A tire pattern image automatic generation method and system based on an LLM model

Abstract

The present application relates to the technical field of intelligent tire design, and particularly relates to a tire pattern image automatic generation method and system based on an LLM model. The method first calculates noise values and total edge lengths of sample patterns, and processes the patterns into 512*512 black and white images. For each image, a structured label is constructed, which includes pattern type, pitch length and width, sea-land ratio, pitch arrangement sequence, noise value and total edge length. Based on samples with N>100, a large model for image generation is fine-tuned and trained, and the model is saved. In the inference stage, a prompt word (including pattern type, pitch parameters, pitch arrangement and noise / edge length target) analyzed by a large language model is input to generate a pattern image. If necessary, rules are projected and corrected to improve manufacturability, and pattern contour data that can be used for molding processing is output, thereby shortening the pattern design cycle, reducing trial and error costs, and realizing controllable optimization of noise and boundary complexity.

Description

Technical Field

[0001] This invention relates to the field of intelligent tire design technology, and in particular to a method and system for automatically generating tire tread patterns based on an LLM model. Background Technology

[0002] As a key structure for radial tires to contact the road surface and achieve comprehensive performance in traction, drainage, braking, wear resistance, and noise reduction, tire tread patterns typically require a trade-off between multiple objectives, including wet grip / drainage, dry handling stability, rolling resistance, wear life, and tread noise. In actual R&D, tread pattern design often revolves around elements such as "tread groove type—tread block shape—pitch sequence—land-sea ratio—sipes / fine tread features," and is completed through CAD / digital modeling parametric modeling, simulation, and prototype testing iterations. Due to the high degree of freedom in tread structure and strong objective conflicts (for example, improving drainage capacity often involves changes in the land-sea ratio and an increase in block boundaries, which may cause coupled changes in noise and wear), the traditional design process relies heavily on experience and repeated trial and error, resulting in long development cycles, heavy workloads for designers, and difficulty in systematically accumulating a reusable knowledge base of "structure—indicators—performance."

[0003] To improve tire tread design efficiency, various digital and algorithmic tools have emerged in existing technologies, supporting the R&D process from the perspectives of parametric modeling of tread geometry, rapid prediction of noise performance, and automatic calculation of key geometric indicators such as the land-sea ratio. For example, Chinese patent document CN106021669A discloses a parametric design system and method for tire tread patterns, which modularizes tread groove design, groove intersection processing, decorative / marking features, and automatic assembly, aiming to standardize the tread design process and improve design efficiency. This type of solution can improve modeling speed and structural consistency through regularized parametric units and interactive modeling, but its essence is still "combining and editing within a predetermined rule base / feature template." For high-degree-of-freedom content such as tread block boundary morphology, groove connectivity topology, and detailed distribution, a large amount of manual intervention by engineers is still required. At the same time, this type of parametric system usually focuses more on "geometric modeling efficiency," and the constraints on performance indicators such as noise often rely on external simulation or empirical criteria, making it difficult to form an end-to-end closed-loop generation capability of "input target → automatically generate a tread pattern that meets the target."

[0004] In tread noise control, pitch sequence is one of the key factors affecting tire driving noise and harmonic peak values. Existing technologies have relatively mature and rapid evaluation methods for the "pitch sequence—excitation—spectrum—noise" chain. Chinese patent document CN108614935A discloses a method for rapidly predicting tire tread pitch noise: after setting basic tire parameters, the pitch sequence is determined, the excitation is extracted and an excitation distribution map is generated, and a Fourier transform is performed on the excitation to obtain a pitch noise spectrum map, supporting the rapid selection of pitch sequence while meeting tread design requirements. This type of solution can significantly reduce the threshold and computational cost of pitch noise analysis, but its focus is mainly on "noise prediction and evaluation for a given pitch sequence," usually still requiring engineers to first propose candidate pitch sequences and candidate tread structures; in other words, it addresses "how to evaluate more quickly" rather than "how to directly generate" a tread pattern that meets the noise target and is consistent with the tread geometry. Furthermore, pitch noise is often closely related to the boundaries of pattern blocks, groove connectivity, and local details. Frequency domain prediction of the pitch sequence alone is insufficient to cover the comprehensive impact of the overall pattern structure on noise. In engineering, multiple rounds of manual adjustment of pattern block / groove details are still required to match the pitch design.

[0005] In the rapid calculation of land-to-sear ratio and related geometric indices, image-based and pixel-statistical methods are widely used due to their ease of implementation. Chinese patent document CN112115549A discloses a two-dimensional land-to-sear ratio calculation method for tire tread patterns based on image recognition and its application in tread pattern design: first, the tire tread pattern is converted into a black-and-white PNG image; then, the total land-to-sear ratio is calculated by recognizing the number of black and white pixels, as well as the land-to-sear ratio being equally divided along the length or width direction, thereby improving tread pattern development efficiency and shortening the development cycle. This type of solution transforms the "calculation of tread pattern geometric indices" from complex geometric intersection / area calculations to image pixel statistics, offering advantages in speed and efficiency; however, it primarily addresses the "indicator measurement" problem, i.e., when the tread pattern image already exists, indicators such as the land-to-sear ratio can be quickly obtained. For the question of "given the target land-sea ratio, target noise and target pitch parameters, how to directly generate a pattern that meets the constraints and is manufacturable", existing methods usually still require designers to first generate / edit the pattern, then repeatedly calculate and modify the indicators, which is an offline iterative process of "post-evaluation - modification", which is difficult to achieve an automated closed loop.

[0006] In summary, the aforementioned existing technologies have improved the efficiency of digital development of tire tread patterns from different aspects: one type improves the level of modeling and structural standardization through parametric design systems (e.g., CN106021669A); another type enables rapid evaluation and screening of pitch sequences through rapid prediction of pitch noise (e.g., CN108614935A); and yet another type enables rapid calculation of key indicators such as the land-sea ratio through image recognition (e.g., CN112115549A). However, in response to the need for "rapid concept generation of new tread patterns," existing solutions still have at least the following shortcomings:

[0007] 1) Fragmented toolchain and lack of end-to-end closed loop: Parametric modeling, noise prediction, and land-sea ratio calculation are mostly independent tools. Engineers usually rely on manually transferring data and making repeated adjustments between multiple software / modules, making it difficult to form a closed loop process of "target indicator driven → automatic pattern generation → automatic verification → automatic correction".

[0008] 2) Difficulty in simultaneously handling multiple objective constraints and structural manufacturability: Pitch noise evaluation focuses on sequence frequency domain characteristics, land-sea ratio evaluation focuses on area proportion, and parametric design focuses on regular modeling; however, actual pattern patterns must simultaneously meet manufacturing constraints such as groove connectivity, minimum glue bridge width, minimum groove width, isolated small block control, and fillet / tool ​​accessibility, and coordinate with objectives such as noise, land-sea ratio, and pitch length and width. Existing technologies often require manual trade-offs between indicators.

[0009] 3) Inefficient knowledge representation and requirement input: Traditional parametric systems rely on professionals understanding a large number of design rules and parameters, and the calculation and prediction methods depend on engineers forming candidate structures before evaluation. Faced with the demands of electrification, segmented scenarios, and rapid product iteration, R&D urgently needs an intelligent method that can uniformly express requirements such as "pattern type, pitch size and sequence, target noise / harmonic constraints, target land-sea ratio, and boundary complexity," and directly drive pattern generation and constraint satisfaction.

