AI-based stuffed toy design method

Abstract

The invention is suitable for the technical field of stuffed toy design control, and provides an AI-based stuffed toy design method which is applied to an AI-based stuffed toy design control system. The system comprises a data acquisition module, a model construction module, a model training module, a model evaluation module, a scheme determination module, a trial production verification module, a wireless communication module, a memory, an alarm, a processing center and an intelligent mobile terminal. The data acquisition module, the model construction module, the model training module, the model evaluation module, the scheme determination module, the trial production verification module and the wireless communication module are respectively connected with the processing center. According to the method, the AI model is constructed and trained by collecting daily plush toy quality data, and a design scheme is matched according to the model, so that various quality abnormalities are prevented, and consistency with customer samples is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of plush toy design control, and particularly to an AI-based plush toy design method in the field of artificial intelligence. Background Art

[0002] The design of plush toys includes processes such as requirement collection, plan formulation, sample making and testing, and small batch verification. However, there are often significant differences between the produced plush toys and the customer samples or drawings, which are all due to the lack of strict control over design links such as fabric selection, detail processing, process flow, parameter control, and their corresponding sample making, resulting in inconsistencies between sample making, trial production, and subsequent mass production. Therefore, strict process control must be carried out to improve the process control ability to ensure that the quality of plush toys meets customer requirements. Summary of the Invention

[0003] In order to solve the deficiencies in the prior art, the purpose of the present invention is to provide an AI-based plush toy design method. By setting up various functional modules such as a data acquisition module, a model construction module, a model training module, a model evaluation module, a scheme determination module, and a trial production verification module, the quality data of daily plush toys is collected to construct and train an AI model, and the design scheme is matched according to the model to prevent various quality anomalies, realizing the improvement of the full-link efficiency from creative design to mass production.

[0004] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0005] An AI-based plush toy design method is applied to an AI-based plush toy design control system. The system includes a data acquisition module, a model construction module, a model training module, a model evaluation module, a scheme determination module, a trial production verification module, a wireless communication module, a memory, an alarm, a processing center, and an intelligent mobile terminal; the data acquisition module, the model construction module, the model training module, the model evaluation module, the scheme determination module, the trial production verification module, the wireless communication module, the memory, and the alarm are respectively connected to the processing center; the intelligent mobile terminal includes a smart phone, a tablet computer, and a smart remote control, and is wirelessly connected to the wireless communication module within the range of a wireless network or the Internet.

[0006] An AI-based plush toy design method provided by the present invention includes the following steps:

[0007] S10. Before trial production, the data acquisition module acquires the historical quality data of similar plush toys in terms of vision, user interaction, market and design preferences, synthesis and enhancement, and function verification, preprocesses and processes it for subsequent model construction, and transfers it to the model construction module;

[0008] S20. The model construction module extracts key features, designs models, integrates and optimizes, initializes parameters, dynamically optimizes, fuses and deploys models based on the verified data to build a model for subsequent model training and verification, and transmits it to the model training module;

[0009] S30. The model training module trains the model according to the collected data and adjusts the model parameters to optimize the model performance for a predetermined plush toy design scheme, ensuring the accuracy of subsequent proofing quality, and transmits it to the model evaluation module;

[0010] S40. The model evaluation module verifies and evaluates the model according to the trained model, and adjusts and optimizes the model according to the verification results to improve the performance and accuracy of the model, and transmits it to the scheme determination module;

[0011] S50. The scheme determination module conducts a proofing test according to the plush toy design scheme predicted by the trained model to obtain the best plush toy design scheme, ensuring the accuracy of subsequent trial production quality, and transmits it to the trial production verification module;

[0012] S60. The trial production verification module conducts automatic small-batch trial production and quality inspection through plush toy automatic equipment according to the set plush toy design scheme to ensure the accuracy and stability of subsequent mass production quality, and transmits it to the processing center;

[0013] S70. The processing center compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: if it meets the standard, it notifies that formal mass production can be carried out; if it does not meet the standard, it transmits it to the alarm and notifies rework or scrapping.

[0014] The system of the present invention further includes a computer-readable storage medium containing a memory; the memory stores a computer program, and when each functional module executes the computer program, the steps of an AI-based plush toy design method described in any one of the above are implemented; a computer program is stored on the computer-readable storage medium, and when the computer program is executed by each functional module, the steps of an AI-based plush toy design method described in any one of the above are implemented; it further includes an AI-based plush toy design control device implemented by the AI-based plush toy design method described above.

[0015] Advantages of the present invention compared with the prior art:

[0016] By setting each functional module, collecting daily plush toy quality data to build and train an AI model, and matching the design scheme according to the model, full-automatic production and artificial intelligence control are realized to prevent various quality anomalies and ensure consistency with customer samples or drawings. (For specific beneficial effects, see the content description in the subsequent "Detailed Description of the Embodiment"). Description of the Drawings

[0017] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or exemplary technical descriptions. The drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 It is a schematic diagram of the system module of the present invention;

[0019] Figure 2 It is a schematic diagram of the data acquisition module of the present invention;

[0020] Figure 3 It is a schematic diagram of the model construction module of the present invention;

[0021] Figure 4 It is a schematic diagram of the model training module of the present invention;

[0022] Figure 5 It is a schematic diagram of the model evaluation module of the present invention;

[0023] Figure 6 It is a schematic diagram of the solution determination module of the present invention;

[0024] Figure 7 It is a schematic diagram of the trial production verification module of the present invention;

[0025] Figure 8 It is a schematic diagram of the method process control program of the present invention;

[0026] Figure 9 It is a schematic diagram of the program of step S10 in the method process of the present invention;

[0027] Figure 10 It is a schematic diagram of the program of step S20 in the method process of the present invention;

[0028] Figure 11 It is a schematic diagram of the program of step S30 in the method process of the present invention;

[0029] Figure 12 It is a schematic diagram of the program of step S40 in the method process of the present invention;

[0030] Figure 13 It is a schematic diagram of the program of step S50 in the method process of the present invention;

[0031] Figure 14 It is a schematic diagram of the program of step S60 in the method process of the present invention. Specific embodiments

[0032] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] The following describes the specific implementation of the present invention in detail with reference to specific embodiments:

[0034] In order to make the technical problems to be solved, technical solutions, and beneficial effects of the present application more clearly understood, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. It should be noted that when a module is said to be "disposed on" another module, it can be directly on the other module or indirectly on the other module. When a module is said to be "connected to" another module, it can be directly connected to the other module or indirectly connected to the other module. In the description of the present application, the meaning of "a plurality" is two or more, unless otherwise specifically defined. The meaning of "several" is one or more, unless otherwise specifically defined.

[0035] In the description of the present application, it should be noted that unless otherwise clearly defined and limited, the terms "connected" and "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to specific circumstances. The terms "including", "comprising", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0036] Please refer to Figure 1 As shown, the present invention provides an AI-based plush toy design method, which is applied to an AI-based plush toy design control system. The system includes a data acquisition module, a model construction module, a model training module, a model evaluation module, a scheme determination module, a trial production verification module, a wireless communication module, a memory, an alarm, a processing center, and an intelligent mobile terminal. The data acquisition module, the model construction module, the model training module, the model evaluation module, the scheme determination module, the trial production verification module, the wireless communication module, the memory, and the alarm are respectively connected to the processing center. The intelligent mobile terminal includes a smart phone, a tablet computer, and a smart remote control, and is wirelessly connected to the wireless communication module within the range of a wireless network or the Internet.

[0037] The wireless communication module is provided with a wireless network unit, which is responsible for the transceiver of wireless signals and automatically forms a network connection with intelligent mobile terminals within an effective network range. The wireless signals include various Internet of Things signals such as MQTT, CoAP, HTTP, REST API, Zagbee, LoRaWAN, NB-IoT, Bluetooth, etc., or 5G, 4G, Wifi network signals. The wireless communication module is effectively connected to each production detection device and is provided with communication protocols such as Bluetooth and Wi-Fi, supporting small program network configuration and APP networking.

[0038] The alarm compares the actual evaluation score of the model with the model evaluation score standard stored in the memory. If it does not meet the standard, it will automatically give a sound alarm and notify to continue training; and compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory. If it does not meet the standard, it will automatically give a sound alarm and notify to adjust the parameters; and compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory. If it does not meet the standard, it will automatically give a sound alarm and notify for repair or upgrade.

[0039] The memory is responsible for storing the information of the data acquisition module, model construction module, model training module, model evaluation module, scheme determination module, trial production verification module, wireless communication module, and alarm, as well as storing the model evaluation score standard and the plush toy quality standard.

[0040] The processing center is responsible for the information transfer of the data acquisition module, model construction module, model training module, model evaluation module, scheme determination module, trial production verification module, wireless communication module, alarm, and memory. It is the hub center of the system, and compares the actual evaluation score of the model with the model evaluation score standard stored in the memory: if it meets the standard, it is the designed scheme for the predetermined plush toy; if it does not meet the standard, it is transferred to the alarm and notifies to continue training; and compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: if it meets the standard, it is set as the plush toy design scheme; if it does not meet the standard, it is transferred to the alarm and notifies to adjust the parameters; and compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: if it meets the standard, it notifies that mass production can be carried out; if it does not meet the standard, it is transferred to the alarm and notifies to rework or scrap.

[0041] Please refer to Figure 2 As shown, the data acquisition module includes an information collection unit, a cleaning and repair unit, a segmentation and extraction unit, a standard enhancement unit, a conversion and encoding unit, and a physical verification unit, which is responsible for obtaining the historical quality data of visual, user interaction, market and design preferences, synthesis and enhancement, and functional verification of similar plush toys, preprocessing and processing, and transferring it to the model construction module.

[0042] Furthermore, the information collection unit obtains various normal and abnormal quality data and image videos of similar plush toys by matching with the industry database according to the customer's needs, and transmits them to the cleaning and restoration unit; the cleaning and restoration unit performs noise elimination, blur restoration, local defect image restoration, 3D model restoration, highlight area processing, and texture consistency optimization, and transmits them to the segmentation and extraction unit; the segmentation and extraction unit performs semantic segmentation and main body separation, reflection and highlight suppression, binarization and edge detection, contour restoration and filling, and cross-modal semantic alignment, and transmits them to the standard enhancement unit; the standard enhancement unit performs size normalization, color space conversion, histogram equalization, binarization processing, geometric transformation enhancement, material and illumination perturbation, and transmits them to the conversion and encoding unit; the conversion and encoding unit performs numerical data standardization, feature encoding, data specification and feature extraction, continuous feature binning, and synthetic feature generation, and transmits them to the physical verification unit; the physical verification unit can avoid physical contradiction design and process conflicts through material parameter verification, mechanical simulation test, ergonomic adaptation, and process constraint mapping.

