Spinning forming automatic modeling simulation method and system driven by large language model

Abstract

This invention discloses an automated modeling and simulation method and system for spinning forming driven by a large language model, specifically relating to the field of intelligent manufacturing. The method includes: acquiring natural language commands for spinning simulation; parsing the natural language commands using a pre-trained large language model to obtain structured parameters; modeling based on geometric modeling parameters to obtain a sheet metal model; dividing the sheet metal model and determining node coordinates; determining adjacent radial layers and adjacent sectors based on the node coordinates, and constructing four-node structured elements to obtain a finite element model; determining the trajectory point sequence based on trajectory parameters and determining the feed rate for each spinning path segment; determining the feed time for each spinning path segment based on the feed rate and the average linear velocity of two adjacent spinning paths, and determining the cumulative time; performing amplitude conversion; and performing spinning forming on the finite element model based on the converted amplitude and cumulative time to obtain the workpiece.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and in particular to an automated modeling and simulation method and system for spinning driven by a large language model. Background Technology

[0002] The spinning process of aluminum alloy curved parts involves complex nonlinear large deformation physical behavior. Although finite element simulation technology has become the mainstream method to support process parameter optimization and forming mechanism analysis, its practical application still faces three significant bottlenecks: First, the modeling process relies heavily on manual experience and is cumbersome, resulting in low modeling efficiency and poor reusability; second, the automated extraction of key forming indicators such as wrinkling and wall thickness distribution in the post-processing stage often requires writing secondary development scripts, which presents a high technical barrier; more importantly, the existing simulation system lacks an "intelligent brain" that can understand user natural language commands and drive the entire process in a coordinated manner, forcing process engineers to manually switch between multiple isolated software modules such as modeling, simulation, and post-processing, making it impossible to achieve end-to-end automation of "command as result".

[0003] Currently, automated modeling, simulation, and post-processing technologies for spinning curved parts are still in their infancy. However, in other similar forming fields, scholars both domestically and internationally have achieved certain results based on parametric programming. For example, a parametric modeling system for shaft parts has enabled rapid modeling and annotation of specific components; and an analysis system for multi-pass roll bending of asymmetric Z-section profiles has been developed based on ABAQUS, achieving automation and parametric processing of pre- and post-processing, and improving the analysis efficiency of multi-pass roll bending experiments. However, the aforementioned existing technologies have significant limitations when applied to spinning scenarios: most of these methods are "hard-coded" systems developed for specific simple geometries (such as shafts and Z-shaped profiles), and their underlying logic relies on a fixed program framework, lacking the versatility and flexibility to handle the varied surface features and complex process paths in spinning forming; in addition, existing research has not yet formed an automated closed loop covering the entire "modeling-simulation-post-processing" chain, and lacks in-depth knowledge mining and intelligent feedback capabilities when faced with massive simulation data; more importantly, existing automation methods still remain in the traditional interaction mode of "GUI interface + manual input of all parameters", which cannot understand abstract process intentions and makes it difficult to directly obtain "generalized natural language instructions - final forming results". Summary of the Invention

[0004] The main purpose of this application is to provide a large language model-driven automated modeling and simulation method and system for spinning forming, which aims to solve the problem that existing automated modeling and simulation technologies cannot be used in the spinning forming process.

[0005] To achieve the above objectives, this application provides a large language model-driven automated modeling and simulation method for spinning forming, comprising: acquiring natural language commands for spinning simulation; parsing the natural language commands using a pre-trained large language model to obtain structured parameters; wherein, the structured parameters include geometric modeling parameters, mesh generation parameters, and process parameters; the process parameters include feed ratio, mandrel rotation speed, and trajectory parameters; modeling based on the geometric modeling parameters to obtain a sheet metal model; acquiring the inner and outer radii of the sheet metal model, and setting the number of radial layers and circumferential segments based on the mesh generation parameters; dividing the sheet metal model according to the number of radial layers and circumferential segments, obtaining multiple sectors in each radial layer; and dividing the sheet metal model according to the number of radial layers, inner radius, and outer radius of the sheet metal model. The node coordinates are determined by the sequence numbers of the radial and radial layers and the sector numbers. Adjacent radial layers and sectors are determined based on the node coordinates, and four-node structured elements are constructed between adjacent radial layers and sectors to obtain the finite element model. The trajectory point sequence is determined based on the trajectory parameters, and the trajectory point coordinates of each spinning path are determined. The feed rate of each spinning path is calculated based on the difference in trajectory point coordinates between two adjacent spinning paths. The instantaneous linear velocity is determined based on the feed ratio and the mandrel rotation speed. The feed time of each spinning path is determined based on the feed rate and the average linear velocity of two adjacent spinning paths, and the cumulative time is also determined. The mandrel rotation speed and feed rate are converted to amplitudes, and spinning is performed on the finite element model based on the converted amplitudes and cumulative time to obtain the workpiece.

[0006] Optionally, the training process of the pre-trained large language model is as follows: Based on knowledge related to spinning forming, a mapping knowledge base of "natural language - computer language" is constructed; wherein, the mapping knowledge base includes multiple core terminology groups; each core terminology group includes a natural language instruction and the corresponding structured parameters; based on the mapping knowledge base, multiple natural language instruction-structured parameter pairs are constructed; multiple real instruction and parameter pairs are extracted from historical simulation projects; the multiple natural language instruction-structured parameter pairs and the multiple real instruction and parameter pairs are combined into a training set, and the training set is input into the large language model for training. During the training process, the LoRA strategy is used to fine-tune the parameters of the large language model to obtain the pre-trained large language model.

[0007] Optionally, the mandrel rotation speed and feed rate are converted into amplitudes, including: converting the mandrel rotation speed into the mandrel angular velocity amplitude; determining the half-cone angle of each spinning path based on the coordinates of the trajectory points of two adjacent spinning paths; performing a rotational mapping on the feed rate based on the half-cone angle and the installation angle of each spinning path to obtain the displacement amplitude of each spinning path; and summing the displacement amplitudes of all spinning paths to obtain the cumulative displacement amplitude.