[0010] Therefore, how to further realize the following on the basis of the existing "parametric design + index calculation / prediction": to unify the modeling of tread pattern and structured labels (including pitch sequence and noise / harmonic index, land-sea ratio, boundary complexity, etc.), form a reusable generative model through data-driven approach, and introduce an automatic correction mechanism for target constraints and manufacturability constraints in the generation process, so as to achieve efficient automatic generation and rapid iteration of tire tread pattern, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0011] The technical objective of this invention is to provide an automatic tire tread pattern generation method and system. By constructing a dataset with structured labels such as tread pattern type, pitch size and pitch arrangement sequence, land-sea ratio, segmented harmonic index, and total side length, a conditional generation model is trained, and a differentiable constraint evaluation and a closed-loop correction mechanism of regular projection / differentiable projection are introduced. This ensures that the generated tread pattern meets target performance requirements such as noise (segmented harmonics), land-sea ratio, and boundary complexity, while also conforming to manufacturability constraints such as minimum groove width, minimum rubber bridge width, connectivity, and fillet radius. The system outputs vector contour data that can be used for mold making, thereby shortening the tread pattern design cycle, reducing manual trial and error costs, and improving design consistency and reusability.

[0012] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution: An automatic tire tread pattern generation method based on an LLM model, comprising the following steps: S1, obtain tire tread pattern samples, process each sample into a black and white image and generate a structured label; the structured label includes at least the tread pattern type, pitch length, pitch width, land-sea ratio, pitch arrangement sequence, segmented harmonic index and total tread pattern side length. S2, establish a pixel-to-millimeter mapping relationship based on pitch length and pitch width, and set manufacturing thresholds, the manufacturing thresholds including at least minimum trench width threshold, minimum adhesive bridge width threshold, minimum isolated block area threshold and minimum rounded corner radius threshold; S3 trains the conditional generation model and the differentiable constraint evaluation model, enabling the conditional generation model to generate candidate pattern images under the generation conditions consisting of pattern type, pitch length, pitch width, land-sea ratio, pitch arrangement sequence, piecewise harmonic index, and total pattern side length. The differentiable constraint evaluation model outputs the predicted piecewise harmonic index vector, predicted land-sea ratio, predicted total pattern side length, and manufacturing violation degree for the candidate pattern images, and is used to construct the differentiable constraint loss. S4: Receive the target constraint set and perform rule projection to obtain legal constraints; S5 generates candidate pattern images under legal constraints using a conditional generation model. Based on the evaluation results of the candidate pattern images using a differentiable constraint evaluation model, it judges the piecewise harmonic constraints, land-sea ratio constraints, total side length constraints, and millimeter-scale manufacturing constraints. It iteratively updates the latent variables and / or generation conditions until they are satisfied, and then outputs the final pattern image. The final pattern image is then extracted into vector contour data for mold making or processing.

[0013] Preferably, in S1, the number of samples is greater than 100, and each sample is processed into a black and white image with a resolution of 512 by 512. In the black and white image, the first pixel value represents the land area of ​​the rubber compound, the second pixel value represents the sea area of ​​the trench, and the land-sea ratio is calculated by the proportion of black and white pixels. And / or, the pitch arrangement sequence is composed of discrete codes of length equal to the number of pitch units; each code is taken from set 1, 2, 3, where codes 1, 2, 3 respectively represent the corresponding pitch units scaled by 0.8 times, 1.0 times, and 1.2 times in the circumferential length direction; And / or, the segmented harmonic index is an index set consisting of multiple circumferential segmented harmonic values, wherein the number of circumferential segments is not less than 3; each circumferential segmented harmonic value is obtained by mapping the pitch arrangement sequence to the circumferential excitation sequence and performing Fourier transform on the corresponding circumferential segment, and the preset frequency band includes block frequency and at least one harmonic frequency band. And / or, the total side length of the pattern is the sum of the perimeters of the boundary between the land and the ditch, and the total side length of the pattern is obtained by extracting the boundary pixel length from the black and white image, converting it to millimeters from the pixel-to-millimeter mapping relationship, and then summing them.

[0014] Preferably, in S2, the pixel-to-millimeter mapping relationship includes at least circumferential scale mapping and lateral scale mapping, which are obtained by combining the pitch length and pitch width with the image pixel size, respectively, and are used to convert the groove width, bridge width, area and corner radius into millimeter scale; And / or, the millimeter-scale manufacturing constraints include at least: the trench width is not less than the minimum trench width threshold, the adhesive bridge width is not less than the minimum adhesive bridge width threshold, the isolated block area is not less than the minimum isolated block area threshold, the minimum fillet radius is not less than the minimum fillet radius threshold, and the trench connectivity meets the preset connectivity requirements.

[0015] Preferably, in S3, the differentiable constraint loss includes at least piecewise harmonic loss, land-sea ratio loss, total side length loss and manufacturing non-compliance loss, wherein the manufacturing non-compliance loss includes at least trench width non-compliance, glue bridge width non-compliance and connectivity non-compliance, and the non-compliance is constructed as a differentiable quantity through a distance transformation field and / or a skeleton approximation field. And / or, the conditional generation model is one of a diffusion model, a generative adversarial network, or an autoregressive generation model; the differentiable constraint evaluation model is one of a convolutional neural network or a visual Transformer network; and the candidate pattern image is a pattern image output by the conditional generation model under the generation conditions.

[0016] Preferably, in S4, the regular projection includes at least: a) Verify the validity of the encoding of the interval arrangement sequence and replace it with the nearest valid sequence if it is invalid; b) Verify the consistency between the target land-sea ratio and the target black-and-white pixel ratio, and correct the target land-sea ratio or the target pixel ratio if the deviation is out of tolerance; c) Verify the consistency of the target pitch length, target pitch width, number of pitch elements, and pitch period. If there is a discrepancy, trim and remap the target pitch length, target pitch width, or number of pitch elements to obtain a valid constraint.

[0017] Preferably, in S5, the segmented harmonic constraint is that the predicted segmented harmonic index vector is not higher than the target segmented harmonic upper limit, the land-sea ratio constraint is that the difference between the predicted land-sea ratio and the target land-sea ratio does not exceed the land-sea ratio tolerance, and the total side length constraint is that the difference between the predicted total side length of the pattern and the target total side length does not exceed the total side length tolerance. And / or, in S5, the iterative update includes: when the predicted segmented harmonic index vector exceeds the target segmented harmonic upper limit, firstly perform local exchange, local rearrangement or scaling encoding replacement on the pitch arrangement sequence to reduce the corresponding segmented harmonic peak value, and then perform gradient update based on differentiable constraint loss on the generated latent variables. And / or, the iteration update of S5 ends when the following stopping conditions are met: the predicted segmented harmonic index vector is not higher than the target segmented harmonic upper limit, the difference between the predicted land-sea ratio and the target land-sea ratio does not exceed the land-sea ratio tolerance, the difference between the predicted total side length of the pattern and the target total side length of the pattern does not exceed the total side length tolerance, and the manufacturing violation degree is 0; or it ends when the preset iteration limit is reached and the current optimal pattern image is output as the final pattern image. And / or, in S5, the vector contour data is a set of closed polylines or spline curves, and before output, the sharp corners are rounded and the tool radius accessibility is checked, so that the minimum fillet radius in the vector contour data is not less than the minimum fillet radius threshold.

[0018] Preferably, the method further includes: receiving natural language design intent text, calling a large language model to parse the natural language design intent text into a set of target constraints, wherein the set of target constraints includes at least one or more of the following: target pattern type, target pitch length, target pitch width, target land-sea ratio, target pitch arrangement sequence, target segmented harmonic upper limit, and target pattern total side length, and the set of target constraints is used as the input for the rule projection in step S4; And / or, after outputting the target constraint set, the large language model performs a consistency check on the target constraint set based on a preset tire design feasible domain rule base and outputs repair constraints. The consistency check includes at least: a) Validity of the target pitch sequence encoding; b) The target land-sea ratio is consistent with the pixel ratio; c) The target pitch length, target pitch width, number of pitch elements, and pitch period are consistent; d) The target segmented harmonic upper limit is consistent with the preset frequency band definition; And replace the target constraint set with the repair constraint in step S4; And / or, when the target constraint set lacks a target pitch arrangement sequence or the target pitch arrangement sequence is incomplete, the large language model generates a candidate pitch arrangement sequence set based on the target pattern type, target pitch length, target pitch width, target land-sea ratio, and target segmented harmonic upper limit. The candidate pitch arrangement sequence set contains no less than two candidate pitch arrangement sequences. The model then calls the pitch harmonic prediction or the differentiable constraint evaluation model to output predicted segmented harmonic indices for each candidate sequence, in order to filter and obtain the target pitch arrangement sequence for the target constraint set. And / or, when the predicted segmented harmonic index vector, the predicted land-sea ratio, or the predicted total side length of the pattern in step S5 does not meet the corresponding target constraints, the deviation of the predicted segmented harmonic index vector relative to the target segmented harmonic upper limit, the deviation of the predicted land-sea ratio relative to the target land-sea ratio, the deviation of the predicted total side length of the pattern relative to the target total side length of the pattern, and the location map of manufacturing violations are input into the large language model. The large language model outputs constraint correction suggestions to update at least one of the target land-sea ratio, the target total side length of the pattern, the target segmented harmonic upper limit, or the target pitch arrangement sequence, and triggers the re-execution of steps S4 and S5.