[0043] Further explanation, the quality data of the plush toy includes: 1) Visual data: image data such as entity photos, line drawings, 3D models, material and texture data such as fluff texture, color gradient, and light and shadow effects, and dynamic interaction data of the user and toy interaction scene; 2) User interaction data: voice data of the user's conversation with the toy, tactile feedback data such as the strength and frequency of the user touching the toy; 3) Market and design preference data: IP element data such as the character image and color scheme of the IP, and user preference data of consumers for the style of plush toys; 4) Synthetic and enhanced data: virtual generation data of toy variants, and annotation data obtained by labeling the collected picture and voice data; 5) Function verification data: behavior log data of abnormal feedback during the use of the toy, and emotional response data obtained by simulating the user's emotions to test the response logic of the toy.

[0044] Please refer to Figure 3 As shown, the model construction module includes a feature extraction unit, a model design unit, an integration and optimization unit, a parameter initialization unit, a dynamic optimization unit, and a network layer unit. It performs key feature extraction, model design, integration and optimization, parameter initialization, dynamic optimization, model fusion and deployment based on the verified data to build a model, and transmits it to the model training module.

[0045] Further, the feature extraction unit performs invalid data filtering, physical attribute extraction, user preference feature encoding, key feature screening, dimensionality compression, and data hierarchical partitioning, and then transfers the results to the model design unit; the model design unit applies generative adversarial networks, fuses multi-modal inputs, designs generator networks, optimizes discriminator networks, and applies transfer learning, and then transfers the results to the integrated optimization unit; the integrated optimization unit embeds physical constraints, performs regularization optimization, enhances interpretability, compresses the model, and conducts quantization-aware training, and then transfers the results to the parameter initialization unit; the parameter initialization unit performs pre-trained transfer, pre-trains word vectors, initializes randomly, initializes according to a distribution, constrains material parameters, and enhances style features, and then transfers the results to the dynamic optimization unit; the dynamic optimization unit configures residual connection parameters, fuses regularization, fine-tunes pre-trained models, differentiates by module, tests initialization sensitivity, and adapts to the production environment, and then transfers the results to the network layer unit; the network layer unit obtains a convolutional neural network model architecture with an input layer, convolutional layers, activation layers, pooling layers, and fully connected layers based on feature and parameter optimization for subsequent model training.

[0046] Please refer to Figure 4 As shown, the model training module includes a convolutional input unit, a pooling conversion unit, a fully connected layer unit, a normalization output unit, an output conversion unit, and a backpropagation unit. It trains the model based on the collected data, adjusts the model parameters to optimize the model performance for a predetermined plush toy design scheme, and then transfers the results to the model evaluation module.

[0047] Further, the convolutional input unit obtains the convolutional input value according to the convolutional input calculation formula "z(t) = ∫x(m)y(t - m)dm" and inputs it to the activation function, and then transfers the results to the pooling conversion unit; the pooling conversion unit obtains the maximum pixel value after pooling according to the average pooling calculation formula "Z(I,j) = mean(X[i*Ps(i + 1)*Ps,j*Ps(j + 1)*Ps])", and then transfers the results to the fully connected layer unit; the fully connected layer unit obtains the output data of the connection layer according to the fully connected layer calculation formula "y = f(∑(Wn*Xn)+b)", and then transfers the results to the normalization output unit; the normalization output unit obtains the Batch Norm output according to the Batch Norm layer calculation formula "h = φ(Bn(Wx + b))", and then transfers the results to the output conversion unit; the output conversion unit obtains the size of the output value after convolution according to the output layer conversion calculation formula "N = (P - F + 2C) / S + 1", and then transfers the results to the backpropagation unit; the backpropagation unit obtains the backpropagation of the convolutional output value according to the backpropagation calculation formula "dz[l] = da[l]*g[l]′(z[l])" to optimize the performance of the network model.

[0048] Please refer to Figure 5As shown, the model evaluation module includes a classification and marking unit, a learning and training unit, a cost verification unit, and an application evaluation unit. It verifies and evaluates the model based on the trained model, adjusts and optimizes the model according to the verification results, and transfers it to the solution determination module.

[0049] Furthermore, the classification and marking unit divides and marks the corresponding data sample categories according to the abnormal categories of the plush toy design quality, and transfers them to the learning and training unit; the learning and training unit determines the training model according to the gradient descent data, inputs the plush toy quality data in the training set into the model for simulation training, and transfers it to the cost verification unit; the cost verification unit obtains the cost function value after model training according to the cost function calculation formula

[0050] "J(θ) = -1 / m ∑m ∑K [y k (i) log(h θ (x(i))k) + (1 - y k (i)) log(1 - (h θ (x(i))k)] + λ / 2m

[0051] ∑L-1 ∑sl ∑sl+1 (θj,i(l))2" and transfers it to the application evaluation unit; the application evaluation unit obtains the comprehensive evaluation score F according to the model evaluation score calculation formula "F = 2(Ac * Re) / (Ac + Re)" and transfers it to the processing center.

[0052] Please refer to Figure 6 As shown, the solution determination module includes a fabric usage unit, a filling weight unit, a seam strength unit, a joint torque unit, a cost accounting unit, a proofing verification unit, and an abnormality identification unit. It conducts a proofing test on the plush toy design solution predicted by the trained model to obtain the best plush toy design solution and transfers it to the trial production verification module.

[0053] Further, the fabric usage unit obtains the fabric usage of the plush toy according to the plush toy fabric usage calculation formula "So = ∑ (Si * Ni) / η", and transmits it to the filling weight unit; the filling weight unit obtains the plush toy filling weight according to the plush toy filling weight calculation formula "Go = V * ρ * (1 + k)", and transmits it to the joint strength unit; the joint strength unit obtains the plush toy joint strength according to the plush toy joint strength calculation formula "F = S * (L / A) * ωo", and transmits it to the joint torque unit; the joint torque unit obtains the plush toy joint strength according to the plush toy movable joint torque calculation formula "τ = ρ * V * g * L * ωa". The plush toy can be moved by the joint torque, and is transmitted to the cost accounting unit; the cost accounting unit obtains the total cost of the plush toy according to the plush toy cost calculation formula "Ct = [(Pf*So) + (Ps*Go) + Ca] * (1 + μ) + Cb + Cm", and transmits it to the proofing verification unit; the proofing verification unit performs a proofing test according to the plush toy design scheme predetermined by the model and its predetermined process parameters and obtains the quality information of the plush toy, and transmits it to the processing center; the abnormality recognition unit matches the plush toy production image acquired in real time with the corresponding image in the trained convolutional neural network model and confirms its abnormal category for the purpose of proofing parameter adjustment.

[0054] See also Figure 7 As shown, the trial production verification module includes a parameter setting unit, a cutting and sewing unit, a filling and shaping unit, and a quality control unit. According to the set plush toy design plan, automatic small-batch trial production and quality inspection are carried out through plush toy automatic equipment, and the products are transmitted to the processing center.

[0055] Furthermore, the parameter setting unit solidifies and synchronizes the core parameters, performs reverse engineering calibration, and presets the multimodal quality inspection parameters according to the set process parameters to ensure the standardization of the trial production and the large-scale flexible production requirements, and transmits them to the cutting and sewing unit; the cutting and sewing unit performs high-precision laser cutting and AI-assisted defect pre-inspection, automated sewing execution and real-time quality monitoring through a laser cutting bed, and transmits them to the filling and shaping unit; the filling and shaping unit performs high-precision material injection, functional module integration, three-dimensional morphology optimization, dynamic shaping correction, real-time data feedback through an automatic filling machine, and transmits them to the collaborative maintenance unit; the collaborative maintenance unit ensures the stability of the trial production through a multi-machine collaborative network, system dynamic scheduling, data-driven maintenance, maintenance task automation, maintenance execution and optimization, and transmits them to the quality control unit; the quality control unit performs real-time calibration and error correction of process parameters, AI visual inspection, stress testing, traceability and improvement through quality monitoring equipment, and transmits them to the collaborative maintenance unit.

[0056] System operation working principle:

[0057] Before trial production, the data acquisition module first obtains the historical quality data of visual, user interaction, market and design preferences, synthesis and enhancement, and function verification of similar plush toys, preprocesses and processes it, and transmits it to the model construction module; secondly, the model construction module extracts key features, designs models, integrates and optimizes, initializes parameters, dynamically optimizes, fuses and deploys models based on the verified data to build a model, and transmits it to the model training module; thirdly, the model training module trains the model according to the collected data, adjusts the model parameters to optimize the model performance for a predetermined plush toy design scheme, and transmits it to the model evaluation module; then, the model evaluation module verifies and evaluates the model according to the trained model, adjusts and optimizes the model according to the verification results, and transmits it to the scheme determination module; then, the scheme determination module obtains the best plush toy design scheme through a proofing test based on the plush toy design scheme predicted by the trained model, and transmits it to the trial production verification module; finally, the trial production verification module conducts automatic small-batch trial production and quality inspection through plush toy automatic equipment according to the set plush toy design scheme, and transmits it to the processing center; the processing center compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: if it meets the standard, it notifies that formal mass production can be carried out, if it does not meet the standard, it transmits it to the alarm and notifies rework or scrapping.

[0058] When the operator or manager is outdoors or in a foreign place, they can use the intelligent mobile terminal to automatically form a network connection with the wireless communication module within the wireless network or the Internet, realizing the combination of the intelligent mobile terminal, the Internet and the Internet of Things technology. Through the APP software on the intelligent mobile terminal or the remote control software on the computer, they can control the automated production of plush toys at close range or remotely. The manager can remotely control and monitor the production quality of plush toys, view the daily operation status, data change status, equipment activity logs, etc., realize remote connection and monitor the improvement of the production quality of plush toys, improve production efficiency and reduce production costs.

