extruded parts generated by different latent points, the corresponding geometric constraint verification results output by the verification and prediction module, and the forming quality prediction score are displayed in association.

[0152] The system receives user selection and annotation operations for specific regions or directions in the potential space from the visualization interface, generates corresponding interactive instructions, and provides these interactive instructions as initial sampling preferences or constraints to the optimization module.

[0153] Specifically, in this embodiment, visualizing and mapping the latent space and associating it with display scheme information means that the module first obtains a set of latent vector samples {z_i} (which may come from initial generation or historical optimization iterations) generated by the conditional variational autoencoder model in the parameter generation module 102 for the same target product. These latent vectors are usually high-dimensional vectors of tens to hundreds of dimensions and cannot be directly visualized. Therefore, the module applies a dimensionality reduction algorithm (such as principal component analysis (PCA) or t-distributed random neighborhood embedding (t-SNE)) to project each high-dimensional latent vector z_i onto a two-dimensional or three-dimensional visualization space to obtain its corresponding low-dimensional coordinate point p_i.

[0154] In the visualization interface, each coordinate point p_i represents a specific process design scheme. When the user hovers the mouse over or clicks a point p_i, the module performs the following related displays: 3D model display: The decoder of the parameter generation module 102 is called in real time to combine the corresponding latent vector z_i with the product SDF condition c, quickly generating (or reading from the cache) the corresponding 3D model of the cold extruded part, and rendering it in the 3D view window. This allows the user to intuitively see the geometry of the billet corresponding to different positions in the latent space. Evaluation information overlay: At the same time, the module obtains the structured verification results and forming quality prediction score Q_score corresponding to the scheme from the verification and prediction module 103. This information is overlaid on the visualization interface in a graphical way: Geometric constraint verification: can be represented by icon color or shape. For example, a green checkmark indicates that all constraints are passed; a red "thickness" icon indicates that the minimum wall thickness does not meet the standard; a yellow "angle" icon indicates that the draft angle is close to the threshold. Quality prediction score: can be displayed by color gradient of points (such as from red [low score] to green [high score]) or by directly labeling numbers. In this way, users can simultaneously observe the distribution (location in latent space), appearance (three-dimensional geometry), and performance (constraint compliance and quality prediction) of design schemes in a unified view, thereby establishing a holistic understanding of the AI-generated design space.

[0155] The "receiving user interaction and converting it into optimization instructions" module provides rich interactive tools, allowing users to actively guide the optimization process based on visual information: Region Selection: Users can use the lasso or box selection tools to select a potential space region of interest on a two-dimensional visual scatter plot. For example, a user might find that although a certain region's points don't have the highest quality score, the resulting billet shape is more regular (e.g., closer to a simple solid of revolution). After selection, the user can annotate "Prefer this region's shape." Directional Annotation: Users can manually draw a vector line in the visual space, or select two significantly different points A and B and annotate "Evolution from A to B." This might represent a beneficial design trend judged by the user based on experience (e.g., "gradually increasing wall thickness" or "gradually optimizing draft angle"). Constraint Setting: Users can annotate a specific solution point that has some advantages but also defects, such as "Retain the overall outline of this solution, but increase the volume by 10%" or "Based on this solution, focus on optimizing its inner draft angle."

[0156] In practical applications, the above interactive operations are transformed into structured interactive instructions by the module. For example: {“Instruction Type”: “Region Preference Sampling”, “Region Coordinates”: [[x_min, y_min], [x_max, y_max]], “Number of Samples”: N}; {“Instruction Type”: “Direction-Guided Optimization”, “Starting Latent Vector”: z_A, “Direction Vector”: d_AB, “Optimization Step Size”: step}; {“Instruction Type”: “Reference Point Constraint Optimization”, “Reference Latent Vector”: z_ref, “Constraint”: {“Volume Adjustment”: +0.1, “Focus Feature”: “Inner Draft Angle”}}. These interactive instructions are sent to the optimization module in real time. The optimization module then translates these high-level instructions into specific adjustments to the latent space sampling or search strategy: for region preference sampling, the optimization module may perform focused sampling within the high-dimensional subspace of the latent space corresponding to that region. For direction-guided optimization, the optimization module can map the d_AB direction back to an approximate direction D in the high-dimensional latent space and perform gradient ascent or interpolation search along this direction, similar to the previously mentioned process. For reference-point-constrained optimization, the optimization module can start from z_ref and, in conjunction with user-specified constraints (such as volume adjustment), perform targeted optimization iterations. This allows the user's domain knowledge and experience to be seamlessly integrated into the automated optimization process, enabling hybrid intelligent decision-making with a human-in-the-loop approach.