[0019] Secondly, the present invention also provides an automatic tire tread pattern generation system, including a processor and a memory, wherein the memory stores instructions that, when executed by the processor, implement the method described above; the system includes at least: a sample and label construction module, a calibration and threshold setting module, and a conditional generation model. Training and generation modules, differentiability constraint evaluation model The module includes a training and evaluation module, a regular projection module, a second-order differentiable projection iterative correction module, and a vector contour output module.

[0020] Thirdly, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0021] Fourthly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the method.

[0022] This invention, by employing the aforementioned technical solution, transforms tire tread patterns from a process of iterative modeling and trial and error relying on manual experience into an automated generation process directly driven by target indicators. On one hand, a generation model is trained based on labels such as pitch length / width, pitch sequence, land-sea ratio, segmented harmonic index, and total tread edge length, enabling the model to quickly output tread patterns that meet the target trend according to given design constraints. On the other hand, a differentiable constraint evaluation model is introduced to quickly predict noise-related segmented harmonics, land-sea ratio, and boundary complexity, combined with pixel-to-millimeter calibration of minimum groove width, minimum rubber bridge width, isolated block area, and corner radius. Manufacturing constraints such as diameter and groove connectivity are designed to be differentiable. Regular projection ensures the legality of input constraints, and iterative correction through differentiable projection ensures that the generated results converge in both performance and manufacturability. This significantly reduces rework caused by "meeting only the appearance of the image without meeting engineering constraints." At the same time, the output pattern can be further vectorized and the accessibility of tools / molds can be verified. It can be directly connected to the pattern CAD / mold manufacturing process, thereby improving design efficiency and consistency while ensuring noise control and consistency of pattern indicators. This shortens the development cycle, reduces the number of simulations and prototypes, and improves the feasibility and batch reusability of the pattern scheme. Attached Figure Description

[0023] Figure 1 A structural block diagram of a system for automatically generating tire tread patterns.

[0024] Figure 2 This is a flowchart of the invention method.

[0025] Figure 3 This is a flowchart of the closed-loop generation and constraint correction process in the reasoning generation stage.

[0026] Figure 4 This is a schematic diagram of the vectorization and manufacturability verification process for pattern images.

[0027] Figure 5 This is a schematic diagram of the patterned blocks used in the training samples.

[0028] Figure 6 This is a schematic diagram of a striped pattern generated based on prompt words. Detailed Implementation

[0029] The following provides a detailed description of the specific implementation of the "Automatic Generation Method and System for Tire Tread Patterns Based on a Large Language Model". Those skilled in the art should understand that the following embodiments are used to explain the present invention, not to limit the scope of protection; without departing from the spirit of the present invention, those skilled in the art can make equivalent substitutions or modifications to the order of steps, model structure, parameter selection, and engineering implementation, all of which should fall within the scope of protection of the present invention.

[0030] For ease of description, the binary representation of the tire tread pattern is referred to as a "black and white image" in the instruction manual, where black and white represent "land areas (rubber solids)" and "marine areas (grooves and voids)," respectively. The correspondence between the two can be interchanged in specific products, but should remain consistent within the same implementation method.

[0031] I. Explanation of Terms and Symbols

[0032] 1. Pattern Image : Refers to the binary representation of the tire tread pattern on a two-dimensional plane, with a resolution of 512×512 pixels.

[0033] 2. Pixel values ​​are divided into two categories: one category represents land areas (rubber material), and the other category represents sea areas (trench).

[0034] 3. Sample size The total number of pattern samples used for training, satisfying... .

[0035] 4. Pattern Types : Pattern category labels, such as all-season tires, winter tires, unidirectional tires, mud tires, agricultural tires, engineering tires, all-steel tires, etc.

[0036] 5. Pitch length The actual length of adjacent pitch units along the tire circumference (rolling direction), in units of .

[0037] 6. Pitch width The actual width of the lateral (axial) pitch unit along the tire tread, in units of .

[0038] 7. Land-sea ratio : The proportion of land area to the total pattern area (dimensionless). In binary images, it can be estimated using pixel area ratio and converted using pixel-to-millimeter calibration.

[0039] 8. Interval arrangement sequence : The pitch coding sequence along the circumferential direction, with a length equal to the number of pitch units. ( (where is a positive integer). In one embodiment, The elements are taken from sets 1, 2, and 3, representing scaling factors of 0.8, 1.0, and 1.2 in the circumferential length direction, respectively.

[0040] 9. Piecewise Harmonic Index Vector : Sequence arranged by interval The harmonic energy / peak value vector obtained by the circumferential excitation signal mapped and subjected to a circumferential piecewise Fourier transform is denoted as... ,in This represents the number of segments in the circumferential direction.

[0041] 10. Total side length of the pattern The sum of the perimeters of the patterned blocks, in units of... .

[0042] 11. Pixel-to-millimeter mapping Convert the length / area of ​​the pixel scale to the actual size. The calibration relationship can include circumferential mapping and lateral mapping respectively.

[0043] 12. Manufacturing threshold:

[0044] Minimum trench width threshold, in mm;

[0045] Minimum adhesive bridge width threshold, in mm;

[0046] Minimum isolated block area threshold, in mm 2 ;

[0047] Minimum fillet radius threshold, in mm.

[0048] 13. Conditional Generative Model : with conditional vector Generate candidate pattern images for input. Image generation models (diffusion model, GAN, autoregressive generation model are all acceptable).

[0049] 14. Differentiable constraint evaluation model : Candidate pattern images Output Predictive Indicators It is a proxy model for manufacturing violations, and its output can be used to construct differentiable constraint losses.

[0050] 15. Set of Objective Constraints With legal constraints : The target constraints are input by the user or given by the upper-level system. To Legal constraints after rule projection.

[0051] 16. Candidate Images With the final image : These are candidate pattern images used in the generation process. The final pattern is designed to meet both the target constraints and manufacturing constraints.

[0052] 17. Target threshold and tolerance:

[0053] : Target piecewise harmonic upper bound vector;

[0054] Target: Land-sea ratio;

[0055] Target total side length;

[0056] Land-sea specific tolerance;

[0057] Total side length tolerance.

[0058] 18. Vector contour data : From the final pattern image The extracted set of closed polylines / spline curves is used for mold making or machining path input.

[0059] II. System Structure of the Invention

[0060] 2.1 System Structure

[0061] like Figure 1 As shown, the system of the present invention can be deployed on a standalone workstation, an enterprise private cloud, or an edge computing server. A typical structure includes:

[0062] 1) Sample and Label Construction Module

[0063] This is used to receive original pattern design drawings (CAD / bitmap / rendered images) and convert them into binary pattern images. and generate structured tags. (including) wait).

[0064] 2) Calibration and manufacturing threshold module

[0065] Used to establish pixel-to-millimeter mapping And set according to enterprise standards / process capabilities .