[0008] Optionally, before dividing the sheet metal model according to the number of radial layers and the number of circumferential segments, the method further includes: adjusting the number of radial layers in the internal region and the flange region of the sheet metal model in contact with the core mold according to a radial progressive densification strategy.

[0009] Optionally, the node coordinates are determined based on the radial layer number, inner radius, outer radius, radial layer number, and sector number of the sheet metal model, including: determining the radial position of each radial layer based on the radial layer number, inner radius, outer radius, and radial layer number of the sheet metal model; determining the circumferential angle of each radial layer based on the number of circumferential segments and sector number; and determining the node coordinates based on the product of the radial position and the circumferential angle.

[0010] Optionally, adjacent radial layers and adjacent sectors are determined based on node coordinates, and four-node structured units are constructed between adjacent radial layers and adjacent sectors to obtain a finite element model. This includes: determining node index logic based on the radial layer number, sector number, and circumferential direction angle; and constructing four-node structured units between adjacent radial layers and adjacent sectors based on the node index logic to obtain a finite element model.

[0011] Optionally, the structural parameters also include the thinning rate and the degree of flange wrinkling; the degree of flange wrinkling is determined based on the maximum amplitude, average peak-to-valley amplitude, variance, and full wavenumber; the maximum amplitude is used to capture extreme deviations in wrinkling; the average peak-to-valley amplitude and variance are used to characterize the typical amplitude and dispersion of the fluctuation; and the full wavenumber is used to quantify the number of wrinkling cycles and the degree of flange wrinkling.

[0012] Optionally, the process for determining the maximum amplitude is as follows: obtain the axial coordinate sequence of the edge nodes of the workpiece, determine the fluctuation variance based on the axial coordinate sequence of the edge nodes of the workpiece, and determine the maximum amplitude based on the maximum and minimum values ​​in the axial coordinate sequence.

[0013] Optionally, the process for determining the average peak-valley amplitude is as follows: A two-point moving average is performed on the axial coordinate sequence of the edge nodes of the workpiece to obtain a smoothed sequence, and the mean of the smoothed sequence is determined based on the smoothed sequence; For each pair of adjacent smoothed sequences, a cross-mean detection is performed, and when the detection result is a cross-mean, a "zero-crossing" index is recorded; Based on the "zero-crossing" index, all smoothed sequences are divided into several peak-valley candidate intervals, and the set of local maxima and local minima within each peak-valley candidate interval is determined; Based on the set of local maxima and local minima of all peak-valley candidate intervals, the average peak-valley amplitude is determined; The complete wavenumber is determined based on the number of "zero-crossing" indices.

[0014] To achieve the above objectives, this application also provides a large language model-driven automated modeling and simulation system for spinning forming, comprising: a language instruction parsing module for acquiring natural language instructions for spinning simulation, parsing the natural language instructions using a pre-trained large language model to obtain structured parameters; wherein, the structured parameters include geometric modeling parameters, mesh generation parameters, and process parameters; the process parameters include feed ratio, mandrel rotation speed, and trajectory parameters; a modeling module for modeling based on the geometric modeling parameters to obtain a sheet metal model; a mesh generation module for acquiring the inner and outer radii of the sheet metal model, and setting the number of radial layers and circumferential segments based on the mesh generation parameters; dividing the sheet metal model according to the number of radial layers and circumferential segments, obtaining multiple sectors in each radial layer; and dividing the sheet metal model according to the number of radial layers and inner radius. The node coordinates are determined by the outer radius, the sequence number of the radial layer, and the sequence number of the sector. Adjacent radial layers and sectors are determined based on the node coordinates, and four-node structured elements are constructed between adjacent radial layers and sectors to obtain the finite element model. The trajectory point feed module determines the trajectory point sequence based on trajectory parameters and the coordinates of the trajectory points for each spinning path. The feed amount for each spinning path is calculated based on the difference in coordinates between the trajectory points of two adjacent spinning paths. The instantaneous linear velocity is determined based on the feed ratio and the mandrel rotation speed. The feed time for each spinning path is determined based on the feed amount and the average linear velocity of two adjacent spinning paths, and the cumulative time is also determined. The amplitude conversion and spinning module performs amplitude conversion on the mandrel rotation speed and feed amount. Based on the converted amplitude and cumulative time, spinning is performed on the finite element model to obtain the workpiece.

[0015] Compared with the prior art, the beneficial effects of this application are as follows:

[0016] This invention presents a large language model-driven automated modeling and simulation method for spinning forming. Through a pre-established mapping knowledge base, it parses natural language commands into structured parameters for spinning simulation based on a pre-trained large language model. Modeling and meshing are performed using these structured parameters to determine the mandrel rotation speed and feed rate controlling the spinning forming process. This achieves automated modeling and simulation from natural language commands to the spinning forming process, improving spinning forming quality and R&D efficiency. The method determines the number of radial layers and circumferential segments using the inner and outer radii of the sheet metal model and meshing parameters. Based on these parameters, it achieves adaptive radial partitioning from the inside out, with multiple layers and ensuring boundary continuity in the circumferential direction. Between adjacent radial layers and adjacent sectors, a four-node structured unit is constructed using node index logic to maintain the topological integrity of the entire mesh. The method dynamically calculates the feed rate and feed time for each spinning path segment based on the feed ratio and mandrel rotation speed, achieving dynamic adaptation of discrete point trajectories and ensuring feed accuracy, thereby improving simulation accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the automated modeling and simulation method for spinning driven by a large language model, as described in this application.

[0018] Figure 2 This is a state diagram of flange fluctuation in the large language model-driven automated modeling and simulation method for spinning in this application.