[0059] Please refer to Figure 8 As shown, a method for designing plush toys based on AI provided by the present invention includes the following steps:

[0060] S10. Before trial production, the data acquisition module obtains the historical quality data of visual, user interaction, market and design preferences, synthesis and enhancement, and function verification of similar plush toys, preprocesses and processes it for subsequent model construction, and transmits it to the model construction module;

[0061] Please refer to Figure 9 As shown, the step S10 includes the following steps:

[0062] S11. The information collection unit obtains various normal and abnormal quality data, images, and videos of similar plush toys by matching with the industry database according to the customer's needs for subsequent data cleaning and transfers them to the cleaning and repair unit.

[0063] Further explanation: According to the customer's order requirements and preferences, input them into the industry database for matching to obtain various quality data and corresponding images or videos of similar plush toys. The quality image and video of the plush toy include: 1) Basic design materials: physical multi-angle photos covering the front, side, top-down, etc. of the plush toy to train the model to understand the three-dimensional shape and structure; as well as hand-drawn line drawings and 3D model data such as the designer's original manuscript and 3D modeling files to assist AI in generating refined line drawings and adapting to production requirements; 2) Material and texture data: Microscopic texture images of fluff with fiber details of different materials such as wool felt and woven yarn are collected through macro photography to enhance the AI's simulation ability of plush texture; and a video of the dynamic change of light and shadow recording the shadow transition and reflection effect of the toy under environments such as natural light and light to optimize the rendering realism; 3) Dynamic interaction scene data: User interaction videos of children or adults touching and squeezing the toy are taken to train the AI to recognize touch behaviors and generate response logic; and the motion capture trajectory of the displacement and deformation data of the toy is extracted through sensors or video analysis to support dynamic behavior modeling; 4) Virtual generated samples: Use tools to batch generate multi-style AI-generated images with styles such as cartoonization and minimalist style to expand design diversity; and combine text descriptions to generate multi-modal synthesis scenes such as a woolen knitted forest and a plush ocean to expand the creative boundary; 5) Annotation and structured data: Semantic segmentation annotation images with pixel-level annotation of details such as the fluff area and suture of the toy image to improve the model generation accuracy; and instance segmentation multi-toy scenes annotating the boundaries of independent entities in complex scenes to support combined design generation.

[0064] S12. The cleaning and repair unit improves the reliability of commercial design through noise elimination, blur repair, local defect image repair, 3D model repair, highlight area processing, and texture consistency optimization, and transfers it to the segmentation and extraction unit.

[0065] Furthermore, after data collection, code is written according to program commands. First, noise elimination and blurring repair are performed. Gaussian filtering is used to eliminate granular noise such as noise introduced by scanning or shooting in the image to retain the soft transition of the plush texture. Median filtering is performed on salt-and-pepper noise such as thread ends or dust near the seams of plush toys to smooth the edges while avoiding blurring details. And models such as ESRGAN are used to magnify low-resolution hand-drawn sketches for super-resolution reconstruction to restore plush details such as fluff bifurcations and fabric fibers, thereby repairing blurred physical photos and enhancing the clarity of hand-drawn line drawings for subsequent AI generation of 3D models. Secondly, local defect repair is carried out. Generative adversarial networks (GANs) or image repair algorithms are used to fill in missing areas such as torn toy ears and background occlusions in the image, and the continuity such as the fluff direction and density matching of the damaged area of the plush material is restored through context encoding. Then, 3D model repair is carried out. Hole filling and surface smoothing are performed on 3D scan data such as the scanned point cloud of the toy head to ensure printability such as uniform wall thickness of the model. Then, high-light area processing is carried out. The overexposed areas of reflective components such as plastic eyes and metal accessories are eliminated through a multi-angle light compensation algorithm. And the Poisson equation is used to adjust the shadow distribution to avoid the fluff texture such as deep shadows at the intersections of the toy's limbs being disturbed by light. Finally, texture consistency optimization is carried out. Local color correction is performed on the repaired image to ensure uniform color of the plush material. In this way, through the AI model, the material characteristics such as the fluff density and suture direction of the plush toy and the structural characteristics such as the limb ratio and accessory position can be learned more accurately, reducing problems such as "material distortion" or "structural deformity" in the generated design and improving the reliability of commercial design.

[0066] S13. The segmentation and extraction unit performs semantic segmentation and separation from the main body, reflection and highlight suppression, binarization and edge detection, contour repair and filling, and cross-modal semantic alignment to provide high-quality data for 3D modeling and transfer it to the standard enhancement unit;

[0067] Further explanation: After cleaning and repairing, the code is written according to the program command. First, semantic segmentation and subject separation are performed, and U-Net and Mask are used. Deep learning models such as R-CNN perform semantic segmentation on the input image to separate the main body of the plush toy from the complex background, and use a multi-channel convolutional network to identify the detailed features of the toy such as the stitching lines and accessories to generate a high-precision mask. In particular, photos of real objects with patterns or gradient backgrounds can generate standardized images with a pure color background. Secondly, reflection and highlight suppression are performed. For reflective parts such as plastic eyes and metal zippers, HDR synthesis technology is used to balance the brightness of the highlight area for lighting compensation to avoid overexposure that interferes with the main features, and the Poisson equation is used to adjust the shadow distribution to eliminate the blurring of the fluff texture caused by uneven lighting. Then binarization and edge detection are performed, and the segmented main image is converted into a grayscale image. A binary image is generated through an adaptive threshold algorithm such as the Otsu algorithm to highlight the stitching lines and contour features, and the Canny edge detection algorithm is combined to optimize the continuity of the outer contour of the plush toy to reduce breakage or burrs. Then contour repair and filling are performed, and holes such as the hollow parts of the toy ears in the segmentation mask are repaired by morphological closing operations to ensure the integrity of the main body, and Flood The Fill algorithm automatically fills in the missing parts such as the toy legs that are blocked by fur due to occlusion; finally, cross-modal semantic alignment is performed, and the CLIP model is used to verify the semantic matching degree between the segmented long-eared rabbit doll and the text description, to avoid image-text conflicts, and to perform secondary corrections or manual annotation intervention on mismatched areas such as accessories that are incorrectly segmented. This allows the AI model to accurately extract structural features such as the proportion of the limbs and the position of accessories, as well as material features such as the density of fur and the direction of sutures of plush toys, providing high-quality input data for subsequent 3D modeling and stylized generation.

[0068] S14, the standard enhancement unit enhances the modeling capability of material diversity and structural complexity through size normalization, color space conversion, histogram equalization, binarization processing, geometric transformation enhancement, material and lighting perturbation, and passes it to the conversion coding unit;

[0069] Further explanation, after segmentation and extraction, code is written according to program commands. First, size normalization is performed to uniformly adjust the input image to a fixed resolution, eliminating size inconsistencies caused by shooting angles or device differences to adapt to the model input requirements, and the bicubic interpolation algorithm is used to maintain the continuity of the fluff texture to avoid jaggedness or blurring after scaling. Second, color space conversion is carried out. The RGB image is converted to a grayscale image to reduce computational complexity, while focusing on textures such as the fluff density of the material and the direction of the suture, and for specific tasks such as color style generation, normalization processing is performed to preserve the three RGB channels but scale the pixel values to the range of 0-1. Third, histogram equalization is carried out. The image contrast is enhanced through global or local histogram equalization (CLAHE) to highlight details such as fluff bifurcations and sutures, and gamma correction is performed on low-light images to restore texture information such as the internal filling structure of the toy in the shadow area. Then, binarization processing is carried out. An adaptive threshold segmentation such as the Otsu algorithm is used to generate a binary image to strengthen the recognition of contours such as the edge between the plush main body and accessories and sutures. Then, geometric transformation enhancement is carried out. The image is randomly rotated by ±30°, horizontally / vertically flipped to simulate the shapes of plush toys from different perspectives, improving the robustness of the model, and affine transformations such as scaling and translation are added to cover the generation requirements of mini dolls and large pillows of different sizes of toys. Finally, material and lighting perturbations are carried out. Gaussian noise or random blur is injected to simulate shooting interferences such as fluff reflection or camera shake in the real scene, and the HSV color space such as saturation ±20% and brightness ±15% is adjusted to enhance the adaptability of the model to color changes such as color differences of different batches of fabrics. This enables the AI model to efficiently process diverse plush toy input data, enhancing the modeling ability for material diversity such as short fluff, long fluff, and sequin decoration, and structural complexity such as multi-limb splicing and accessory embedding, laying a foundation for generating design solutions that meet commercial standards.

[0070] S15. The conversion and coding unit provides reliable input for intelligent generation and optimization through numerical data standardization, feature coding, data specification and feature extraction, continuous feature binning, and synthetic feature generation, and transmits it to the physical verification unit;

[0071] For further explanation, after the standard is enhanced, code is written according to program commands. First, numerical data is standardized. The dimensional parameters such as the height and arm length of plush toys are standardized by Z-score transformation into a distribution with a mean of 0 and a standard deviation of 1 to eliminate the influence of dimensional differences on the model. Continuous features such as fluff density and filling weight are scaled by min-max normalization and normalized to the 0-1 interval to preserve the original data distribution relationship, especially to unify the numerical ranges of parameters such as suture line spacing and accessory size, and improve the model convergence speed. Secondly, feature encoding is performed. The material types such as short fluff, long fluff, and sequins of plush toys are one-hot encoded and converted into binary vectors to avoid misleading weights introduced by numerical magnitudes. For high-dimensional classification features such as the toy style labels "cartoon" and "realistic", embedding encoding is implemented for low-dimensional dense vector representation to capture semantic correlations. Especially for the "bear doll with a bow", the multi-label design requirement is processed through encoding to achieve multi-attribute joint modeling. Then, data normalization and feature extraction are carried out. The RGB image pixel values of plush toys are normalized to the [-1,1] or [0,1] interval to adapt to the input specifications of deep learning frameworks. The voxelization of 3D model point cloud data is performed to convert the surface structure into a regular grid representation, and dimensionality reduction is carried out through principal component analysis (PCA) to extract key geometric parameters such as the symmetry and aspect ratio of the toy body contour. Then, continuous feature binning is carried out. The weight or price range of plush toys is binned by equal width / equal frequency to generate discretized features such as lightweight and medium weight, which are suitable for decision tree models. The suture line length is dynamically binned through clustering algorithms such as K-means to improve feature interpretability. Finally, synthetic feature generation is carried out. The material type and color features are encoded by Cartesian product to generate combined attributes such as red short fluff and blue long fluff, and non-linear features such as the square term of size are generated through polynomial transformation to enhance the model's fitting ability for complex design rules. This makes the feature variance ratio > 85%, enabling the AI model to efficiently process heterogeneous data such as numerical parameters, images, and classification labels in plush design, and establish the correlation between features such as comfort score and production cost and design goals, providing reliable input for intelligent generation and optimization.