[0066] 3) Condition generation module (model) )

[0067] Used to train and invoke an image generation model, under given conditions or Output candidate pattern images .

[0068] 4) Differentiable constraint evaluation module (model) )

[0069] Used for predict It outputs the manufacturing violation degree (such as width violation, connectivity violation, isolated block violation, etc.) in order to form a differentiable closed-loop correction.

[0070] 5) Regular projection module

[0071] Used to constrain user goals Legalized as This avoids infeasible inputs that could lead to divergent or unmanufacturable outputs.

[0072] 6) Second-order differentiable projection iterative correction module

[0073] Used to iteratively update latent variables and / or conditions during the generation phase, so that the output simultaneously satisfies performance constraints and manufacturing constraints.

[0074] 7) Vectorization and Modeling Output Module

[0075] Used to convert the final binary pattern Perform boundary extraction, corner rounding, and tool reachability verification to output vector contours. (e.g., DXF / SVG / custom spline formats).

[0076] 8) LLM semantic parsing and closed-loop suggestion module

[0077] Used to parse natural language design intent into Furthermore, it provides feasible domain repair and strategy suggestions when there are constraint conflicts or generation deviations, forming an interactive closed loop of "human-LLM-generative model-evaluation model".

[0078] III. Specific Technical Route for Implementing the Method of the Invention

[0079] like Figure 2 As shown, an automatic tire tread pattern generation method based on an LLM model is proposed, which includes the following steps: S1, obtain tire tread pattern samples, process each sample into a black and white image and generate a structured label; the structured label includes at least the tread pattern type, pitch length, pitch width, land-sea ratio, pitch arrangement sequence, segmented harmonic index and total tread pattern side length. S2, establish a pixel-to-millimeter mapping relationship based on pitch length and pitch width, and set manufacturing thresholds, the manufacturing thresholds including at least minimum trench width threshold, minimum adhesive bridge width threshold, minimum isolated block area threshold and minimum rounded corner radius threshold; S3 trains the conditional generation model and the differentiable constraint evaluation model, enabling the conditional generation model to generate candidate pattern images under the generation conditions consisting of pattern type, pitch length, pitch width, land-sea ratio, pitch arrangement sequence, piecewise harmonic index, and total pattern side length. The differentiable constraint evaluation model outputs the predicted piecewise harmonic index vector, predicted land-sea ratio, predicted total pattern side length, and manufacturing violation degree for the candidate pattern images, and is used to construct the differentiable constraint loss. S4: Receive the target constraint set and perform rule projection to obtain legal constraints; S5 generates candidate pattern images under legal constraints using a conditional generation model. Based on the evaluation results of the candidate pattern images using a differentiable constraint evaluation model, it judges the piecewise harmonic constraints, land-sea ratio constraints, total side length constraints, and millimeter-scale manufacturing constraints. It iteratively updates the latent variables and / or generation conditions until they are satisfied, and then outputs the final pattern image. The final pattern image is then extracted into vector contour data for mold making or processing.

[0080] 3.1 Step S1: Sample set construction, binarization, and structured label generation

[0081] 3.1.1 Pattern Data Source and Standardization

[0082] Sample patterns can be derived from: (1) CAD drawings, two-dimensional unfolded drawings, and mold pattern drawings of historical mass-produced tire patterns; (2) concept pattern drawings in the R&D stage (including hand-drawn / vector sketches / simulation drawings); (3) pattern drawings from public information or databases.

[0083] Data from different sources were unified into a binary pattern image. The basic process is as follows:

[0084] 1) If the input is a vector outline: rasterize and render to 512×512, and fill the groove area as "sea area" and the rubber area as "land area";

[0085] 2) If the input is a grayscale / color bitmap: perform grayscale conversion, threshold segmentation, and morphological denoising (such as opening and closing operations) to obtain a clean binary image;

[0086] 3) For pattern boundaries with break lines or spike noise, the edges can be refined (skeletonized) or smoothed to reduce label calculation errors.

[0087] 3.1.2 Sea-to-land ratio calculate

[0088] Sea and land ratio In this embodiment, it is defined as the proportion of land area to the total area. It can be approximately calculated using pixel count:

[0089] ;

[0090] in: The land area The area represents the sea region; both can be calculated by multiplying the number of corresponding pixels in the binary image by the area of ​​a single pixel (via...). (Conversion) If the millimeter scale is not yet calibrated, pixel ratio can be used as a preliminary method. The dimensionless estimate of .

[0091] 3.1.3 Total side length of the pattern calculate

[0092] The total side length of the pattern is used to characterize boundary complexity and potential noise trends. The boundary pixel set is obtained through boundary extraction (e.g., Canny or morphological boundary operators). And convert its pixel length to millimeters:

[0093] ;

[0094] in: The length of the boundary pixels (can be obtained by accumulating the chain code length or pixel adjacency distance); The length conversion factor from pixels to millimeters (by...) (Given).

[0095] 3.1.4 Pitch Arrangement Sequence Piecewise harmonic index

[0096] Interval arrangement sequence It can be given directly by the design end (e.g., "231211..."), or it can be inferred by circumferential slicing of the image, matching unit templates, or based on repeating structure detection.

[0097] In one embodiment, Elements belonging to 1, 2, and 3 represent circumferential scaling factors of 0.8, 1.0, and 1.2 for that pitch element, respectively. This is used to obtain the piecewise harmonic indices. First Mapped to a circumferential excitation sequence For example, using the pitch element boundary or block boundary as the excitation source, a discrete excitation signal is constructed; then the circumferential direction is divided into... For each segment, perform a Fourier transform and calculate the energy / peak value within a preset frequency band to obtain... The preset frequency band can cover the block frequency and at least one harmonic to be associated with noise control objectives.

[0098] 3.2 Step S2: Pixel-to-millimeter calibration and manufacturing threshold setting

[0099] 3.2.1 Pixel-to-Millimeter Mapping

[0100] Since the pattern image is used for generation and manufacturing, the pixel scale must be aligned with the actual size. This embodiment uses a method calibrated by the actual pitch size: the pitch length corresponding to the pattern in the actual tire deployment is known. With pitch width And the image resolution is known to be Then, pixel scales in the circumferential and horizontal directions can be established separately:

[0101] ;

[0102] in: The length per pixel in the circumferential direction ( ), The length per pixel in the horizontal direction ( ); The number of pixels corresponding to the circumferential / horizontal direction (which can be taken as...) in common implementations Alternatively, the value can be taken from the effective pixel count after ROI cropping. Therefore... This can be represented as the distance of any pixel. Mapped to millimeter distance The relationship.

[0103] 3.2.2 Manufacturing Threshold Setting

[0104] Manufacturing thresholds are used to ensure that the generated patterns can be processed in the field and meet engineering requirements such as strength, drainage, and abrasion resistance. Typical thresholds include:

[0105] Minimum groove width (mm) to prevent the groove from being too narrow, which could lead to difficulties in mold processing or insufficient drainage;

[0106] Minimum adhesive bridge width (mm) to prevent the adhesive bridge from being too narrow, which could cause the block to tear easily or have insufficient durability;

[0107] Minimum isolated block area (mm²) 2 This prevents the formation of excessively small isolated blocks that could lead to delamination or burrs during processing.

[0108] Minimum fillet radius (mm) to prevent stress concentration and machinability issues caused by sharp corners.

[0109] The above thresholds can be determined based on enterprise standards, mold and tool radii, and experience with rubber flow and vulcanization. For example: Desirable , Desirable , Related to tool radius, can be taken , It can be combined with pixel-to-millimeter mapping to convert into an area threshold. The threshold is not fixed and is a configurable parameter.

[0110] 3.3 Step S3: Model Training

[0111] The core of step S3 lies in coupling "generation" and "performance / manufacturability constraint evaluation" into an optimizable closed loop. Traditional methods often separate generation and evaluation: generation occurs first, followed by offline simulation / manual review; this invention trains a differentiable evaluation model. And construct a differentiable constraint loss to make the generative model Both the training and inference phases can be constrained, significantly improving the probability of satisfying constraints and the convergence speed.