[0019] Figure 3 This is an interface diagram of the language instruction parsing module in the large language model-driven automated modeling and simulation system for spinning forming in this application;

[0020] Figure 4 This is an interface diagram of the automated modeling and simulation module in the large language model-driven automated modeling and simulation system for spinning in this application.

[0021] Figure 5 This is another interface diagram of the automated modeling and simulation module in the large language model-driven automated modeling and simulation system for spinning in this application;

[0022] Figure 6 This is an interface diagram of the automated post-processing module in the large language model-driven automated modeling and simulation system for spinning forming in this application.

[0023] Figure 7 The image shows a finite element model obtained from an embodiment of the large language model-driven automated modeling and simulation method for spinning forming in this application.

[0024] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] The first embodiment of the present invention provides an automated modeling and simulation method for spinning driven by a large language model, specifically including the following steps:

[0027] Step S1: Obtain the natural language instructions for the spinning simulation. Parse the natural language instructions using a pre-trained large language model to obtain structured parameters. The structured parameters include geometric modeling parameters, mesh generation parameters (such as mesh density and fine-grained areas), and load parameters. The geometric modeling parameters include sheet dimensions (such as sheet diameter and thickness), mandrel profile, material properties (such as material grade and constitutive model parameters), and contact parameters (such as friction coefficient). The load parameters include feed ratio, mandrel rotation speed, and trajectory parameters.

[0028] Furthermore, the training process of the pre-trained large language model is as follows:

[0029] Based on knowledge related to spinning forming, a mapping knowledge base of "natural language - computer language" is constructed. This knowledge base includes multiple core terminology groups; each core terminology group includes a natural language instruction and its corresponding structured parameters. Based on the mapping knowledge base, multiple natural language instruction-structured parameter pairs are constructed, and multiple real instruction and parameter pairs are extracted from historical simulation projects. These natural language instruction-structured parameter pairs and the real instruction and parameter pairs are combined to form a training set. This training set is then input into a large language model for training. During training, the LoRA strategy is used to fine-tune the parameters of the large language model, resulting in a pre-trained large language model. To ensure that the training set contains as much spinning forming-related knowledge as possible, this embodiment also expands the training set using data augmentation methods such as synonym replacement and word order adjustment.

[0030] Specifically, the mapping knowledge base of "natural language-computer language" adopts the construction method of "annotation-semi-automatic alignment-synonym expansion". Specifically, each natural language phrase (such as "feed speed faster") is annotated with its corresponding computer-executable parameters (structured parameters) (such as feed ratio = 1.5 mm / r, feed speed = 2 mm / s) to form an initial mapping seed set; using a thesaurus and rule-based template induction, the seed set is expanded to cover different colloquial expressions with the same semantics as each natural language phrase, forming the mapping knowledge base. In constructing the mapping knowledge base, this embodiment relies on the following data sources: professional literature, technical manuals, and standards in the field of spinning forming, such as "Spinning Forming Technology" and "High-Intensity Spinning Technology," from which the definitions, physical meanings, and typical value ranges of core process parameters are extracted; patent literature and academic papers related to spinning processes, summarizing commonly used terminology and their corresponding computer-resolvable parameters; the experience and knowledge of senior spinning process experts, collecting empirical correspondences between qualitative process descriptions and quantitative modeling parameters; and a historical spinning simulation project case library, including simulation input files (.inp) and corresponding natural language task descriptions, forming positive and negative examples.

[0031] The LoRA strategy freezes the backbone weights of the large language model and adds parallel low-rank trainable branches (rank r=8, α=16) only to the Q and V matrices of the self-attention layer. For example, the large language model can be Qwen, Llama 3.2, GLM-4, etc., as long as it possesses basic semantic understanding and instruction execution capabilities. In this embodiment, Qwen-7B is used as the base model, with 7 billion parameters. While ensuring strong semantic understanding capabilities, it has moderate computational resource requirements, facilitating deployment and fine-tuning on ordinary servers. A training set of 3500 parameter pairs is used, with 5 training epochs, a batch size of 4, a learning rate of 2e-4, and the AdamW optimizer. Evaluation on the test set shows that the model's generated JSON parameters achieve an accuracy of over 90% perfect matching with the standard answer, and its semantic understanding accuracy is 95%. The above fine-tuning strategy enables the large language model to accurately grasp the physical meaning and parameter association logic of spinning-specific terms, and strictly follow the preset data output format specifications, thereby ensuring that it can stably and accurately generate structured configuration information to drive the operation of automated programs.

[0032] In this embodiment, a mapping dataset covering spinning process rules and material constitutive properties is constructed through a pre-trained large language model. This establishes the conversion logic from "qualitative process requirements" to "quantitative modeling parameters," enabling the model to understand the user's design intent. Based on the built-in knowledge of the pre-trained large language model, it automatically infers and completes key parameters such as missing geometric dimensions, mesh density, and boundary conditions, thereby establishing a logical bridge between natural language and simulation parameters.

[0033] Step S2: Model the sheet metal according to the geometric modeling parameters to obtain the sheet metal model;

[0034] Step S3: Obtain the inner and outer radii of the sheet metal model, and set the radial layer number and circumferential segment number based on the mesh generation parameters; wherein, the mesh generation parameters include mesh density;

[0035] Specifically, the radial layer number and circumferential segment number are preset parameters based on simulation accuracy requirements, rather than derived values ​​obtained through calculation. The radial layer number controls the radial mesh density, and the circumferential segment number controls the circumferential mesh density. Therefore, setting the radial layer number and circumferential segment number is based on the mesh density to obtain the final mesh generation parameters. When the simulation accuracy requirement increases, the radial layer number and circumferential segment number are increased accordingly to achieve mesh refinement. Considering that the flange region (i.e., the region where the radius of the sheet metal model exceeds the radius of the mandrel) is the main location for wrinkling and other defects during spinning, and its fluctuation behavior needs to be accurately captured, a denser radial layer division is used for this region. The internal region of the sheet metal model (i.e., the overlapping region between the sheet metal model and the mandrel) has relatively stable deformation, and a sparser radial layer can be used to save computational resources. That is, a radial progressive refinement strategy is adopted to adjust the radial layer number in the internal region and the flange region of the sheet metal model. This is specifically achieved through a segmented density function: uniformly dividing the internal region with a smaller number of radial layers, i.e., the internal radial layer number... The outer flange region is evenly divided into a large number of radial layers, i.e., the number of outer radial layers. The two layers are naturally connected by a transition layer. This strategy significantly reduces the overall mesh size and improves solution efficiency while ensuring the computational accuracy of the flange region.