[0072] S16. The physical verification unit can avoid physical contradiction design and process conflicts through material parameter verification, mechanical simulation testing, ergonomic adaptation, and process constraint mapping, so as to ensure that the generated scheme is both innovative and feasible.

[0073] For further explanation, after converting the encoding, code is written according to the program commands. First, material parameter verification is carried out. Through finite element simulation, it is verified whether parameters such as fluff density and filler elastic modulus meet the actual production standards, and it is detected whether the suture tension parameter is within the safety threshold to avoid the risk of unsewing during toy use. Secondly, mechanical simulation tests are carried out. Load simulations such as gravity and extrusion are carried out on the 3D model to verify the anti-deformation ability of the installation points of accessories such as the joints of the toy's limbs and buttons, and the topology optimization algorithm is used to detect whether the ear support skeleton meets the support structure redundancy such as the minimum volume constraint. Then, ergonomic adaptation is carried out. It is checked whether the dimensions such as the ratio of the toy's head to the body meet the child safety standards such as the EU EN71-1 regulations on sharp edges of toys, and it is verified through cluster analysis whether the design parameters (such as the limb length) are within the reasonable distribution range of historical successful cases. Then, process constraint mapping is carried out. It is verified whether the complexity of the suture is adapted to the precision requirements such as the sewing machine stitch pitch error < 0.5mm, and it is detected whether the dimensions and installation positions of accessories such as plastic eyes meet the requirements of the injection mold opening. In this way, the safety compliance rate is < 1%, and the simulation error rate is < 5%. The AI model can avoid physical contradiction designs such as unstable standing caused by overweight fillers and process conflicts such as complex curves exceeding the sewing machine precision, ensuring that the generated scheme is both innovative and feasible.

[0074] S20. The model construction module extracts key features, designs the model, integrates and optimizes, initializes parameters, dynamically optimizes, fuses and deploys the model based on the verified data to build a model for subsequent model training and verification, and transfers it to the model training module;

[0075] Please refer to Figure 10 As shown, the step S20 includes the following steps:

[0076] S21. The feature extraction unit filters out invalid data, extracts physical attributes, encodes user preference features, screens key features, compresses dimensions, and divides data into layers to improve the model training speed and transfers it to the model design unit;

[0077] For further explanation, the code written according to the program commands after physical verification first filters out invalid data, eliminates sample 3D model files with damaged or blurred suture breaks in the design drawings, repairs missing parameters such as null values in the filler density field, and unifies the elastic modulus of the material parameters to MPa and the tear strength to N / mm 2Measurement units such as these are used to avoid multi-source data conflicts. Secondly, physical property extraction is carried out. Mechanical parameters such as suture tension and filler resilience coefficient are extracted through finite element simulation tools and used as input features for structural stability prediction. Geometric features such as fluff density distribution and fitting installation point coordinates are extracted from the design sketches using image segmentation algorithms. Thirdly, user preference feature encoding is performed. User comment tags such as cute style and retro style are converted into semantic vectors through Word2Vec and mapped to the design style generation module. Popular color schemes such as the proportion of pink series ≥ 60% are extracted using RGB histogram analysis, and the color combination is optimized by associating with market sales data. Then, key feature screening is carried out. Redundant features are removed based on the Pearson correlation coefficient, such as the strong correlation between toy weight and size, and only size is retained as the main feature. The importance of features is evaluated using the random forest algorithm, and features with a high impact on target variables such as user collection rate are preferentially retained, such as head roundness score and suture complexity. Then, dimensionality reduction is carried out. Principal component analysis is used for high-dimensional material parameters such as the elastic modulus of 50 fillers, and the dimension is reduced to 5 - 8 dimensions to improve the training efficiency. Finally, data stratified division is carried out. Stratified sampling is carried out according to animal shapes such as felines and bears to ensure the balanced shape distribution of the training set and the test set. The time window division method is used, such as 2022 - 2024 as the training set and Q1 of 2025 as the test set, to verify the generalization ability of the model to the popular trend. In this way, feature engineering transforms raw data such as design drawings and user feedback into structured features that can be efficiently utilized by AI models, laying a foundation for generating design solutions that are both innovative and feasible for production.

[0078] S22. The model design unit provides directions for subsequent model training through the application of generative adversarial networks, multi-modal input fusion, generator network design, discriminator network optimization, and transfer learning application, and transmits them to the integrated optimization unit;

[0079] Further explanation: After feature extraction, code is written according to program commands. First, an adversarial network application is generated. The generator of the GAN framework outputs a design sketch based on tags such as cute style and retro style input by the user. The discriminator evaluates the consistency between the generated result and the real design, driving style imitation and innovation. For complex structures such as movable joints, a conditional generative adversarial network is introduced, and the generation direction is constrained by mechanical parameters such as tensile strength ≥ 40N. Secondly, multi-modal input fusion is carried out, integrating multi-modal input layers of text such as user preferences, images such as historical design drawings, and numerical values such as material properties. The different modal features are dynamically weighted through the attention mechanism to enhance the diversity and practicality of the design scheme. Then, the generator network design is carried out. A generator with a U-Net architecture is constructed. The skip connection is used to retain details such as fluff bifurcation and suture position in the sketch, supporting progressive generation from low resolution to high resolution of ≥ 512 * 512 pixels. And a style control module is added to adjust the generated feature distribution through label vectors such as the user's round head and short limbs to achieve directional design output. Then, the discriminator network is optimized. A multi-scale discriminator is used to evaluate the overall style consistency (global discriminator) and local detail accuracy (local discriminator) of the design drawing respectively, such as verifying the coordination of facial expressions and body proportions. Finally, transfer learning is applied. The convolutional layer parameters of the pre-trained model are reused to initialize the generator, accelerating the feature extraction ability of small sample data, retaining the general texture generation characteristics, and loading the NLP pre-trained model for the text encoding layer to map user descriptions such as soft touch and safe and non-toxic into semantic vectors, reducing the need for labeled data.

[0080] S23. The integrated optimization unit speeds up the subsequent model training through physical constraint embedding, regularization optimization, interpretability enhancement, model compression, and quantization-aware training, and transmits it to the parameter initial unit;

[0081] For further illustration, after the model design, code is written according to program commands. First, physical constraint embedding is performed. A structural stability index is introduced into the loss function. For example, the tensile strength at the joints of the limbs is calculated through the simulation module to ensure that the generated design complies with the EN71-1 safety standard. Second, regular optimization is carried out. L2 regularization (λ = 0.01) is applied to the material parameter prediction task to limit the reasonable fluctuation range of parameters such as filling density and elastic modulus. Then, interpretability enhancement is implemented. The SHAP tool is integrated to visualize the influence weights of preference labels such as the user's color scheme and accessory layout on the generation results, supporting the backtracking of the design logic. Next, model compression is performed. Knowledge distillation is used to compress the complex generator into a lightweight version, reducing the inference time to ≤5 minutes per scheme while maintaining the generation quality (SSIM≥0.85). Finally, quantization-aware training is carried out. Quantization-aware training is implemented for the 3D model generation module to convert floating-point parameters into 8-bit integers to adapt to the deployment requirements of embedded design terminals. This enables the AI model to generate plush toy design schemes with high precision (<25), high compliance (safety test pass rate≥98%), and high customization (user demand matching degree≥80%).

[0082] S24. The parameter initial unit provides efficiency for subsequent model training through pre-training transfer, word vector pre-training, random initialization, distribution initialization, material parameter constraint, and style feature enhancement, and transfers it to the dynamic optimization unit;

[0083] For further illustration, after the integrated optimization, code is written according to program commands. First, pre-training transfer is performed. The pre-trained convolutional layer parameters of ResNet-50 or VGG-16 are loaded for the image feature extraction module of the generator, retaining the general texture generation ability such as fluff density and suture direction, and reducing the overfitting risk under small sample data. Second, word vector pre-training is carried out. For the text encoding module of user preference labels such as "round shape, environmentally friendly material", the pre-trained word vector parameters of BERT are reused to quickly capture semantic relevance. Third, random initialization is performed. The output layer of the GAN generator uses He initialization (standard deviation = 0.02) to accelerate the convergence of sketch details such as the position of the accessory installation point. Then, distribution initialization is carried out. The fully connected layer of the discriminator is initialized with the Xavier uniform distribution to balance the gradient propagation stability of different style design drawings such as "cute style" and "realistic style". Then, material parameter constraint is carried out. The weight of the filling density prediction layer is initialized to a mean of 0.5, with the unit of g / cm 3 , and the range matches the common PP cotton of 0.03 - 0.08 g / cm 3 and 0.12 - 0.15 g / cm 3Memory foam and other filling materials are used to narrow the parameter search space. For mechanical simulation modules such as tensile strength prediction, the bias term is initialized to the industry safety threshold to guide the model to quickly approach the compliant design. Finally, style feature enhancement is performed. In the style control module of the generator, the embedding vector of the user preference label is initialized as a one-hot encoding. For example, "retro style" corresponds to [1, 0, 0], enhancing the controllability of the design direction. This increases the convergence speed of model training by 30 - 50%.