[0112] 3.3.1 Dataset Construction and Data Format

[0113] 3.3.1.1 Sample Composition and Quantity

[0114] Sample set Where N>100, preferably N=150-2000.

[0115] It is a binary pattern image with a fixed resolution of 512×512; the pixel value can be 0, 1 or -1, 1. For structured tags, at least the following must be included: in: Pattern type; Pitch dimension (mm); Land-sea ratio; Interval sequence; Noise level (dB); Total side length (mm).

[0116] 3.3.1.2 Tag Numericalization and Normalization

[0117] To ensure training stability, continuous variables are normalized to... or :

[0118] For any continuous variable : in This data was obtained from statistics on the training set.

[0119] 3.3.1.3 Pitch Sequence Encoding method

[0120] For length A discrete sequence, with elements taken from 1, 2, 3 (representing circumferential scaling of 0.8 / 1.0 / 1.2 times respectively). Two implementable encoding methods are provided (either one can be used to reproduce the result):

[0121] Option A: Sequence embedding (Recommended)

[0122] Will Input Embedding layer This yields the sequence embedding:

[0123] Then, a fixed-dimensional vector is obtained through 1D convolution or a Transformer encoder. As a condition.

[0124] Option B: Histogram / Statistical Feature Coding (Simplified)

[0125] statistics Construct feature vectors by considering the proportions and number of transitions of each encoding element. (For example) and concatenate with continuous variables.

[0126] 3.3.1.4 Data Partitioning and Augmentation

[0127] Division: Training / Validation / Test = 80% / 10% / 10% (or 70 / 15 / 15).

[0128] Enhancements: Only enhancements that do not change the physical meaning are allowed, such as slight boundary smoothing and 1-pixel-level morphological denoising; large-angle rotations / mirroring are not recommended (as they may destroy the meaning of the pitch direction), and if rotation is necessary, it should be updated synchronously. The direction is defined.

[0129] 3.3.2 Generative Model Structure

[0130] 3.3.2.1 Generating Targets and Inputs / Outputs

[0131] Generative models Distribution of learning conditions , where the condition vector is:

[0132] ;

[0133] The output is a candidate pattern image. (Both are 512×512 binary domains, and can be represented in the continuous domain during training, and then thresholded after inference).

[0134] 3.3.2.2 Optimal Structure: Conditional Diffusion Model

[0135] (1) Forward diffusion (training noise injection)

[0136] For real images conduct Step-by-step noise addition:

[0137] ;

[0138] in: for A continuous representation (e.g., mapping 0,1 to [-1,1]); This is the cumulative noise schedule factor.

[0139] (2) Reverse Network (U-Net)

[0140] Using U-Net as the noise predictor :

[0141] Encoder / decoder layers: 4-5 levels;

[0142] Basic number of channels: 64 or 128;

[0143] Normalization: GroupNorm;

[0144] Activation: SiLU;

[0145] Attention layer: , Add Self-Attention or Cross-Attention at the resolution.

[0146] (3) Conditional injection method

[0147] Conditions Mapped to condition vector And inject into U-Net:

[0148] type Embedding obtained ;

[0149] sequence : Obtained according to 3.3.1.3 ;

[0150] Continuous variables: obtained by concatenation and MLP. ;

[0151] Overall conditions:

[0152] ;

[0153] Injection method: Use FiLM modulation or Cross-Attention in each layer of U-Net:

[0154] FiLM: for features implement ,Then

[0155] Cross-Attention: Inject attention layers as key / value pairs.

[0156] Note: The invention can be achieved by any of the above injection methods; in actual engineering, both methods can be used together to enhance controllability.

[0157] 3.3.3 Constraint Evaluation Model Structure

[0158] To train / inference Get it quickly and (optional) manufacturing violation degree Training and evaluating models .

[0159] 3.3.3.1 Input / Output Definitions

[0160] Input: Candidate pattern images (or real during training) ).

[0161] Output: in This is a piecewise harmonic vector.

[0162] 3.3.3.2 Preferred Structure: CNN Regression Network or ViT

[0163] CNN Solution: Modified ResNet-18 / 34 regression head;

[0164] ViT Solution: ViT-S / 16+ Regression MLP Header;

[0165] Multi-task output: Shared backbone + multiple output heads ( (Each one end), loss weighted.

[0166] 3.3.4 Training Methods

[0167] 3.3.4.1 Training sequence (three stages recommended to ensure reproducibility)

[0168] Phase A: Training and evaluating the model (Return to Supervision)

[0169] Training on real samples This enables it to have stable predictive capabilities.

[0170] Phase B: Training / Fine-tuning the Generative Model (Conditional diffusion)

[0171] Training in Phase A Based on this, make minor adjustments Furthermore, constraint losses can be introduced to improve controllability.

[0172] Phase C: Joint Fine-tuning (Optional)

[0173] Small learning rate joint fine-tuning and To reduce domain bias.

[0174] 3.3.4.2 Loss function and training details

[0175] For each sample The monitoring signal is (Calculated or calibrated in step 1). Loss:

[0176] ;

[0177] in: As weight, preferred (Noise is more important and can be improved) ).

[0178] Recommended training hyperparameters (ready to be implemented directly):

[0179] Optimizer: AdamW; Learning Rate Weight decay ;

[0180] Batch size (Depending on video memory);

[0181] Number of training rounds ;

[0182] Early stopping: Validation set metrics are continuous If no improvement is achieved within a certain number of epochs, the process will stop.

[0183] Input normalization: mapping pixels to... And fix the mapping method.

[0184] 3.3.4.3 Basic diffusion loss

[0185] The noise prediction loss commonly used in diffusion models is:

[0186] .

[0187] 3.3.4.4 Introducing auxiliary loss based on tag controllability

[0188] To strengthen "input" "Can control output", can add consistency loss (through) Evaluation of generated results):

[0189] ;

[0190] in: Depend on get; From conditions ; Weights (preferred) ).

[0191] Total loss of the generative model:

[0192] ;

[0193] in: Preferred Adjustments are made based on the amount of data and controllable requirements.

[0194] 3.3.5 Key Parameter Selection and Recommended Range

[0195] To ensure practical implementation, the following default configurations and optional ranges are provided:

[0196] 3.3.5.1 Diffuse Noise Schedule and Sampling

[0197] Diffusion Steps: Training Step (or );

[0198] Schedule: Linear Or cosine schedule;

[0199] Inference sampling: DDIM 20-50 steps (speed / quality trade-off).

[0200] 3.3.5.2 Generative Model Network scale

[0201] U-Net hierarchy: 4 levels (feature map size) );

[0202] Basic channels: 64 (for smaller data volumes) or 128 (for larger data volumes);

[0203] Attention: and Add 1-2 attention blocks to the scale;

[0204] Conditional vector dimension: or .

[0205] 3.3.5.3 Optimizer and Training

[0206] Optimizer: AdamW;

[0207] Learning rate: (Fine-tuning) or (Start training from the beginning);

[0208] Batch size: (Diffusion model has higher memory requirements);

[0209] Number of training rounds: (Fine-tuning) or higher (start training from scratch);

[0210] Gradient clipping: ;

[0211] Mixed precision: FP16 / BF16 (recommended).

[0212] 3.3.6 Dataset Usage and Prompt Word Templates

[0213] 3.3.6.1 Tag serialization into prompt words

[0214] Input prompts during the inference phase can use a fixed template, for example:

[0215] "Tire tread pattern, type = " Pitch length = mm, pitch width = mm, sea-to-land ratio = Pitch arrangement = Noise = dB, total side length = mm".

[0216] During training, the same template can be used as conditional text input (if a text encoder is used), or the numerical conditional vectors can be directly input (if a text encoder is not used). Either method can be used for deployment.