[0036] Step S4: Divide the sheet metal model according to the number of radial layers and the number of circumferential segments, and obtain multiple sectors in each radial layer;

[0037] Specifically, the sheet metal model is divided based on the number of circumferential segments and the adjusted number of radial layers, resulting in multiple sectors in each radial layer. Furthermore, the meshing parameters can include a densification region, which users can adjust via natural language commands, such as "double the mesh density in the flange area." The pre-trained large language model parses this command and automatically adjusts the mesh on the sheet metal model accordingly. and The proportion.

[0038] Step S5: Determine the node coordinates based on the radial layer number, inner radius, outer radius, radial layer number, and sector number of the sheet metal model;

[0039] Specifically, in step S51, the radial position of each radial layer is determined based on the radial layer number, inner radius, outer radius, and radial layer sequence number of the sheet metal model. The calculation formula is:

[0040] ( );

[0041] In the formula, Let the inner radius be , The outer radius is The radial layer number, The number is the radial layer number, which can be directly obtained from the partitioned model.

[0042] Step S52: Determine the circumferential direction angle of each radial layer based on the number of circumferential segments and the sector number. The calculation formula is:

[0043] ( );

[0044] In the formula, The sector number can be directly obtained from the partitioned model; This represents the number of segments in the circumference.

[0045] Step S53: Determine the node coordinates based on the product of the radial position and the circumferential angle. .

[0046] Step S6: Determine the adjacent radial layers and adjacent sectors based on the node coordinates, and construct four-node structured elements between the adjacent radial layers and adjacent sectors to obtain the finite element model;

[0047] Specifically, the node index logic is determined based on the radial layer number, sector number, and circumferential angle: ,in , ;

[0048] Based on the node index logic in adjacent radial layers ( and ) and adjacent sectors ( and Between these elements, four-node structured elements are constructed to obtain the finite element model. At this point, the finite element model consists of multiple four-node structured elements, each element... The corresponding four node indices are:

[0049] Node 1: ;

[0050] Node 2: ;

[0051] Node 3: ;

[0052] Node 4: ;

[0053] The above node indexing logic ensures the periodic closure in the circumferential direction and automatically generates all quadrilateral units without manual intervention.

[0054] Step S7: Determine the trajectory point sequence based on the trajectory parameters, and determine the trajectory point coordinates for each spinning path segment; wherein, the trajectory parameters may include trajectory type, surface geometric parameters, or trajectory point coordinates;

[0055] Step S71: When the trajectory parameter is a trajectory type, i.e., the natural language command describes the trajectory type, and the trajectory type is a typical curved surface (such as a cone, hemisphere, parabola, etc.), since the simulation system pre-builds a trajectory library initially, which includes trajectory types and corresponding trajectory point sequences, the system will automatically generate the corresponding trajectory point coordinates based on the trajectory type during use. In this case, the trajectory data can be obtained by acquiring the trajectory point sequence corresponding to the trajectory type. If the trajectory type is a non-typical curved surface and includes the geometric parameters of the surface, the trajectory point coordinates are automatically calculated based on the surface equation. The method of automatically calculating trajectory point coordinates based on the surface equation is a conventional method of trajectory point calculation and is not an improvement of this application; as long as the trajectory point coordinate calculation can be achieved, it is acceptable, and will not be elaborated further here. If the trajectory parameter is trajectory point coordinates, such as "trajectory points from (0, 0) to (100, 50) and then to (200, 30)," it is automatically parsed into a trajectory point sequence.

[0056] Step S72: Divide the trajectory point sequence into multiple spinning paths and obtain the trajectory point coordinates of each spinning path.

[0057] Step S8: Determine the feed rate for each spinning path segment based on the difference in coordinates of the trajectory points between two adjacent spinning paths; i Feed rate of the spinning path The expression is:

[0058] ;

[0059] In the formula, , These are the coordinates of the trajectory points.

[0060] Specifically, let the trajectory point number be... The trajectory segment number is , of which Segment connection point With point The expression for instantaneous linear velocity is:

[0061] ;

[0062] Feed time The expression is:

[0063] ;

[0064] In the formula, For the first The feed rate of the segment trajectory. This represents the total number of trajectory points.

[0065] Cumulative Time The expression is:

[0066] ;

[0067] In the formula, To reach the The cumulative time for each trajectory point This is the initial delay time. This refers to the initial settling time required for the mandrel to accelerate from a standstill to a stable rotational speed. Physically, it ensures the mandrel reaches its set rotational speed, achieves stable contact, and avoids initial impact. Based on spinning process experience, The default value is 5 seconds; users can also adjust it via natural language commands (such as "set initial delay to 3 seconds"). If not specified, the default value of 5 seconds will be used.

[0068] Step S10: The rotational speed and feed rate of the mandrel are converted into amplitudes. Based on the corresponding amplitudes and cumulative time after conversion, spinning is performed on the finite element model to obtain the workpiece.