[0157] This application's solution, by introducing a human-computer interaction and decision visualization module, breaks through the limitations of the black-box decision-making in traditional automated systems, achieving transparency and guidance in the intelligent process design process. Visual mapping transforms the high-dimensional, abstract latent space into two-dimensional / three-dimensional charts that engineers can intuitively perceive and understand. Correlation display tightly integrates the geometric form of the solution with engineering evaluation indicators, greatly improving the efficiency of solution review and comparison. More importantly, the interactive function empowers users (process experts) with the ability to actively intervene and guide based on AI's automatic exploration. Experts can quickly identify potentially valuable solution areas or optimization directions that AI might overlook, based on their experience, and convey their intentions to the system through simple interactive operations. The system then translates these intentions into specific mathematical instructions, driving the optimization module to perform targeted searches. This collaborative model of "AI wide-area exploration + human expert key guidance" leverages AI's advantage of rapidly traversing and discovering patterns in massive amounts of data, while also incorporating human wisdom in complex decision-making, experience-based judgment, and risk control. This enables a more efficient and reliable convergence of high-quality process solutions that meet both engineering constraints and expert intuition, achieving deep complementarity and integration of human and machine intelligence.

[0158] Specifically, in some of the above embodiments, the specific implementation steps of the human-computer interaction and decision visualization module can be further refined. The specific working steps of the human-computer interaction and decision visualization module include: data preparation and dimensionality reduction steps: Latent vector collection: During the design optimization process for a specific product, continuously record all latent vectors z_i generated by the parameter generation module and the optimization module, and their corresponding scheme IDs. Feature association: Associate each z_i with its corresponding cold-extruded part SDF data (or its feature thumbnail), a detailed list of verification results, and a forming quality prediction score. Dimensionality reduction calculation: Periodically or as needed, perform dimensionality reduction on the collected latent vector set {z_i}. For example, using the PCA algorithm: p_i = PCA.transform(z_i), obtain two-dimensional coordinates (u_i, v_i). To maintain the relative stability of the dimensionality reduction results, incremental PCA can be used, or a fixed PCA model pre-trained based on a large amount of historical data can be used. Visual Interface Rendering and Interaction Steps: Main View (Scatter Plot): A scatter plot is drawn in the center of the interface, with the horizontal and vertical axes representing the first and second principal components (or t-SNE coordinates) after dimensionality reduction, respectively. Each point p_i is color-mapped (chroma key) according to its Q_score, and a special icon is marked according to the most serious failure in its verification results (e.g., ▲ for wall thickness issues, ● for volume issues, ◆ for draft angle issues). Related Views: 3D Model View: When the user selects a point, the corresponding 3D model of the cold-extruded part is immediately loaded and rendered in the side window, supporting rotation, scaling, and sectioning. Evaluation Panel: Detailed data of the scheme is displayed synchronously: Q_score value, specific calculated values ​​and thresholds for volume / wall thickness / draft angle, and billet process parameters (diameter, length, etc.). Interactive Toolset: Region Selection Tool: Provides rectangular and lasso selection tools. After selecting a group of points, the user can click the "Set as Preferred Region" button. Directional annotation tool: Users can hold down the Shift key and click on two points A and B. The system will draw an arrow from A to B and pop up a dialog box for the user to input their understanding of this direction (e.g., "direction for increasing wall thickness"). Constraint annotation tool: After selecting a point and viewing its 3D model, users can click on specific parts of the model (using ray picking) and add text annotations such as "the fillet here is too small," or set desired parameter adjustment values ​​through a form. Instruction generation and transmission steps: Instruction encoding: After the interactive operation is triggered, the module encodes the user input into optimization instructions. For region preference: {"type": "region_bias", "points_indices": [idx1, idx2, ...], "action": "sample_around"}. The optimization module will calculate the center or distribution of the latent vectors corresponding to these points and use this distribution as the prior for subsequent sampling.For directional guidance: {"type":"direction_guide", "z_start": z_A, "z_end": z_B, "strength": 0.5}. The optimization module will calculate the direction vector d = (z_B - z_A) and use this direction as a bias term for gradient ascent in subsequent optimizations, or directly explore interpolation along this direction. For local constraints: {"type": "refine_constraint", "z_ref": z_ref, "target_volume_change": +0.05, "note": "enlarge_inner_fillet"}. The optimization module will start with z_ref as the initial point and initiate a multi-objective optimization that combines volume adjustment instructions with possible sensitivity to specific local features based on GNN. Instruction transmission and response: The generated instructions are sent to the optimization module through an internal message queue. After receiving the instructions, the optimization module adjusts its internal state or sampling strategy and immediately begins a new round of optimization iterations. Newly generated solution points are added to the visualized scatter plot in real time, forming dynamic feedback. Users can observe whether the optimization process is proceeding in the expected direction and can make new interventions at any time.