[0217] 3.3.6.2 Handling Missing Conditions During Training

[0218] To enhance robustness, some conditions can be randomly masked (e.g., masking with a 10% probability). or The trained model can still generate reasonable patterns even under missing conditions; if the user does not provide [information] during inference... The model will then be generated by default according to type / pitch.

[0219] 3.4 Step S4: Regular Projection

[0220] The purpose of rule projection is to constrain potentially conflicting or infeasible objectives. Transform into legal constraints This avoids oscillations, divergence, or the generation of unmanufacturable structures caused by unattainable goals during the generation phase.

[0221] Regular projections must include at least:

[0222] Encoding validity: Ensure Each element in the set belongs to an allowed set (e.g., ),length Within the permitted range;

[0223] Consistency with pixel ratio: If the target If there is a significant conflict with the target image resolution / pitch size combination, then... Make trimmings or suggest adjustments;

[0224] Consistency with period: Ensure pitch length / width and pitch element number The combination of these can form a reasonable cycle;

[0225] Consistency with frequency band definition: Ensure target harmonic upper limit and number of segments The frequency band allocation must be consistent; otherwise, adjust. Or fill in the missing dimensions.

[0226] Projection output Then, the closed-loop generation and correction steps begin.

[0227] 3.5 Step S5: Closed-loop generation and iterative correction of second-order differentiable projection

[0228] like Figure 3 As shown, step S5 is the key difference between this invention and "generating pattern images directly based solely on prompts": it not only generates images, but also automatically corrects deviations during the generation process, so that the results simultaneously meet the requirements of noise-related segmented harmonic indices, land-sea ratio, boundary complexity, and millimeter-scale manufacturability, and can output manufacturable vector contours.

[0229] 3.5.1 Target Constraints and Stopping Conditions

[0230] The target constraints include:

[0231] Piecewise harmonic constraints: ;

[0232] Sea-to-land ratio constraint: ;

[0233] Total side length constraint: ;

[0234] Manufacturing constraints: trench width , width of rubber bridge Area of ​​isolated blocks Corner radius Connectivity meets the standard.

[0235] When all the above constraints are satisfied, or the preset maximum number of iterations is reached. Stop when stopped. Output when stopped. (Output the optimal solution when the constraints are satisfied; output the candidate solution with the lowest violation degree when the constraints are not fully satisfied and record the reason for the violation for engineers to review.)

[0236] 3.5.2 Basic Idea of ​​Second-Order Differentiable Projection

[0237] The generation process can be viewed as searching for solutions that satisfy the constraints in the latent variable space (or diffusion inversion space). For each generated candidate... ,Depend on Output And construct the correction objective function:

[0238] ;

[0239] in: The objective function for iterative correction; As weight; For the first Upper limit of segmented harmonics; other symbols are defined as before.

[0240] By generating latent variables (or the noise variable or conditional injection vector of the diffusion model) undergo gradient descent or approximate projection to make The value is gradually reduced to approximate the pattern that satisfies the constraints.

[0241] 3.5.3 Priority Update Strategy of "Sequence First, Geometry Later"

[0242] Tire noise is highly correlated with the pitch sequence; therefore, this invention adopts a priority strategy of "adjusting the pitch first, then adjusting the geometric details."

[0243] 1) If the main violation is a piecewise harmonic ( (Exceeding limits), priority will be given to Perform local swaps, local rearrangements, or code substitutions within the allowed set (e.g., replacing some "3"s with "2"s and compensating in adjacent positions) to reduce the harmonic peak values ​​of the corresponding segments;

[0244] 2) If the main violation is the sea-to-land ratio deviation, then priority should be given to guidance. Make global scale adjustments to the trench width / block fill;

[0245] 3) If the main violation is the total side length deviation, the boundary complexity should be adjusted by reducing / increasing the number of small cutting lines, merging small boundaries, or increasing the density of cutting lines.

[0246] 4) If the main violation is a manufacturing threshold (such as too narrow glue bridge or too small isolated block), it should be handled by local morphology repair (thickening glue bridge, filling small holes, and rounding sharp corners) while keeping the macroscopic target intact.

[0247] The above strategy avoids the structural noise and instability caused by "relying on image pixel-level gradient updates for all violations", making the iteration more controllable and more in line with engineering intuition.

[0248] 3.5.4 Details of Differentiable Approximation Implementation of Manufacturing Constraints

[0249] 1) Trench width and adhesive bridge width

[0250] The range field is approximated by a differentiable distance transformation of the binary image. and , representing the distance from the sea pixel to the boundary and the distance from the land pixel to the boundary, respectively. The local minimum width can be approximated as... The local minimum. The violation is constructed after converting the millimeter threshold to a pixel threshold:

[0251] ;

[0252] in: For the corresponding direction, pixels to millimeter scale (optional) or (or a combination thereof) This represents the statistical average or quantile for a sea area. The same logic applies to the Jiaoqiao area. .

[0253] 2) Area of ​​isolated blocks

[0254] This can be approximated by soft connected components: Differentiable smoothing and thresholding of the land region yields a connected response map; then, area penalties are applied to small connected response regions to form... In engineering, non-differentiable connected component analysis can be used to identify violation locations, and then a differentiable local filling loss can be applied to the violation locations (this combination of "non-differentiable detection + differentiable repair" is also an optional implementation of this invention).

[0255] 3) Corner radius and sharp corner

[0256] After boundary extraction, estimate local curvature or use corner detection to obtain a sharp corner probability map. If the sharp corner radius is less than... Then apply a smoothing / rounding loss to this area to form .

[0257] 4) Connectivity

[0258] For trench connectivity, a "circumferential connectivity probability" index can be constructed: start and end boundaries are set in the circumferential direction, the soft probability of a connectivity path existing in the sea area is calculated and compared with a threshold, and the result is obtained. In engineering, differentiable connectivity indices can be obtained through multi-scale skeletonization approximation.

[0259] 3.5.5 Output Vector Profile Verification of processing accessibility

[0260] When the final pattern is obtained Then, convert it into a vector profile. :

[0261] 1) Closed contours are obtained using contour tracing (marching squares) or boundary tracing;

[0262] 2) Simplify the contour (Douglas–Peucker) to reduce redundant points;

[0263] 3) Round off sharp corners to ensure the minimum fillet radius is not less than [amount missing]. ;

[0264] 4) Check tool accessibility (e.g., tool radius) satisfy (Or check according to process rules) to ensure it is processable.

[0265] IV. Further Constraints on the LLM Model

[0266] This section corresponds to the further improvement and introduction of a large language model. Its function is not to replace or Instead, it enhances the engineering usability and stability of "constraint generation, conflict resolution, and closed-loop hints".

[0267] 4.1 LLM Semantic Parsing Generates Target Constraints

[0268] When a user inputs their natural language requirements (such as "Make an all-season tire with a land-sea ratio of 0.72, lower noise, and a pitch width of around 150mm"), LLM... Parse the text into structured constraints At least includes One or more of the following, and can be completed. The candidate set.

[0269] 4.2 LLM Consistency Verification and Feasible Domain Repair

[0270] LLM Consistency checks can be performed in conjunction with a pre-defined "feasible domain rule base":

[0271] 1) If and If the combination makes it geometrically difficult to achieve (e.g., the land-sea ratio is too high but the trench width is required to be no less than a certain width), then it is recommended to adjust the design. Or relax a certain threshold;

[0272] 2) If and If it doesn't match, then a suggestion is needed. Reasonable range or reset ;

[0273] 3) If With the number of segments If the dimensions are inconsistent, then fill in the missing dimensions or adjust them. .

[0274] Output after repair Then enter the rule projection module to obtain .

[0275] 4.3 LLM generates and filters candidate interval sequences

[0276] When missing At that time, LLM based on Generate candidates ( ), and can invoke fast harmonic prediction (or call (Harmonic prediction capability) to select the closest The sequence as This reduces the cost of engineers manually testing sequences.