[0069] Specifically, in step S101, the core mold rotation speed is adjusted. n Convert to core mold angular velocity amplitude Its expression is:

[0070] ;

[0071] Step S102: Based on the coordinates of the trajectory points of two adjacent spinning paths, determine the half-cone angle of each spinning path. i Segment half cone angle The expression is:

[0072] ;

[0073] Step S103, based on the half-cone angle and mounting angle of each spinning path segment By rotating the feed rate, the displacement amplitude of each spinning path segment is obtained. );

[0074] ;

[0075] ;

[0076] Step S104: Sum the displacement amplitudes of all spinning paths to obtain the cumulative displacement amplitude. );

[0077] ;

[0078] ;

[0079] Step S105: Based on the angular velocity amplitude of the mandrel, the displacement amplitude of each spinning path, the cumulative displacement amplitude, and the cumulative time, spin forming is performed on the finite element model to obtain the workpiece.

[0080] Furthermore, to ensure the accuracy of the simulated workpiece, the structured parameters also include the thinning rate and the degree of flange wrinkling. After obtaining the workpiece, the thinning rate and the degree of flange wrinkling are automatically determined. Among them, the maximum thinning rate is determined based on the wall thickness distribution; the degree of flange wrinkling is determined based on the maximum amplitude, average peak-to-valley amplitude, variance, and full wavenumber; the maximum amplitude is used to capture extreme deviations in wrinkling; the average peak-to-valley amplitude and variance are used to characterize the typical amplitude and dispersion of the fluctuation; and the full wavenumber is used to quantify the number of wrinkling cycles and the degree of flange wrinkling.

[0081] For a specific schematic diagram of flange undulation, see [link to diagram]. Figure 2 , Figure 2 In the image, 'a' represents the point cloud diagram of the workpiece. The extracted node changes of the flange are shown in the figure. Figure 2 b. From Figure 2 From this, we can conclude that flange wrinkling manifests as axial periodic fluctuations at the edge nodes, which can be quantified by extracting feature points on the outermost contour. The edge nodes of the workpiece refer to the outermost nodes in the finite element model of the workpiece, along the radial direction. The calculation method for the degree of flange wrinkling is as follows:

[0082] Obtain the axial coordinate sequence of the edge nodes of the workpiece The fluctuation variance is determined based on the axial coordinate sequence. The maximum amplitude is determined based on the maximum and minimum values ​​in the axial coordinate sequence. :

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, The average value of the axial coordinate sequence. The node number;

[0087] To suppress high-frequency noise while preserving edge ripple details, this invention proposes a three-stage filtering and quantization method: "smoothing-zero crossover-local extrema." Specifically, a two-point moving average is performed on the axial coordinate sequence of the workpiece's edge nodes to obtain a smoothed sequence. :

[0088] ;

[0089] And determine the mean of the smoothed sequence based on the smoothed sequence. ;

[0090] ;

[0091] For each pair of adjacent smoothed sequences, a cross-mean detection is performed sequentially, meaning the differences between the two smoothed sequences and the mean have opposite signs. The specific detection formula is as follows:

[0092] ;

[0093] When the detection result crosses the mean (i.e., the product is less than zero), a "zero-crossing" index is recorded. Based on the "zero-crossing" index, all smoothed sequences are divided into several peak-valley candidate intervals, and the set of local maxima {M} within each peak-valley candidate interval is determined. p} and the set of local minima {m p After excluding incomplete intervals at the beginning and end, the average peak-valley amplitude is determined based on the set of local maxima and local minima of all candidate peak-valley intervals. :

[0094] ;

[0095] In the formula, J To ensure the complete number of peak-valley pairs, this method uses a two-layer filtering approach (moving average and interval extreme values) to accurately extract the flange fluctuation amplitude, combining the advantages of noise suppression and detail preservation.

[0096] Traditional peak counting is susceptible to misjudgment due to local fluctuations. This invention introduces a period division strategy based on zero crossovers to ensure the accuracy and repeatability of wavenumber statistics. The complete wavenumber is determined based on the number of zero crossover indices.

[0097] Specifically, let L be the total number of zero cross-indexes. If L < 2, it is considered that a complete peak-valley-peak or valley-peak-valley cycle has not been formed, and a unified output of the complete wave number is given. If it is 1; otherwise, every two zero cross-indexes correspond to a complete round-trip cycle, therefore the definition is: When L is odd, the tail segment is incomplete, but this method still calculates the number of complete cycles by rounding down, and ensures that the results are stable under a small amount of jitter through the aforementioned degradation judgment.

[0098] It should be noted that spinning includes various process types such as ordinary spinning, shear spinning, and high-strength spinning. The forming mechanisms of wrinkling and wall thickness reduction differ under different process types, and their quality judgment standards also vary significantly. In addition, the material type (such as aluminum alloy, titanium alloy, high-strength steel), the purpose of the part, and the industry standards applied will also affect the selection of the judgment threshold. Therefore, this embodiment does not preset a single fixed judgment threshold.

[0099] Furthermore, the automated modeling and simulation method of this invention supports multi-condition parametric analysis and has a complete batch simulation organization and scheduling mechanism. For large language models, the processing of multi-conditions is the same as that of single-conditions, both parsing natural language instructions into structured parameters; for the modeling and simulation process, general simulation software also supports multi-conditions, and this embodiment does not improve the condition processing process. This embodiment is specifically illustrated through the following examples.

[0100] Natural language commands include, for example: Parameter range scan: "Sheet diameter from 200mm to 300mm, step size 20mm", which automatically generates 6 sets of diameter values ​​after parsing by a pre-trained large language model; Orthogonal experiment: "Generate 9 sets of working conditions according to the L9 (3^4) orthogonal array", which combines parameters such as feed ratio, rotational speed, and thickness after parsing; Combination list: "Simulate for thicknesses of 2mm, 3mm, and 4mm respectively", which generates three sets of tasks. See Table 1 for details.

[0101] Table 1. Examples of mapping natural language commands to batch experiment parameters

[0102]

[0103] The management mechanism during the modeling and simulation process, namely task scheduling and management, is as follows:

[0104] Task queue: A FIFO queue is used to manage multiple groups of tasks, which are submitted to the solver in sequence to avoid resource conflicts.