[0159] Secondly, this application provides an automated design method for cold-extruded camera housings based on artificial intelligence, such as... Figure 2 This includes the following steps:

[0160] The system receives a 3D product model of the vehicle-mounted camera housing, performs geometric repair, coordinate alignment, and scale normalization on it, and converts the processed model into a signed distance field with a fixed resolution to obtain the product SDF data.

[0161] The product SDF data is input into a pre-trained conditional variational autoencoder model, which encodes the product SDF data into a latent space. Based on the decoder corresponding to the encoder, latent vectors are sampled from the latent space, and the product SDF data is used as a generation condition input into the decoder to reconstruct the corresponding cold extrusion part SDF data. Simultaneously, through the parameter regression branch in the decoder, a set of billet process parameters associated with the cold extrusion part SDF data is output synchronously.

[0162] The SDF data of the cold-extruded part is analyzed to calculate its volume, minimum wall thickness, and draft angle characteristics, and compared with the preset process constraint thresholds to generate manufacturability verification results. At the same time, the set of process parameters of the blank is input into the performance prediction model to obtain the forming quality prediction score.

[0163] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the method described in the second aspect.

[0164] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described in the second aspect.

[0165] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An automated system for cold extrusion camera housing design based on artificial intelligence, characterized in that, include: The preprocessing module receives the 3D product model of the vehicle-mounted camera housing, performs geometric repair, coordinate alignment and scale normalization, and converts the processed model into a signed distance field with a fixed resolution, outputting the product SDF data. The parameter generation module is used to encode the product SDF data into the latent space based on a pre-trained conditional variational autoencoder model; Based on the latent vectors sampled from the latent space by the decoder corresponding to the conditional variational autoencoder model and the product SDF data, the corresponding cold extrusion part SDF data is reconstructed, and the associated set of billet process parameters is output simultaneously through a parametric regression branch. The verification and prediction module is used to analyze the SDF data of the cold extruded part to determine whether its volume, minimum wall thickness, and draft angle meet the preset process constraint thresholds and generate verification results. It also receives the set of process parameters for the billet and outputs a forming quality prediction score.

2. The automated system for cold extrusion camera housing design based on artificial intelligence driven according to claim 1, characterized in that, The system also includes an optimization module; the optimization module is communicatively connected to the parameter generation module and the verification and prediction module, respectively. The optimization module is configured to: receive the verification result and the forming quality prediction score; when the verification result does not meet the preset standard, generate a feedback signal based on the non-compliance item, and convert the feedback signal into an adjustment instruction for the latent space sampling strategy based on the conditional variational autoencoder model, so as to optimize and generate a new set of billet process parameters that simultaneously meet geometric constraints and high forming quality probability.