[0277] 4.4 LLM driver closed-loop correction prompt

[0278] When a deviation vector appears in step S5 Alternatively, when creating a map of violation locations, input it into the LLM, and the LLM will output "priority correction suggestions":

[0279] 1) Priority adjustment (If the main issue is harmonic over-limit);

[0280] 2) Prioritize adjusting the trench width / glue bridge (if the main issue is manufacturing violations);

[0281] 3) Prioritize adjusting the cutting line density (if the main issue is the total side length deviation). Then, translate the recommendations into updates to the conditions or iteration weights, thereby accelerating convergence and reducing invalid iterations.

[0282] V. Application Example 1: Pilot Demonstration of Controllable Generation of Tire Tread Noise in All Seasons

[0283] 1. Experimental Objective

[0284] Verify the invention in a given pattern type Pitch size Interval arrangement sequence and target noise value Under the given conditions, a pattern image is generated using a finely tuned image generation model. Through mold making, tire manufacturing, and bench / road testing, it was demonstrated that the measured tread noise can decrease in the target direction and the tread geometry label (land-sea ratio) can be improved. Total side length The input is consistent with the data.

[0285] 2. Test Subjects and Sample Preparation

[0286] Tire specifications: Taking passenger car all-season tires as an example (e.g., 205 / 55R16; specifications can be replaced according to the company's project).

[0287] Comparison of options:

[0288] Comparative Example 1 (Manual Baseline Tread Pattern): R&D engineers manually modified the existing all-season tire tread pattern to meet the requirements. Compared to basic land and sea targets, but without using the "pitch sequence + noise tag driven generation" of this invention.

[0289] Example 1 (Pattern Generation in this Invention): Patterns are generated according to steps S1–S5 of this invention and used for mold making and die making.

[0290] Number of prototype tires prepared for each protocol A single sample (with the same formula, the same vulcanization conditions, and the same tire structure) was used for noise statistical testing.

[0291] like Figure 4 As shown, the mold is implemented by generating... Binary pattern image Vectorization yields the contour. (Closed polyline / spline), rounded (minimum fillet radius) Post-processing mold pattern blocks / inserts; mold processing accuracy, surface roughness, and tool radius shall be performed in accordance with enterprise standards.

[0292] 3. Implement the steps (corresponding to steps S1–S5)

[0293] 3.1 Step 1: Performance Calculation (Noise and Total Side Length)

[0294] (1) Total side length calculate

[0295] For binary graphs Extracting Boundaries And convert it to millimeters:

[0296] ;

[0297] in: The length of the boundary pixels; This is the conversion factor from pixels to millimeters.

[0298] (2) Noise value Calculation (based on interval sequence)

[0299] Arrange the interval sequence Mapped to a circumferential excitation sequence Calculate the frequency band energy and map it to noise values:

[0300] ;

[0301] in: for The discrete Fourier amplitude; Preset frequency band (including block frequency and multiplier); This is the calibration constant.

[0302] Key points for project implementation: Alignment and calibration can be performed using "measured noise of known patterns". It is monotonically correlated with platform or road noise.

[0303] 3.2 Second step: Binarization and size unification of the pattern image

[0304] The original pattern images (CAD / vector / bitmap) are uniformly processed into... Black and white images It is agreed that white represents the land (rubber) area and black represents the trench (sea) area (or vice versa, but must be consistent).

[0305] 3.3 Third step: Tag construction

[0306] Create labels for each pattern image : ;in: Pattern type; Actual pitch length / width (mm); The land-sea ratio (the percentage of land area); This is a sequence of pitch arrangements (in this example, 1 / 2 / 3 represent the circumferential direction, respectively). (scaling by multiple times); Noise level (dB); The total side length of the pattern is (mm).

[0307] 3.4 Step 4: Fine-tune the training and save the model

[0308] Select The generative model was obtained by fine-tuning and training using images of patterns and labels. After training, save the model and label serialization rules (e.g., ...). Tokenize by string (Normalization).

[0309] 3.5 Step 5: Input prompt words and generate patterns

[0310] Input prompts: "Tire tread pattern, all-season tire, actual pitch length" Pitch width sea-land ratio Pitch arrangement noise value dB".

[0311] Model output generates patterns (with attached) Figure 6 The striped continuous pattern is consistent in shape, and after vectorization and manufacturing verification, it is used for mold making.

[0312] 4. Test conditions for actual noise measurement

[0313] 4.1 Indoor Rolling Noise Test Bench

[0314] Equipment: Indoor tire noise drum machine / rolling noise test bench.

[0315] Operating condition: speed ; load According to standard load Tire pressure Use recommended tire pressure; ambient temperature Measurements were taken after each tire was preheated for 20 minutes.

[0316] Indicator: A-weighted rolling noise (dB(A)), each tire was measured 3 times and the average was taken.

[0317] 4.2 Outdoor noise

[0318] Conduct a passing noise test on a standardized test surface (such as a noise-sensitive surface as specified in ISO / GB) to obtain the vehicle passing noise. This is used to cross-validate the trend consistency with bench test results.

[0319] 5. Experimental Results and Data

[0320] 5.1 Consistency Verification of Generated Pattern Labels

[0321] Pattern generated in Example 1 Posterior calculation yielded:

[0322] Land-sea ratio: (Target ,deviation );

[0323] Total side length: ;

[0324] Interval sequence consistency: Completely identical to the input (either output from the token or obtained from parsing after generation).

[0325] 5.2 Noise Measurement Results

[0326] Table 1 presents the measured indoor rolling noise results for Comparative Example 1 and Example 1 (for each group) The value is the mean. (Standard deviation).

[0327] Table 1 Comparison of measured indoor rolling noise levels

[0328]

[0329] Adopting a double-sided Test, Example 1 compared to Comparative Example 1 Significant difference ( This demonstrates that the pattern generated by the present invention can achieve stable noise reduction while maintaining the structural and dimensional constraints of all-season tires.

[0330] Figure 5 The training set contains typical seasonal tire pitch block pattern samples with clear block boundaries, suitable for calculation. And establish a "segment sequence" -noise "Mapping; Figure 6 The strip pattern generated after fine-tuning under the given prompt is shown in Table 1. The main groove and the cutting line are continuous, which meets the manufacturability requirements, and the noise is measured to decrease.

[0331] 6. Conclusion on Technical Effectiveness

[0332] As shown in Table 1, compared to the artificial reference pattern without the present invention (Comparative Example 1), the present invention, through "pattern image binarization and structured labeling (including pitch arrangement)", achieves better results. ,noise Total side length The method of "fine-tuning the large image generation model" enables the generative model to obtain the correspondence between the target constraints and the pattern structure. After inputting prompts such as target noise and pitch sequence, it can directly generate pattern images that meet the requirements of structural continuity and manufacturability, and achieve approximately [percentage missing] in indoor rolling noise tests. The stable reduction of dB(A) and the good consistency between geometric labels such as land-sea ratio and target significantly reduce the number of times it is necessary to manually modify drawings, conduct simulations and prototypes, shorten the pattern development cycle and improve design consistency.