[0105] Parallel computing: Users can specify the maximum number of parallel tasks (e.g., running 3 simulations simultaneously). The system dynamically allocates resources based on the number of available CPU cores and submits multiple inp files to the Abaqus solver in parallel.

[0106] Status monitoring: The GUI interface displays the status of each task in real time (waiting, running, completed, failed), and users can pause, cancel, or retry any task.

[0107] Results archiving: The output files of each task (.odb, .dat, .msg, etc.) are named according to "task ID_parameter combination" and automatically stored in the specified folder for easy batch post-processing.

[0108] Through the above mechanism, the present invention achieves flexible expansion from single working condition to multiple working conditions, and greatly improves the efficiency and automation level of spinning simulation analysis.

[0109] In this embodiment, a pre-established mapping knowledge base enables the pre-trained large language model to parse natural language instructions into structured parameters for spinning simulation. Based on these structured parameters, modeling and mesh generation determine the core mold rotation speed and feed rate controlling the spinning process, achieving full-process intelligentization from natural language instructions to automated modeling and simulation, thus improving forming quality and R&D efficiency. The number of radial layers and circumferential segments is determined by the inner and outer radii of the sheet metal model and the mesh generation parameters. Based on these parameters, adaptive radial partitioning from the inside out and across multiple levels is achieved, ensuring boundary continuity in the circumferential direction. Between adjacent radial layers and adjacent sectors, a four-node structured unit is constructed using node index logic, maintaining the topological integrity of the entire mesh. The feed rate and feed time are dynamically calculated for each spinning path segment based on the feed ratio and core mold rotation speed, achieving dynamic adaptation of discrete point trajectories and ensuring feed accuracy, thereby improving simulation accuracy.

[0110] The second embodiment of the present invention provides an automated modeling and simulation system for spinning driven by a large language model, including a language instruction parsing module, a modeling module, a mesh generation module, a trajectory point feed module, and an amplitude conversion and spinning module; the language instruction parsing module is used to acquire natural language instructions for spinning simulation, and parses the natural language instructions through a pre-trained large language model to obtain structured parameters; wherein, the structured parameters include geometric modeling parameters, mesh generation parameters, and load parameters; the load parameters include feed ratio, mandrel rotation speed, and trajectory parameters.

[0111] The modeling module is used to create a sheet metal model based on geometric modeling parameters.

[0112] The mesh generation module is used to obtain the inner and outer radii of the sheet metal model and, combined with mesh generation parameters, couples the number of radial layers and the number of circumferential segments. The sheet metal model is then divided according to the number of radial layers and circumferential segments, resulting in multiple sectors in each radial layer. Node coordinates are determined based on the number of radial layers, inner and outer radii, the radial layer number, and the sector number. Adjacent radial layers and sectors are determined based on the node coordinates, and four-node structured elements are constructed between adjacent radial layers and sectors to obtain the finite element model.

[0113] The trajectory point feed module is used to determine the trajectory point sequence based on trajectory parameters, determine the trajectory point coordinates of each spinning path segment; determine the feed amount of each spinning path segment based on the difference in trajectory point coordinates between two adjacent spinning paths segment; determine the instantaneous linear velocity based on the feed ratio and the mandrel rotation speed; and determine the feed time of each spinning path segment based on the feed amount and the average linear velocity of two adjacent spinning paths segment, and determine the cumulative time.

[0114] The amplitude conversion and spinning module is used to convert the core mold rotation speed and feed rate into amplitudes. Based on the corresponding amplitude and cumulative time after conversion, spinning is performed on the finite element model to obtain the workpiece.

[0115] It also includes an automated post-processing module for determining the workpiece thinning rate and the degree of flange wrinkling.

[0116] In this embodiment, the modeling and simulation system adopts a layered architecture design based on the existing Abaqus finite element software: the bottom layer consists of pre-packaged automated modeling, simulation, and post-processing programs, while the top layer is a pre-trained large language model, specifically a large language model for spinning that has been fine-tuned using knowledge from the spinning domain. This invention innovatively constructs a driving mechanism of "natural language intent recognition - parameter mapping - script interface call," utilizing the pre-trained large language model as an intelligent scheduling hub to parse fuzzy user commands and automatically configure parameters, thereby accurately calling the underlying scripts and achieving full-process automation from process design to result analysis.

[0117] Specifically, the system can be divided into three core functional modules:

[0118] The intelligent scheduling module for the spinning-specific large language model, also known as the language instruction parsing module, transforms unstructured requirements into a structured set of standard parameters and automatically generates calling instructions for the underlying script library based on the task type. The pre-trained large language model obtained in this embodiment, i.e., the spinning LLM assistant, has a UI diagram shown below. Figure 3 .

[0119] The automated modeling and simulation module includes a modeling module, a mesh generation module, and a trajectory point feeding module. It receives a structured parameter set from a pre-trained large language model, drives a script to generate a standard keyword input file (.inp), and automatically submits it to the solver for computation, requiring no manual intervention. The interface of the automated modeling and simulation module can be found here. Figure 4-5 .

[0120] The automated forming information extraction module, also known as the automated post-processing module, responds to the call commands of the pre-trained large language model, automatically accesses the result database, executes the extraction program for key forming information (such as calculating the wrinkling degree based on the peak-to-trough difference and mapping the overall wall thickness distribution), and feeds back the extracted key indicator data to the front end. The interface of the automated post-processing module is shown below. Figure 6 .

[0121] Example

[0122] The natural language command is: Perform spinning simulation on 2219 aluminum alloy sheet with a diameter of 300mm and a thickness of 2mm, using a hemispherical trajectory, a feed ratio of 1.2mm / r, a rotation speed of 60rpm / min, and a finer mesh. Analyze the wrinkling situation in batches when the sheet thickness is 1.8mm, 2.0mm, and 2.2mm.