[0084] S25. The dynamic optimization unit provides safety performance for subsequent model training through residual connection parameter configuration, regularization fusion, pre-training fine-tuning, module differentiation, initialization sensitivity testing, and production environment adaptation, and transmits it to the network level unit;

[0085] Further explanation: After the parameters are initialized, code is written according to the program commands. First, residual connection parameter configuration is performed. For the skip connection layer of the U-Net generator, the scaling factor is initialized to 0.1 to suppress the gradient explosion caused by the difference in feature map sizes in the initial stage of training. And in the multi-scale module of the discriminator, orthogonal initialization is applied to the convolution kernels of the low-resolution branch to improve the discrimination accuracy of local details such as the eye suture point. Second, regularization fusion is performed. A Dropout layer is added to the material prediction branch of the generator, with an initial dropout rate = 0.3 to prevent overfitting of filler parameters such as elastic modulus. And in the attention module of the discriminator, the spectral normalization parameters are initialized to constrain the Lipschitz constant of the weight matrix and improve the stability of adversarial training. Third, pre-training fine-tuning is performed. The pre-trained parameters of the bottom convolutional layer of the generator are fixed, and the top deconvolutional layer is randomly initialized, taking into account both the reuse of general features and the innovation of design details. Then, module differentiation is performed. The text encoding module is initialized with BERT to maintain semantic understanding ability, and the image generation module is initialized with GAN-focused to enhance the texture generation effect. Then, initialization sensitivity testing is performed to compare the impact of different initialization methods on the generation quality, such as the difference in FID scores between He initialization and Kaiming initialization, ΔFID < 5, and monitor the loss curve in the initial stage of training. If the convergence speed is delayed by more than 20%, adjust the learning rate or the initialization distribution range. Finally, production environment adaptation is performed. For embedded design software such as AutoCAD plugins, the quantized model parameters are calibrated and initialized to reduce the accuracy loss caused by 8-bit integer conversion. This makes the compliance safety parameter compliance rate of production ≥ 95%.

[0086] S26. The network level unit obtains the convolutional neural network model architecture of the input layer, convolutional layer, activation layer, pooling layer, and fully connected layer according to feature and parameter optimization for subsequent model training.

[0087] For further illustration, during the dynamic optimization of various parameters, the number of network layers, the size of convolutional kernels, the stride, the padding method are adjusted according to the network hierarchical structure, and the loss function and optimization algorithm are selected to ensure the performance and accuracy of the model. Then, by adding zero-padding layers and normalization layers before and after the convolutional layer respectively, the size of the input data is adjusted and normalized to improve the training speed and stability of the model. Then, according to the ratios of the training set, validation set, and test set determined by the preselected segmentation strategy, and ensuring that the data distributions among the subsets are as consistent as possible to avoid introducing biases. In some application scenarios, due to the time series characteristics of the data, the dataset is divided into a training set for training the model, a validation set for adjusting the model parameters and selecting the best model, and a test set for evaluating the final performance of the model according to the time series by the written code to ensure the temporal consistency of the training and test sets. Finally, the dataset is used to establish the training set and test set required for training the prediction model for identifying quality anomalies in radiator production control in different environments in a 9:1 relationship through deep learning of convolutional neural networks, thus establishing a deep convolutional neural network CNN model. Since the convolutional kernels in the convolutional layer contain weight coefficients while the pooling layer does not, the pooling layer is not an independent layer. The convolutional layer is used to extract local features of the image, the pooling layer is used to reduce the dimension of the feature map, and the fully connected layer is used to integrate features and perform classification.

[0088] S30. The model training module trains the model based on the collected data, adjusts the model parameters to optimize the model performance according to the predetermined plush toy design scheme, ensures the accuracy of the subsequent proofing quality, and transfers it to the model evaluation module;

[0089] Please refer to Figure 11 As shown in the figure, the step S30 includes the following steps:

[0090] S31. The convolution input unit obtains the convolution input value according to the convolution input calculation formula "z(t) = ∫x(m)y(t - m)dm, where z(t) is the output function, x(t) is the input function, y(t) is the convolution kernel function, m is the integration variable, and dm is the integration element", activates the function input, and transfers it to the pooling conversion unit;

[0091] For further explanation, after normalizing the image, the data is first filled according to the data padding size calculation formulas "Ph = [(Ho - 1) * Sh + Kh - Hi] / 2, Pw = [(Wo - 1) * Sw + Kw - Wi] / 2, where Ph and Pw are the padding sizes in the height / width directions respectively, Hi and Wi are the height / width of the input feature map respectively, Kh and Kw are the height / width of the convolution kernel respectively, Sh and Sw are the values of the stride in the height / width directions respectively, and Ho and Wo are the height / width of the output feature map" to obtain the padding data and input it. Then, the convolution input value is obtained through the above convolution input calculation formula; the integration element dm is the "width" or "increment" when integrating each infinitesimal interval of the variable m. The cumulative effect z(t) on the domain is obtained by accumulating the function values on these infinitesimal intervals. dm is the integration element for integrating the variable m; and the convolution value is obtained through the convolution calculation and enters the convolution layer. The activation function value is obtained through the ReLU function calculation formula "f(x) = max(0, x + Y)" and enters the activation function layer. In the formula, f(x) is the activated function, x is the eigenvalue output by the convolution layer or the fully connected layer, and Y is a random variable; the output function z(t) is the convolution result of the input and the convolution kernel, that is, the feature map after the signal is filtered, which is realized by discrete summation or FFT acceleration; the input function x(t) is the input signal, which is real-time data such as sound and tactile signals collected by sensors or preprocessed feature data; the convolution kernel function y(t) is a weight parameter, which is iteratively optimized through the backpropagation algorithm and training data, and finally converges to a form that can extract effective features; the integration variable m is the sliding position of the convolution kernel on the input signal; the integration element dm is the integration element for the variable m, which is an infinitesimal quantity in the continuous domain and corresponds to the sampling interval after discretization.

[0092] S32. The pooling conversion unit obtains the maximum pixel value after pooling according to the average pooling calculation formula "Z(I, j) = mean(X[i * Ps(i + 1) * Ps, j * Ps(j + 1) * Ps])), where Z(I, j) is a pixel value in the output feature map after pooling, (i, j) is the coordinate in the output feature map, mean() is to obtain the average value of all pixel values in the input window, X is the input feature map, and Ps is a hyperparameter" to retain important feature information and transfer it to the fully connected layer unit;

[0093] Further explanation: After obtaining the average pooling pixel value through the average pooling calculation formula, the maximum pixel value after max pooling is obtained according to the max pooling calculation formula "Z(I,j) = max(X[i*Ps(i + 1)*Ps,j*Ps(j + 1)*Ps])", where max in the formula is to find the maximum value among all pixel values in the input window; X is the output feature map of the previous layer, generated by the convolutional layer or the activation function layer; the hyperparameter Ps is the window size of the pooling operation, preset manually according to the model requirements, generally 2×2 or 3×3, used to balance feature retention and computational efficiency; (i,j) is the coordinate in the output feature map, determined by the pooling window sliding stride and the input size; mean() is to sum all pixel values in the input window and then divide by the total number of pixels P s 2 , and backpropagation is performed, that is, the average pooling layer gradient calculation needs to evenly distribute the upper layer error to each pixel in the original window; Z(i,j) is the dimensionality-reduced feature, used to reduce the number of parameters, prevent overfitting, and at the same time retain the main features, and is the input to the subsequent fully connected layer or classification layer, supporting tasks such as instruction recognition or sentiment analysis.

[0094] S33. The fully connected layer unit obtains the output data of the connection layer according to the fully connected layer calculation formula "y = f(∑(Wn*Xn)+b), where y is the output result, f is the activation function, Wn is the weight of the nth input feature, Xn is the n input features, n is the dimension of the input features, and b is the bias", for subsequent output calculation, and is passed to the normalized output unit;

[0095] Further explanation: The calculation formula of the fully connected layer is deduced from "y = f(W*X + b)" to "y = f(W1*X1 + W2*X2 + … + Wn*Xn + b)"; the output result y is the final output after the fully connected layer is processed by linear transformation and non-linear activation function, which can be the class probability of the classification task, the predicted value of the regression task, etc., specifically depending on the subsequent processing of the model and the task type; Xn is the feature vector output by the previous layer (such as the convolutional layer or the pooling layer), generally a 512-dimensional speech spectrum feature, used to transfer abstract features; Wn is randomly initialized and optimized through training data, generally with an initial value of He normal distribution, used to control the feature weights; b is initially 0 or a small random number, and is optimized synchronously with the weights, generally with an initial value of 0.01, used to adjust the output baseline; f is preset manually and selected according to the task type, generally ReLU (hidden layer), used to introduce non-linearity; y is the output after weighted summation through the activation function, generally a Softmax probability distribution, used for classification / regression results.

[0096] S34. The normalization output unit obtains the output of the BN layer according to the Batch Norm layer calculation formula "h = φ(Bn(Wx + b)), where h is the output of the fully connected layer, φ is the activation function, Bn is the operator for batch normalization, W is the weight parameter, X is the input of the fully connected layer, and b is the bias parameter" for the output of the model data and passes it to the output conversion unit;

[0097] Further explanation: The Batch Norm layer calculation formula "h = φ(Bn(Wx + b))" is derived from "Bn(x) = γ⊙[(X - μb) / σb] + β". In the formula, μb is the mean, σb is the standard deviation, γ is the stretching parameter, β is the offset parameter, and (X - μb) / σb is the standardized normal distribution; The Batch Norm layer (abbreviated as "BN layer") is placed between the affine transformation and the activation function in the fully connected layer to standardize the output value of each layer in the network, making it more obedient to the normal distribution, which can accelerate the training speed of the neural network and help prevent gradient disappearance and gradient explosion; The h is obtained from the output after BN and the activation function, generally a Softmax probability distribution, used for classification / regression results; The φ is obtained from a preset type (such as ReLU / Sigmoid), generally ReLU (hidden layer), used to introduce non-linearity; The W is initialized with He / Xavier and optimized through training data, generally the initial value is a He normal distribution, used for linear combination of features; The x is obtained from the feature vector output by the previous layer (such as the convolutional layer), generally a 256-dimensional image feature vector, used to transfer high-order abstract features; The b is initialized to 0 and updated during training, generally the initial value is 0, used to adjust the output baseline; The γ and β are initially 1 and 0 respectively and optimized through training. Generally, after training, γ ≈ 0.8, used to restore the data distribution expression ability; The μ and σ2 are calculated as batch statistics during training and obtained using moving averages during inference, generally μglobal, used to standardize data to accelerate convergence.