[0333] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0334] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0335] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0336] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0337] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0338] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0339] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0340] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A method for automatically generating tire tread patterns based on an LLM model, characterized in that, The method includes the following steps: S1, obtain tire tread pattern samples, process each sample into a black and white image and generate a structured label; the structured label includes at least the tread pattern type, pitch length, pitch width, land-sea ratio, pitch arrangement sequence, segmented harmonic index and total tread pattern side length. S2, establish a pixel-to-millimeter mapping relationship based on pitch length and pitch width, and set manufacturing thresholds, the manufacturing thresholds including at least minimum trench width threshold, minimum adhesive bridge width threshold, minimum isolated block area threshold and minimum rounded corner radius threshold; S3 trains the conditional generation model and the differentiable constraint evaluation model, enabling the conditional generation model to generate candidate pattern images under the generation conditions consisting of pattern type, pitch length, pitch width, land-sea ratio, pitch arrangement sequence, piecewise harmonic index, and total pattern side length. The differentiable constraint evaluation model outputs the predicted piecewise harmonic index vector, predicted land-sea ratio, predicted total pattern side length, and manufacturing violation degree for the candidate pattern images, and is used to construct the differentiable constraint loss. S4: Receive the target constraint set and perform rule projection to obtain legal constraints; S5 generates candidate pattern images under legal constraints using a conditional generation model. Based on the evaluation results of the candidate pattern images using a differentiable constraint evaluation model, it judges the piecewise harmonic constraints, land-sea ratio constraints, total side length constraints, and millimeter-scale manufacturing constraints. It iteratively updates the latent variables and / or generation conditions until they are satisfied, and then outputs the final pattern image. The final pattern image is then extracted into vector contour data for mold making or processing. In S3, the differentiable constraint loss includes at least piecewise harmonic loss, land-sea ratio loss, total side length loss and manufacturing violation loss, wherein the manufacturing violation loss includes at least trench width violation, glue bridge width violation and connectivity violation, and the manufacturing violation loss is constructed as a differentiable quantity through distance transformation field and / or skeleton approximation field. In S4, the regular projection includes at least: a) Verify the validity of the encoding of the interval arrangement sequence and replace it with the nearest valid sequence if it is invalid; b) Verify the consistency between the target land-sea ratio and the target black-and-white pixel ratio, and correct the target land-sea ratio or the target pixel ratio if the deviation is out of tolerance; c) Verify the consistency of the target pitch length, target pitch width, number of pitch elements, and pitch period. If there is a discrepancy, trim and remap the target pitch length, target pitch width, or number of pitch elements to obtain a valid constraint.

2. The method according to claim 1, characterized in that, In S1, the number of samples is greater than 100. Each sample is processed into a black and white image with a resolution of 512*512. In the black and white image, the first pixel value represents the land area of ​​the rubber material and the second pixel value represents the sea area of ​​the trench. The land-sea ratio is calculated by converting the proportion of black and white pixels. And / or, the pitch arrangement sequence is composed of discrete codes of length equal to the number of pitch units; each code is taken from set 1, 2, 3, where codes 1, 2, 3 respectively represent the corresponding pitch units scaled by 0.8 times, 1.0 times, and 1.2 times in the circumferential length direction; And / or, the segmented harmonic index is an index set consisting of multiple circumferential segmented harmonic values, wherein the number of circumferential segments is not less than 3; each circumferential segmented harmonic value is obtained by mapping the pitch arrangement sequence to the circumferential excitation sequence and performing Fourier transform on the corresponding circumferential segment, and the preset frequency band includes block frequency and at least one harmonic frequency band. And / or, the total side length of the pattern is the sum of the perimeters of the boundary between the land and the ditch, and the total side length of the pattern is obtained by extracting the boundary pixel length from the black and white image, converting it to millimeters from the pixel-to-millimeter mapping relationship, and then summing them.

3. The method according to claim 1, characterized in that, In S2, the pixel-to-millimeter mapping relationship includes at least circumferential scale mapping and lateral scale mapping, which are obtained by combining the pitch length and pitch width with the image pixel size, respectively, and are used to convert the groove width, adhesive bridge width, area and corner radius into millimeter scale; And / or, the millimeter-scale manufacturing constraints include at least: the trench width is not less than the minimum trench width threshold, the adhesive bridge width is not less than the minimum adhesive bridge width threshold, the isolated block area is not less than the minimum isolated block area threshold, the minimum fillet radius is not less than the minimum fillet radius threshold, and the trench connectivity meets the preset connectivity requirements.

4. The method according to claim 1, characterized in that, In S3, the conditional generation model is one of a diffusion model, a generative adversarial network, or an autoregressive generation model; the differentiable constraint evaluation model is one of a convolutional neural network or a visual Transformer network; and the candidate pattern image is the pattern image output by the conditional generation model under the generation conditions.

5. The method according to claim 1, characterized in that, In S5, the segmented harmonic constraint is that the predicted segmented harmonic index vector is not higher than the target segmented harmonic upper limit, the land-sea ratio constraint is that the difference between the predicted land-sea ratio and the target land-sea ratio does not exceed the land-sea ratio tolerance, and the total side length constraint is that the difference between the predicted total side length of the pattern and the target total side length does not exceed the total side length tolerance. And / or, in S5, the iterative update includes: when the predicted segmented harmonic index vector exceeds the target segmented harmonic upper limit, firstly perform local exchange, local rearrangement or scaling encoding replacement on the pitch arrangement sequence to reduce the corresponding segmented harmonic peak value, and then perform gradient update based on differentiable constraint loss on the generated latent variables. And / or, the iteration update of S5 ends when the following stopping conditions are met: the predicted segmented harmonic index vector is not higher than the target segmented harmonic upper limit, the difference between the predicted land-sea ratio and the target land-sea ratio does not exceed the land-sea ratio tolerance, the difference between the predicted total side length of the pattern and the target total side length of the pattern does not exceed the total side length tolerance, and the manufacturing violation degree is 0; or it ends when the preset iteration limit is reached and the current optimal pattern image is output as the final pattern image. And / or, in S5, the vector contour data is a set of closed polylines or spline curves, and before output, the sharp corners are rounded and the tool radius accessibility is checked, so that the minimum fillet radius in the vector contour data is not less than the minimum fillet radius threshold.

6. The method according to claim 1, characterized in that, The method further includes: receiving natural language design intent text, calling a large language model to parse the natural language design intent text into a set of target constraints, wherein the set of target constraints includes at least one or more of the following: target pattern type, target pitch length, target pitch width, target land-sea ratio, target pitch arrangement sequence, target segmented harmonic upper limit, and target pattern total side length, and the set of target constraints is used as the input for the rule projection in step S4; And / or, after outputting the target constraint set, the large language model performs a consistency check on the target constraint set based on a preset tire design feasible domain rule base and outputs repair constraints. The consistency check includes at least: a) Validity of the target pitch sequence encoding; b) The target land-sea ratio is consistent with the pixel ratio; c) The target pitch length, target pitch width, number of pitch elements, and pitch period are consistent; d) The target segmented harmonic upper limit is consistent with the preset frequency band definition; And replace the target constraint set with the repair constraint in step S4; And / or, when the target constraint set lacks a target pitch arrangement sequence or the target pitch arrangement sequence is incomplete, the large language model generates a candidate pitch arrangement sequence set based on the target pattern type, target pitch length, target pitch width, target land-sea ratio, and target segmented harmonic upper limit. The candidate pitch arrangement sequence set contains no less than two candidate pitch arrangement sequences. The model then calls the pitch harmonic prediction or the differentiable constraint evaluation model to output predicted segmented harmonic indices for each candidate sequence, in order to filter and obtain the target pitch arrangement sequence for the target constraint set. And / or, when the predicted segmented harmonic index vector, the predicted land-sea ratio, or the predicted total side length of the pattern in step S5 does not meet the corresponding target constraints, the deviation of the predicted segmented harmonic index vector relative to the target segmented harmonic upper limit, the deviation of the predicted land-sea ratio relative to the target land-sea ratio, the deviation of the predicted total side length of the pattern relative to the target total side length of the pattern, and the location map of manufacturing violations are input into the large language model. The large language model outputs constraint correction suggestions to update at least one of the target land-sea ratio, the target total side length of the pattern, the target segmented harmonic upper limit, or the target pitch arrangement sequence, and triggers the re-execution of steps S4 and S5.

7. A tire tread pattern automatic generation system, characterized in that, The system includes a processor and a memory, wherein the memory stores instructions that, when executed by the processor, implement the method described in any one of claims 1 to 6; the system includes at least: a sample and label construction module, a calibration and threshold setting module, a conditional generation model training and generation module, a differentiable constraint evaluation model training and evaluation module, a rule projection module, a second-order differentiable projection iterative correction module, and a vector contour output module.

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

9. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 6.

Images