[0123] Step 1: Obtain the natural language commands for the spinning simulation. Parse the commands using a pre-trained large language model to obtain structured parameters: sheet diameter 300mm, base thickness 2mm, material 2219 aluminum alloy (automatically matched Johnson-Cook constitutive model parameters), trajectory type hemispherical, feed ratio 1.2mm / r, rotational speed 200rpm, mesh refinement (automatically set to twice the refinement in the contact area), batch parameters are thicknesses of 1.8mm, 2.0mm, and 2.2mm, and the evaluation standard is HB 5800. For parameters not explicitly input, such as initial delay time and friction coefficient, the pre-trained large language model automatically completes them based on the built-in process knowledge base: initial_delay is set to the default value of 5 seconds, and the contact friction coefficient is set to 0.15. Finally, all parameters are integrated into a standard JSON format parameter configuration file and passed to the lower-level module.

[0124] Step 2: Invoke the automated modeling and simulation program: Read the parameter configuration file generated in Step 1 and start the automated modeling Python script. The script first accesses the built-in trajectory library to read the trajectory point sequence of the hemispherical curved surface part; at the same time, it parses the batch parameters to generate three sets of simulation task parameter sets corresponding to the thickness values.

[0125] For each task group, the script sequentially executes steps S2-10, namely geometric modeling, mesh generation, and application of the wheel trajectory load. Mesh generation employs an automatic ring mesh generation method, controlled by four-dimensional parameters: inner radius, outer radius, radial layer number, and circumferential segment number. According to the "mesh refinement" command, the script uses a sparse mesh in the internal region where the sheet metal contacts the mandrel, and a refined mesh (with the radial layer number doubled) in the external flange region, automatically adjusting the radial layer spacing using an exponential density function. Following node index logic, four-node structured elements are automatically constructed to ensure topological integrity; the resulting finite element model is shown below. Figure 7 .

[0126] Load application: After reading the trajectory point sequence, the script calculates the feed rate, instantaneous linear velocity, and required time for each segment, with an initial delay of 5 seconds. The physical feed is converted to Abaqus amplitude definitions: the mandrel angular velocity amplitude is converted from a rotational speed of 60 rpm to 1 rad / s, and the translational amplitude is based on a rotational mapping between the half-cone angle and the installation angle of each segment, accumulating to generate a sequence of displacement amplitudes in the X and Y directions. Finally, a keyword input file (.inp) conforming to the solver specification is generated, named "task_thickness_value.inp".

[0127] Three sets of simulation tasks enter the task queue. Since the user hasn't explicitly specified the number of parallel tasks, the system defaults to running two tasks simultaneously, submitting them sequentially to the Abaqus / Explicit solver for computation. The GUI interface displays the real-time status of each task: the 1.8mm thickness task is running, the 2.0mm thickness task is waiting, and the 2.2mm thickness task is waiting. Once all simulations are complete, the output files are automatically archived with names such as "task_1.8mm.odb", "task_2.0mm.odb", and "task_2.2mm.odb".

[0128] Step 3: Interactive forming information extraction and quality assessment based on a pre-trained large language model (a large language model specifically for spinning): After the simulation calculation is completed, the user submits a specific query request to the system. The natural language command is: "Which thickness group has the most severe wrinkling? What is the maximum thinning rate?"

[0129] The spinning-specific large language model maps this query requirement to pre-defined post-processing function call instructions in the background, driving the automated post-processing script to access the ODB result file. The script first extracts the axial coordinate sequence of the edge nodes of the workpiece for each task, and calculates the degree of flange wrinkling: maximum amplitude, average peak-to-valley amplitude, variance, and complete wavenumber. At the same time, it extracts the wall thickness distribution and calculates the maximum thinning rate.

[0130] After summarizing the results, the system provides feedback to the user in natural language using a dedicated spinning language model: "Wrinkling is most severe at a thickness of 1.8mm, with an average wave height of 1.2mm, a maximum wave height of 2.1mm, and a wave number of 5."

[0131] Through the above steps, this invention achieves full automation from natural language command input to simulation result output. Users do not need to master Abaqus GUI operation or Python programming skills. They can complete the simulation and analysis of spinning forming of complex curved parts simply by describing the process requirements in natural language. This significantly reduces the threshold for using professional software and improves the modeling efficiency and the level of intelligence in result extraction.

[0132] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A large language model driven spinning forming automation modeling and simulation method, characterized in that, include: Natural language instructions for spinning simulation are obtained, and the natural language instructions are parsed using a pre-trained large language model to obtain structured parameters; The structured parameters include geometric modeling parameters, mesh generation parameters, and process parameters; the process parameters include feed ratio, mandrel rotation speed, and trajectory parameters. Model the sheet metal based on the geometric modeling parameters to obtain the sheet metal model; Obtain the inner and outer radii of the sheet metal model, and set the number of radial layers and the number of circumferential segments based on the mesh generation parameters; The sheet metal model is divided according to the number of radial layers and the number of circumferential segments, resulting in multiple sectors in each radial layer; The node coordinates are determined based on the radial layer number, inner radius, outer radius, radial layer number, and sector number of the sheet metal model. The adjacent radial layers and adjacent sectors are determined based on the node coordinates, and four-node structured elements are constructed between the adjacent radial layers and adjacent sectors to obtain the finite element model. The trajectory point sequence is determined based on the trajectory parameters, and the coordinates of the trajectory points for each spinning path are determined. The feed rate of each spinning path is calculated based on the difference in coordinates of the trajectory points of two adjacent spinning paths. The instantaneous linear velocity is determined based on the feed ratio and the core die rotation speed; the feed time of each spinning path is determined based on the feed amount and the average linear velocity of two adjacent spinning paths, and the cumulative time is also determined. The rotational speed and feed rate of the mandrel are converted into amplitudes. Based on the corresponding converted amplitudes and cumulative time, spinning is performed on the finite element model to obtain the workpiece.