3. The automated system for cold extrusion camera housing design based on artificial intelligence driven according to claim 2, characterized in that, The verification and prediction module includes a graph neural network model; The process of receiving the set of process parameters for the billet and outputting a forming quality prediction score includes: The three-dimensional voxel mesh corresponding to the SDF data of the cold-extruded part is transformed into a graph structure; wherein each voxel is a node of the graph, and the node features include the SDF value, coordinate position and local gradient of the voxel; the set of billet process parameters is fused with the features of each node as global graph features; After message passing and node feature updating through a multi-layer graph convolutional network in the graph neural network model, global information is aggregated through a graph pooling layer, and finally the forming quality prediction score is obtained through regression by a fully connected layer.

4. The automated system for cold extrusion camera housing design based on artificial intelligence driven according to claim 3, characterized in that, The decoder corresponding to the conditional variational autoencoder model adopts a multi-task learning architecture, which includes an SDF reconstruction branch and a parametric regression branch. The SDF reconstruction branch is used to predict signed distance values ​​on a voxel-by-voxel basis to form the SDF of the cold extruded part. The parametric regression branch shares hidden layer features with the SDF reconstruction branch in the intermediate layer of the decoding process, and uses an independent regression head network to map the shared features to the billet process parameter set. The billet process parameter set includes at least the billet diameter, length, volume compensation coefficient, and material identification code.

5. The automated system for cold extrusion camera housing design based on artificial intelligence driven according to claim 4, characterized in that, The optimization module is configured to perform the following specific optimization process: If the verification result shows that the volume does not match, the ratio of the current cold extrusion volume to the target volume is calculated, and vector interpolation or perturbation sampling is performed in the potential space of the parameter generation module along the direction that makes the ratio approach 1 to generate a new set of billet process parameters. If the forming quality prediction score is lower than a preset threshold, the gradient of the forming quality prediction score with respect to the potential vector is calculated using the differentiable property of the prediction model in the verification and prediction module. Then, an iterative search is performed in the potential space along the gradient ascent direction to find potential vectors that can generate higher prediction scores, thereby generating an optimized set of billet process parameters.

6. The automated system for cold extrusion camera housing design based on artificial intelligence as described in claim 1, characterized in that, The scale normalization process includes: Calculate the maximum outer envelope size of the three-dimensional product model in the three-dimensional space along the three coordinate axes, and use the length of the longest side in the maximum outer envelope size as the normalization reference. Scale the coordinates of all vertices of the three-dimensional product model proportionally so that the length of the longest side of the scaled model is equal to the preset reference length value.

7. The automated system for cold extrusion camera housing design based on artificial intelligence driven according to any one of claims 1-6, characterized in that, The system also includes a human-computer interaction and decision visualization module, which is communicatively connected to the parameter generation module, the verification and prediction module, and the optimization module, and is configured to execute: The latent space based on the conditional variational autoencoder model in the parameter generation module is dimensionality-reduced and mapped. In the visualization interface, the three-dimensional models of cold extruded parts generated by different latent points, the corresponding geometric constraint verification results output by the verification and prediction module, and the forming quality prediction score are displayed in association. The system receives user selection and annotation operations for specific regions or directions in the potential space from the visualization interface, generates corresponding interactive instructions, and provides these interactive instructions as initial sampling preferences or constraints to the optimization module.

8. An automated design method for cold-extruded camera housings based on artificial intelligence, characterized in that, Includes the following steps: The system receives a 3D product model of the vehicle-mounted camera housing, performs geometric repair, coordinate alignment, and scale normalization on it, and converts the processed model into a signed distance field with a fixed resolution to obtain the product SDF data. The product SDF data is input into a pre-trained conditional variational autoencoder model, which encodes the product SDF data into the latent space. Based on the decoder corresponding to the encoder, latent vectors are sampled from the latent space, and the product SDF data is input to the decoder as a generation condition to reconstruct the corresponding cold extrusion SDF data; at the same time, through the parameter regression branch in the decoder, the set of billet process parameters associated with the cold extrusion SDF data is output synchronously. The SDF data of the cold-extruded part is analyzed to calculate its volume, minimum wall thickness, and draft angle characteristics, and compared with the preset process constraint thresholds to generate manufacturability verification results. At the same time, the set of process parameters of the blank is input into the performance prediction model to obtain the forming quality prediction score.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 8.

Images