[0098] S35. The output conversion unit obtains the size of the output value after convolution according to the output layer conversion calculation formula "N = (P - F + 2C) / S + 1, where N is the size of the output after convolution, P is the size of the input before convolution, F is the size of the convolution kernel, C is the number of layers of 0 added around the image, and S is the step size" for subsequent model analysis and training and passes it to the backpropagation unit;

[0099] Further explanation: The output layer transformation calculation formula is used to calculate the size of the output feature map after convolution operation. In a convolutional neural network, features are extracted by performing convolution operations on the input feature map. The N represents the size of the feature map obtained after convolution operation, which represents the side length of the output feature map in two-dimensional convolution. The P represents the size of the width or height of the input image or feature map. By the above calculations, the padding size Ph in the height direction and the padding size Pw in the width direction are obtained respectively, which represent the side length of the input feature map in two-dimensional convolution. The F is the matrix obtained by sliding the convolutional kernel on the input feature map and performing convolution operation, that is, the side length of the convolutional kernel. The number of layers C of adding 0 around the image is used to control the size of the output feature map and is also obtained by the above network construction. The C is the padding data, that is, in order to keep the size of the output feature map close to that of the input feature map before convolution operation, or to meet the requirements of specific network design, zero-value pixels are added around the input feature map, that is, the number of layers of added zero-value pixels. The S is the distance at which the convolutional kernel slides on the input image or feature map and is obtained by the above calculations. The larger the S, the smaller the size of the output feature map.

[0100] S36. The backpropagation unit performs backpropagation of the convolution output value according to the backpropagation calculation formula "dz[l] = da[l] * g[l]′(z[l])", where dz[l] is the input gradient, da[1] is the output gradient, g[l]′(z[l]) is the derivative of the activation function, and z[l] is the input of the activation function, so as to optimize the performance of the network model.

[0101] Further explanation: dz[l] is the gradient of the input z[l] of the activation function of the l-th layer. During the backpropagation process, the dz[l] of each layer needs to be calculated to update the weights, biases, and transmit the gradients. da[l] is the gradient of the output a[l] of the activation function of the l-th layer, which is obtained through backpropagation of the gradients of the output layer and intermediate layers calculated by the loss function. Generally, it is the cross-entropy loss output gradient y′ - y, which is used to guide the direction of error backpropagation. g[l]′(z[l]) is the derivative of the activation function g[l] of the l-th layer at z[l], which is calculated according to the type of activation function and obtained by caching z[l] during forward propagation. Generally, it is 0 or 1 of the ReLU derivative, which is used to calculate the input gradient of the current layer. z[l] is the input of the activation function of the l-th layer, which is the linear output (W[l]a[l - 1] + b[l]) of forward propagation and needs to be explicitly cached. Generally, it is a quantized INT8 tensor, which is used as the basis for calculating the derivative of the activation function. Through the above backpropagation calculation formula, the convolution output values are obtained as dw[l] = dz[l] * a[l - 1], db[l] = dz[l], and da[l - 1] = W[l]T * dz[l]. Through the backpropagation of the convolutional layer, the error term da[l] is passed to the previous layer through a convolutional operation, the convolutional kernel is flipped and applied to the error map, and the local gradient is calculated through the derivative of the activation function, which is implemented through matrix operations. The input and convolutional kernel are deformed into matrices for matrix multiplication, that is, the numbers in the convolutional window are pulled into a row to form a column vector for matrix multiplication.

[0102] S40. The model evaluation module verifies and evaluates the model based on the trained model, and adjusts and optimizes the model according to the verification results to improve the performance and accuracy of the model, and transmits it to the solution determination module;

[0103] Please refer to Figure 12 As shown, the step S40 includes the following steps:

[0104] S41. The classification and marking unit classifies and marks the corresponding data sample categories according to the abnormal categories of the plush toy design quality for subsequent machine learning of the model, and transmits it to the learning and training unit;

[0105] Further explanation: Classification is carried out according to the abnormal phenomena of the design quality of plush toys: "Chemical safety abnormality" is category 1, "Mechanical and physical defects" is category 2, "Material and process defects" is category 3, and "Labeling and compliance defects" is category 4; When category 1 is the positive sample, the remaining categories are negative samples, that is, the labeled data is (1, 0, 0, 0). When category 2 is the positive sample, the remaining categories are negative samples, that is, the labeled data is (0, 1, 0, 0). When category 3 is the positive sample, the remaining categories are negative samples, that is, the labeled data is (0, 0, 1, 0). When category 4 is the positive sample, the remaining categories are negative samples, that is, the labeled data is (0, 0, 0, 1); The above four abnormal categories are relatively typical abnormal classifications for the design quality control of plush toys, including but not limited to the above four categories. One or more of these categories can also be reclassified into more categories, and more refined and in-depth classification can be carried out according to the actual design quality effect. The above classification is not fixed. After each category change, the model is retrained to meet the needs of subsequent production practice; The abnormal design quality of plush toys includes: 1) Chemical safety abnormality: Caused by excessive formaldehyde, or due to harmful substances such as phthalates and lead in the filling or printing ink; 2) Mechanical and physical defects: Caused by insufficient pulling force of plastic eyes / buttons, or due to the rope length exceeding the standard > 220 mm or the packaging bag being too thin < 0.038 mm, or due to insufficient seam strength resulting in the leakage of the filling; 3) Material and process defects: Caused by uneven distribution of the filling resulting in the collapse or deformation of the head / limbs, or due to plush shedding, the broken hair rate > 5% or poor color fastness, and color fading after friction; 4) Labeling and compliance defects: Caused by the lack of applicable age warning or the execution standard number, or due to false labeling of "Class A baby products".

[0106] S42. The learning and training unit determines the training model according to the gradient descent data, and inputs the quality data of the plush toys in the training set into the model for simulation training to ensure the accuracy of model recognition, and transmits it to the cost verification unit;

[0107] Further explanation: After fully communicating with the customer to obtain the specific needs of the customer, abnormal prevention measures are obtained through model training and matching for layout to avoid various production quality abnormalities. The prevention measures for the design quality abnormalities of plush toys include: 1) Intelligent design verification mechanism: First, multi-model collaborative verification is carried out, using double verification of "3D modeling engine + physical simulation system" to achieve real-time interception of 12 types of design defects such as suture overlap error > 0.3mm and disproportionate facial features, and enabling the virtual assembly system to simulate the deformation effects under different filling densities to identify weak points in the support structure in advance; Second, historical defect atlas matching is carried out, calling the design defect database in the past 5 years (including more than 3000 cases), and automatically identifying high-risk design schemes through semantic matching algorithms to trigger correction reminders; 2) Dynamic parameter closed-loop control: First, core parameter threshold management is carried out, setting dynamic adjustment intervals for hair texture parameters, creativity values, and three-view alignment degrees, automatically calibrating and generating optimization logs for each iteration. When the user-defined size > 100cm, the support grid density is forced to increase and a reinforcement plan is generated; Second, a foolproof design rule library is built, with 30 anti-error rules built in, including that special-shaped parts need to be embedded with anti-dropping buckles and redundant sewing lines are reserved at the seams; 3) Design process collaborative optimization: First, a manufacturability prediction system is used, and the process simulation engine automatically detects the complexity of the sewing path, generates a simplified alternative plan for high-difficulty designs, and enables the material adaptation system to recommend the optimal combination in real time to avoid physical property conflicts; Second, reverse engineering calibration is carried out. The first version of the sample generates point cloud data through 3D scanning, and reversely corrects the initial line drawing parameters, with an error compensation accuracy of ±0.1mm; 4) Quality data iterative upgrade: First, the defect feature library is dynamically updated, with more than 200 user feedback data added every week, which is used to retrain the weights of the ControlNet model, and the prediction accuracy of design defects is increased to 99%; Second, blockchain evidence storage and traceability are carried out. Each design plan is bound with a unique ID, associated with design parameters, correction records, and quality inspection reports, and the quality problem traceability time limit < 5 minutes; 5) Strengthening of personnel capabilities: First, AI-assisted design training is carried out, and real-time design specification prompts are pushed through AR glasses. The error operation interception response time < 10 seconds, and monthly AI simulation abnormal drills are carried out, with the emergency handling efficiency increased by 50%. This can increase the first version design passing rate of plush toys by 35%, reduce the number of user modifications / schemes by 75%, reduce the design defect outflow rate by 90%, reduce the design rework cost by 82%, and increase the customer customization satisfaction rate to 98.6%.

[0108] S43. The cost verification unit calculates according to the cost function formula "J(θ) = -1 / m∑m∑K[y k (i)log(h θ (x(i))k)+(1-y k (i))log(1-(h θ(x(i))k)] + λ / 2m ∑L-1 ∑sl ∑sl+1(θj,i(l))2, where J(θ) is the cost function, m is the total number of training samples, K is the total number of classes, y k (i) is the true label value of the sample, h θ (x(i))k is the predicted probability of the sample, λ is the regularization parameter, L is the total number of layers of the neural network, sl is the number of neurons in the l-th layer, and θj,i(l) is the weight parameter between connected neurons. "Obtain the value of the cost function after model training to ensure the accuracy of model evaluation and pass it to the application evaluation unit."

[0109] Furthermore, the calculation formula of the cost function obtains the minimum cost function of the model to measure the difference between the model prediction result and the true label, which can optimize the parameters of the model and improve the prediction accuracy of the model; the θ is the model parameter, including weights and biases, which are continuously adjusted during training to minimize the cost function; the total number of classes K in a multi-classification problem, each sample will be assigned a class label; the true label value of the sample "y k (i)" is the value of the true label of the i-th sample on the k-th class. If it is a binary classification problem, its value is 0 or 1. If it is a multi-classification problem, its value will be a one-hot encoded vector, that is, the probability that the sample belongs to a certain class; the predicted probability of the sample "h θ (x(i))k" is the probability that the model predicts the i-th sample as the k-th class; the regularization parameter λ controls the strength of the regularization term to prevent model overfitting, which is achieved by penalizing large weight parameters; the weight parameter "θj,i(l)" between connected neurons is the weight parameter between the i-th neuron in the l-th layer and the j-th neuron in the l+1-th layer; the "-1 / m ∑m ∑K[y k (i) log(h θ (x(i))k) + (1 - y k (i)) log(1 - (h θ (x(i))k)]" is the cross-entropy loss function to measure the difference between the model prediction result and the true label; the "λ / 2m

[0110] ∑L-1 ∑sl ∑sl+1(θj,i(l))2" is the regularization term to prevent model overfitting. By minimizing the cost function, the parameters of the model can be optimized and the generalization ability of the model can be improved.