2. The automated modeling and simulation method for spinning driven by a large language model according to claim 1, characterized in that, The training process of a pre-trained large language model is as follows: Based on knowledge related to spinning forming, a knowledge base mapping "natural language - computer language" is constructed; The mapping knowledge base includes multiple core terminology groups; each core terminology group includes a natural language instruction and the corresponding structured parameters of that natural language instruction. Based on the mapping knowledge base, construct multiple natural language instruction-structured parameter pairs; Extract multiple real instruction-parameter pairs from historical simulation projects; A training set is composed of multiple natural language instruction-structured parameter pairs and multiple real instruction-parameter pairs. The training set is then input into the large language model for training. During the training process, the LoRA strategy is used to fine-tune the parameters of the large language model, resulting in a pre-trained large language model.

3. The automated modeling and simulation method for spinning driven by a large language model according to claim 1, characterized in that, Amplitude conversion is performed on the core mold rotation speed and feed rate, including: Convert the core mold rotation speed into the core mold angular velocity amplitude; Determine the semi-cone angle of each spinning path based on the coordinates of the trajectory points of two adjacent spinning paths; The feed rate is rotated and mapped based on the half-cone angle and installation angle of each spinning path segment to obtain the displacement amplitude of each spinning path segment; the displacement amplitudes of all spinning paths segment are summed to obtain the cumulative displacement amplitude.

4. The automated modeling and simulation method for spinning driven by a large language model according to claim 1, characterized in that, Before dividing the sheet metal model according to the number of radial layers and the number of circumferential segments, the following is also included: Based on the radial progressive densification strategy, the number of radial layers in the internal and flange regions of the sheet metal model is adjusted.

5. The automated modeling and simulation method for spinning driven by a large language model according to claim 1, characterized in that, The node coordinates are determined based on the radial layer number, inner radius, outer radius, radial layer number, and sector number of the sheet metal model, including: The radial position of each radial layer is determined based on the radial layer number, inner radius, outer radius, and radial layer number of the sheet metal model. The circumferential direction angle of each radial layer is determined based on the number of circumferential segments and the sector number. The node coordinates are determined by the product of the radial position and the circumferential angle.

6. The automated modeling and simulation method for spinning driven by a large language model according to claim 5, characterized in that, Based on the node coordinates, adjacent radial layers and adjacent sectors are determined. Four-node structured elements are then constructed between adjacent radial layers and adjacent sectors to obtain the finite element model, which includes: Based on the radial layer number, sector number, and circumferential angle, the node index logic is determined; based on the node index logic, four-node structured elements are constructed between adjacent radial layers and adjacent sectors to obtain the finite element model.

7. The automated modeling and simulation method for spinning driven by a large language model according to claim 1, characterized in that, Structural parameters also include thinning rate and flange wrinkling degree; The degree of flange wrinkling is determined based on the maximum amplitude, average peak-to-valley amplitude, variance, and complete wavenumber. Maximum amplitude is used to capture extreme deviations from wrinkling; average peak-to-trough amplitude and variance are used to characterize the typical amplitude and dispersion of fluctuations; full wave number is used to quantify the number of wrinkling cycles and the degree of flange wrinkling.

8. The automated modeling and simulation method for spinning driven by a large language model according to claim 7, characterized in that, The process for determining the maximum amplitude is as follows: obtain the axial coordinate sequence of the edge nodes of the workpiece, determine the fluctuation variance based on the axial coordinate sequence of the edge nodes of the workpiece, and determine the maximum amplitude based on the maximum and minimum values ​​in the axial coordinate sequence.

9. The automated modeling and simulation method for spinning driven by a large language model according to claim 7, characterized in that, The process of determining the average peak-valley amplitude is as follows: perform a two-point moving average on the axial coordinate sequence of the edge nodes of the workpiece to obtain a smooth sequence, and determine the mean of the smooth sequence based on the smooth sequence. For each pair of adjacent smooth sequences, a cross-mean detection is performed sequentially. When the detection result is a cross-mean, a "zero cross" index is recorded. Based on the "zero crossover" index, all smooth sequences are divided into several peak-valley candidate intervals, and the set of local maxima and local minima within each peak-valley candidate interval are determined. The average peak-valley amplitude is determined based on the set of local maxima and local minima of all peak-valley candidate intervals. The complete wave number is determined based on the number of "zero crossover" indices.

10. A large language model-driven automated modeling and simulation system for spinning forming, characterized in that, include: The language instruction parsing module is used to obtain natural language instructions for spinning simulation. It parses the natural language instructions using a pre-trained large language model to obtain structured parameters. The structured parameters include geometric modeling parameters, mesh generation parameters, and process parameters; the process parameters include feed ratio, mandrel rotation speed, and trajectory parameters. The modeling module is used to create a sheet metal model based on geometric modeling parameters. The mesh generation module is used to obtain the inner and outer radii of the sheet metal model, and set the number of radial layers and the number of circumferential segments based on the mesh generation parameters. The sheet metal model is then divided according to the number of radial layers and circumferential segments, resulting in multiple sectors in each radial layer. Node coordinates are determined based on the number of radial layers, inner and outer radii, the radial layer number, and the sector number. Adjacent radial layers and sectors are determined based on the node coordinates, and four-node structured elements are constructed between adjacent radial layers and sectors to obtain the finite element model. The trajectory point feed module is used to determine the trajectory point sequence based on trajectory parameters and to determine the coordinates of the trajectory points in each spinning path; to calculate the feed amount of each spinning path based on the difference in coordinates between the trajectory points of two adjacent spinning paths; to determine the instantaneous linear velocity based on the feed ratio and the mandrel rotation speed; and to determine the feed time of each spinning path based on the feed amount and the average linear velocity of two adjacent spinning paths, and to determine the cumulative time. The amplitude conversion and spinning module is used to convert the core mold rotation speed and feed rate into amplitudes. Based on the corresponding amplitude and cumulative time after conversion, spinning is performed on the finite element model to obtain the workpiece.

Images