[0111] S44. The application evaluation unit obtains the comprehensive evaluation score F according to the model evaluation score calculation formula "F = 2(Ac * Re) / (Ac + Re), where F is the evaluation score of the trained model, Ac is the precision rate, and Re is the recall rate", and passes it to the processing center;

[0112] Further explanation: Ac and Re are respectively calculated through "Ac = (Tp + Tn) / (Tp + Fp + Fn + Tn), Re = Tp / (Tp + Fn)", where Tp is the number of true positive samples, Fp is the number of actually false positive samples, Fn is the number of false negative samples, and Tn is the number of true negative samples; after the model training is completed, tests are carried out to verify the accuracy and reliability of the model, and problems existing in the model may be found during the test for optimization and adjustment; the evaluation score of the training model, that is, the F value, is the harmonic mean of precision and recall, which can evaluate the performance of the classification model, and its value range is from 0 to 1, where 1 is the best performance and 0 is the worst performance. The higher the F value, the better the prediction effect of the model, and vice versa. Therefore, if the F value is close to 1, it indicates that the model performs well in maintaining the balance between precision and recall; adjust the hyperparameters according to the performance of the model such as the learning rate and regularization coefficient to optimize the model performance.

[0113] S45. The processing center compares the actual evaluation score of the model with the evaluation score standard of the model stored in the memory: if it meets the standard, it is the pre-determined plush toy design plan; if it does not meet the standard, it is transmitted to the alarm and notifies to continue training.

[0114] Further explanation: The test evaluation score standard of the plush toy design quality anomaly identification training model is best when it is greater than 0.8, and is determined by the following factors: 1) Material selection such as fabric type, filling quality, and chemical safety; 2) Structural design such as the firmness of accessories, sewing strength, and safety structure; 3) Production processes such as cutting accuracy, sewing process, and filling process; 4) Safety compliance such as national standards and label specifications; 5) Appearance and details such as design creativity and detail handling; 6) Quality control and inspection such as raw material inspection, finished product inspection, and market feedback. The plush toy design plan includes: 1) Intelligent design core architecture: AI design engine, 3D modeling and simulation; 2) Whole-process quality prevention and control system: design verification system, reverse engineering calibration; 3) Customer customization and experience optimization: interactive design platform, intelligent recommendation algorithm; 4) Data-driven iterative system: first, quality data closed-loop, device linkage control; 5) Production processes, etc. (see the description in step S57 for details). After being trained, the model is automatically matched and applied in production to avoid the occurrence of various quality anomalies; if the evaluation score of this model does not meet the standard, it returns to step S42 to retrain and optimize and adjust the parameters until it meets the standard.

[0115] S50. The plan determination module conducts a proofing test according to the plush toy design plan predicted by the trained model to obtain the best plush toy design plan, so as to ensure the accuracy of the subsequent trial production quality, and transmits it to the trial production verification module;

[0116] Please refer to Figure 13 As shown, the said step S50 includes the following steps:

[0117] S51. The fabric consumption unit obtains the fabric consumption of the plush toy according to the fabric consumption calculation formula for plush toys "So = ∑(Si * Ni) / η, where So is the total fabric area of the plush toy (m 2 ), Si is the single-piece area of the i-th cut piece of the plush toy (m 2 ), Ni is the number of the i-th cut piece of the plush toy, η is the fabric utilization rate of the plush toy, and N is the total number of types of cut pieces of the plush toy", in order to balance cost and efficiency and transmit it to the filling weight unit;

[0118] Further explanation: The fabric area So of the plush toy is the total fabric area required for producing the plush toy, including cutting losses; the single-piece area Si of the i-th cut piece of the plush toy needs to calculate the area by measuring the length and width of the cut piece through CAD software or manually. For complex shapes, it is disassembled into simple geometric figures for summation, or an integral tool is used; the number Ni of the i-th cut piece of the plush toy is the number of uses of the i-th type of cut piece in a single plush toy, which is directly counted according to the structural design such as the number of parts like the limbs and head of the toy; the fabric utilization rate η of the plush toy is the proportion of the fabric effectively utilized after nesting of the plush toy, which is negatively correlated with cutting losses, and the actual utilization rate is calculated by simulating nesting through software such as Optitex. Generally, for simple cut pieces, it is ≥85%, and for special-shaped cut pieces, it is ≥70%; N is the number of types of cut pieces with different shapes / sizes of the plush toy, which is directly counted for the independent cut piece types in the design drawing.

[0119] S52. The filling weight unit obtains the filling weight of the plush toy according to the filling weight calculation formula for plush toys "Go = V * ρ * (1 + k), where Go is the filling weight of the plush toy (g), ρ is the filling density of the plush toy (g / m 3 ), V is the volume of the plush toy (m 3 ), and k is the processing loss rate of the plush toy", in order to control the filling cost and transmit it to the seam strength unit;

[0120] Further explanation: V * ρ is the theoretical filling weight of the plush toy, that is, the weight when there is no loss in the processing of the plush toy; 1 + k is the loss compensation factor of the plush toy, which is used to make up for material waste such as uneven compression and operation losses during the filling process; the volume V of the plush toy is calculated by immersing the finished plush toy in water before design and proofing, and calculating according to Archimedes' principle through the drainage volume. If the toy is approximately a regular geometric body, the volume formula is used (if it is a sphere, then V = 4 / 3πr 3 ), and if it is after design, the volume data is directly exported through 3D modeling design software; the filling density ρ of the plush toy refers to the reference filling density values such as PP cotton and pearl cotton in the material parameter table provided by the supplier, and the typical range is 20 - 50 kg / m 3 , generally, the density of PP cotton is 0.03 g / cm 3, The down cotton is 0.02 g / cm 3 , If the supplier does not provide a specific material parameter table, the weight of the filling per unit volume is measured by the experimental method. If the mass of the filling in 1 m 3 is 30 kg, then ρ = 30 kg / m 3 ; The k is usually 0.1 - 0.2 (i.e., 10 - 20%), which specifically depends on the filling process and equipment accuracy, and is obtained by statistically analyzing the deviation ratio between the actual usage and the theoretical usage in the historical production data in the historical production records.

[0121] S53. The seam strength unit obtains the seam strength of the plush toy according to the plush toy seam strength calculation formula "F = S * (L / A) * ωo, where F is the force borne by the plush toy seam (N), S is the single - needle seam strength of the plush toy (N / st), L is the length of the plush toy seam (cm / st), A is the stitch pitch of the plush toy (cm), and ωo is the production safety factor of the plush toy", to avoid toy damage or safety hazards caused by stitch breakage, and transmits it to the joint torque unit;

[0122] Further explanation, it is obtained by substituting "No = L / A" into "F = S * No * ω" through the plush toy seam strength calculation formula. In the formula, No is the number of stitches of the plush toy seam (st), to avoid toy damage or safety hazards caused by stitch breakage; S * N is the theoretical seam strength, that is, the maximum bearing capacity when all stitches are defect - free; the single - needle seam strength S is the maximum tensile force that a single stitch can withstand before breaking, which is obtained by tensile testing the breaking strength of a single stitch with a tensile testing machine or by querying the reference stitch specification parameter table provided by the supplier. Generally, the polyester thread is 15 N / st, the nylon thread is 20 N / st, and the cotton thread is 10 N / st; L is the actual stitching length of the seam, which is measured segmentally and accumulated for curved seams such as arc parts. The theoretical value in the design drawing has a sewing error of ±5% from the actual value; A is the distance between each stitch. The industry requirement is A ≤ 0.3 cm / st to prevent filling leakage; ωo is to compensate for risks such as stitch aging, sewing process fluctuations, and material non - uniformity. Generally, ω is 1.5 - 3.0, but for children's toys, ω ≥ 2.0, with a safety redundancy of 100%. However, for high - risk scenarios such as hanging toys, it is increased to 3.0. Affected by factors such as stitch type, stitch pattern type, and dynamic load, if the cotton thread has low strength and S is 10 N / st, then the number of stitches or the safety factor is increased. If it is a high - strength thread and S > 30 N / st, then the number of stitches is reduced, but the cost is higher; if the flat seam has stable strength, but for the double - line structure, it is calculated according to the actual effective number of stitches. If the chain stitch is prone to thread shedding, then ω is increased to more than 2.5; for seams at joints that are frequently bent, etc., the fatigue strength is additionally considered, and ω ≥ 2.5.

[0123] S54. The joint torque unit obtains the torque of the movable joints of the plush toy according to the formula for calculating the torque of the movable joints of the plush toy "τ = ρ * V * g * L * ωa", where τ is the torque of the plush toy joint (N·m), ρ is the density of the plush toy filling (g / m 3 ), V is the volume of the plush toy (m 3 ), g is the gravitational acceleration of the plush toy (m / s 2 ), L is the arm length of the plush toy (m), and ωa is the safety factor for using the plush toy, so as to improve the mobility and durability of the plush toy and transmit it to the cost accounting unit;

[0124] Further explanation, the formula for calculating the torque of the movable joints of the plush toy is obtained by substituting "G = ρ * V" into "τ = G * g * L * ωa". In the formula, G is the weight of the plush toy limb (kg), which is obtained by direct weighing or marked on the design drawing, reflecting the rotational effect generated by gravity of the limb. A reverse torque is provided by joint structures such as buttons and Velcro or drive devices such as motors to maintain balance. The total weight of the limb G = the weight of the filling Go + the weight of the fabric + the weight of the accessories. Usually, the proportion of the weight of the filling Go is ≥ 50%; the τ is calculated through the formula for the torque of the movable joints, and the unit of τ is kg·m 2 / s 2 , and 1 kg·m 2 / s 2 = 1 N·m. Therefore, the unit of τ is N·m, which is the torque required for joint rotation. When the joint friction torque ≥ 50 mN·m, the current design does not meet the standard, and it is necessary to increase the friction or shorten the arm length. According to the international toy standard, the torque of the movable joint ≤ 0.5 N·m to avoid the risk of pinching. If τ is too small, the limb will droop and the joint will be loose, making the arm unable to stay lifted. If τ is too large, the joint will have difficulty moving, the material will fatigue and break, and frequent bending will cause the wire to break. Therefore, reduce the weight Gm, select lightweight filling materials, and shorten the arm length L to optimize the limb shape so that the center of gravity is close to the joint; the g is the gravitational acceleration constant on the earth's surface, and the general standard value g = 9.8 m / s 2 ; the L is the vertical distance from the joint rotation center to the center of gravity of the limb, which is obtained by measurement or three-dimensional modeling analysis; the ωa is the additional safety loss due to dynamic forces such as shaking and impact during actual play of the plush toy, and generally ωa = 1.5 - 2.0.

[0125] S55. The cost accounting unit obtains the total cost of plush toys according to the plush toy cost calculation formula "Ct = [(Pf*So) + (Ps*Go) + Ca] * (1 + μ) + Cb + Cm, Ct is the total cost of plush toys (CNY), Pf is the unit price of plush toy fabric (CNY / ㎡), So is the amount of plush toy fabric used (㎡), Ps is the unit price of plush toy filling (CNY / kg), Go is the weight of plush toy filling (kg), Ca is the cost of plush toy auxiliary materials (CNY), μ is the loss rate of plush toys, Cb is the labor cost of plush toys (CNY), Cm is the apportionment cost of plush toy molds (CNY)", so as to quickly estimate the standardized production cost of plush toys and pass it to the proofing verification unit;

[0126] Further explanation: the cost calculation formula of the plush toy is obtained by substituting "Cr=Cf+Cs+Ca" into "Ct=Cr*(1+μ)+Cb+Cm" to obtain "Ct=(Cf+Cs+Ca)*(1+μ)+Cb+Cm", and then substituting "Cf=Pf*So" and "Cs=Ps*Go" into it to obtain, where Cr is the material cost of the plush toy, Cf is the fabric cost of the plush toy, and Cs is the filling cost of the plush toy; the fabric unit price Pf is determined by the supplier's quotation, and it is noted that the price differences of different materials such as short plush and coral fleece are distinguished; the fabric consumption So is calculated through the above step S51; the filling unit price Ps is determined according to the market price of filling types such as PP cotton and down cotton, generally PP cotton≈30kg / m 3 , down cotton ≈ 20kg / m 3 ; The filler weight Go is calculated through the above step S52; the auxiliary material cost Ca includes the costs of sutures, buttons, decorations, etc., which are counted according to the purchase list; the loss rate μ is set according to production experience, taking into account the losses such as fabric cutting and filler compression, generally 5-15%, of which fabric is 5-10% and auxiliary materials are 3-5%. The fabric cutting and filler compression losses need to be adjusted according to production experience; the labor cost Cb is calculated by process hours * local wage standard; the mold sharing cost Cm is the total mold cost ÷ production quantity, which is suitable for complex parts that need to be produced by mold opening.

[0127] S56, the proofing verification unit performs a proofing test according to the plush toy design scheme predetermined by the model and its predetermined process parameters and obtains quality information of the plush toy to ensure the accuracy of the plush toy design scheme, and transmits it to the processing center;

[0128] Further explanation: According to the predetermined scheme and process parameters of the model, 3D modeling and parameter initialization are first performed, and the initial 3D model is generated based on the AI design engine (such as ControlNet), and the suture overlap (1.5-3mm) and filling density (28-32kg / m3 ) and other core parameters, verify the deformation amount (early warning threshold ±3mm) through physical simulation, and call the historical defect database (including more than 3,000 cases) to automatically identify high-risk designs (such as multi-curved surface splicing, special-shaped parts), and generate forced correction plans such as anti-detachment buckles (diameter ≥5mm); secondly, conduct virtual assembly verification, use the Blender+AI plug-in to simulate the complexity of the stitching path (curvature radius ≥5mm), automatically generate a simplified version of the topological structure or replace standard parts for high-difficulty structures (such as animal joints), and preview the filler distribution through the physical engine. If the deformation of the support structure >3mm, trigger the mesh density strengthening instruction (+15%-25%); thirdly, drive the trial production through the AI model, start a small batch of injection molding equipment, use high-precision molds (error ≤0.2mm) to generate key components, complete the multi-curved surface splicing assembly through the robotic arm, and implant the "technology core" module (including voice interaction, health management functions), and synchronously test the sensor sensitivity and circuit stability; then conduct reverse engineering calibration, perform 3D scanning (accuracy ±0.1mm) on the first version of the sample, reverse-compensate the design error (such as the deviation of facial features symmetry >0.3mm), and generate a correction parameter log; then conduct multi-modal quality inspection, detect 12 types of defects such as suture breakage and uneven filling through the AI vision system, evaluate the softness of the fluff (friction coefficient ≤0.3) in combination with the tactile sensor, and perform a tensile test (≥50N) to verify the firmness of small parts. If the detachment risk >5%, trigger the structural re-design; then conduct dynamic parameter iteration, update the ControlNet model weights according to the test results, optimize the suture path algorithm to increase the efficiency by 20%, and record the design parameters and correction records of each version through the blockchain evidence storage system to support tracing the root cause of quality problems within 5 minutes; finally, through 3D scanning+AI vision comparison, make the suture alignment accuracy ≤0.3mm, and use X-ray fluoroscopy imaging to make the filler distribution uniformity, so that the density difference ≤±2kg / m 3 , use human-computer interaction pressure testing to make the response speed of the functional module, so that the voice command delay <0.5 seconds; during this process, compare and adjust the process parameters according to the data obtained by calculating the above steps S51-S54 to ensure the optimal production quality information.

[0129] S57. The processing center compares the quality information of the plush toy sample with the quality standard of the plush toy stored in the memory: if it meets the standard, it is set as the plush toy design plan; if it does not meet the standard, it is transmitted to the alarm and the parameter adjustment is notified.

[0130] Further description, the plush toy design solution includes: 1) Intelligent core architecture: First, the AI design engine, based on Stable Diffusion, generates the initial line drawing, combines with ControlNet to precisely control the position of the sewing line and the proportion of facial features, and uses the GAN network to generate the hair texture and color gradient effect, which is a multi-modal generation model; manages the core parameter library such as size, filling density, and suture overlap amount, supports dynamic adjustment range. Secondly, 3D modeling and simulation, automatically generates a manufacturable three-dimensional model through Blender + AI plug-in, real-time detects the feasibility of curved surface sewing, and physically simulates the distribution and deformation of the filling, and automatically strengthens the parametric design system of the support structure grid density. 2) Whole-process quality prevention and control system: First, the design verification system. First is the virtual assembly test, that is, previews the sewing machine stitch pitch and the sewing path, generates multiple alternative solutions for complex structures, and conducts material compatibility testing. Then is the defect prediction model, that is, trains a CNN classifier based on more than 3000 historical defect cases, and intercepts 12 types of high-risk designs in real time. Secondly, process collaborative optimization, conducts the DFM rule library, that is, generates a simplified topology structure for multi-curved surface splicing with a sewing path complexity > Lv.5, and replaces special-shaped parts with a mold opening cost > ￥500 with standard parts + 3D printing solutions. Secondly, reverse engineering calibration, that is, 3D scans the first version of the sample, reversely compensates for the design parameter errors, and the iteration time-consuming < 20 minutes. 3) Customer customization and experience optimization: First, the interactive design platform. The Web side provides a material editor, supports real-time adjustment of more than 20 parameters such as the fluff direction and glossiness, and the rendering delay < 0.5 seconds, thus forming a real-time rendering system, and AR virtual preview, that is, projects a 1:1 toy model through the mobile phone APP, and supports the simulation of the lighting environment. Secondly, the intelligent recommendation algorithm. First, according to the user's historical data, recommends similar best-selling models through collaborative filtering, and the click-through conversion rate is increased by 40%. Then is the mandatory safety design rule, that is, automatically locks the size of small parts for children's toys and disables sharp edges. 4) Data-driven iteration system: First, the quality data closed-loop, stores each version of the design data on the blockchain, and the traceability time limit for quality problems < 3 minutes. Secondly, updates the defect feature library every week, retrains the weights of the AI model, and the monthly average prediction accuracy is increased by 1.2%. Secondly, device linkage control, that is, the design parameters are directly passed to the production equipment, reducing manual transcription errors, and the data transfer accuracy reaches 100%. 5) Production process: Material adaptation parameters such as the rotation speed of the short plush fabric penetrator and the PP cotton filling loss rate, sewing control parameters such as the intelligent sewing stitch pitch range, the tension at the limb joints, the fabric utilization rate, and the dynamic filling amount, quality inspection and compliance parameters such as the diameter of the facial features parts, the formaldehyde content, the seam strength, the defect detection rate, and the suture deviation rate, efficiency and flexible production parameters such as the daily production capacity of a single machine, the delivery rate of customized orders, and the comprehensive energy consumption reduction rate.If the quality does not meet the standard (the detection standard is detailed in the subsequent step S70), then proceed to the following step S58 to identify the abnormal category and readjust the parameters and improve the design plan. If the quality meets the standard, it proves that the design plan of the proofing is qualified and trial production can be carried out to confirm the stability of its quality.

[0131] S58. The abnormal recognition unit matches the real-time obtained plush toy production image with the corresponding image in the trained convolutional neural network model and confirms its abnormal category for proofing parameter adjustment.

[0132] Further explanation: According to the program command, the image is input into the model, and the trained model is used to predict the new monitoring image. According to the output result, the abnormal type of the plush toy production quality is judged. If the judgment of the plush toy scattered production quality abnormality by the model reaches a certain threshold, the warning system is triggered and corresponding solutions are taken in a timely manner; the solutions to the abnormal plush toy design quality include: 1) Design verification and defect prediction: First, multi-model collaborative verification is carried out. The rationality of the design is verified through the double engines of the "line drawing generation model + 3D modeling model". 100% prediction and interception of 12 types of design defects such as abnormal limb proportions and suture overlap errors > 0.3mm are achieved, and a virtual assembly system is established to simulate the fluff density of 30kg / m 3Filling deformation under certain conditions can identify weak points of the support structure in advance; secondly, historical defect map matching, calling up the design defect database containing more than 2000 abnormal cases in the past three years, automatically matching similar risks through the cosine similarity algorithm, and automatically adding correction labels to high-risk designs; 2) Dynamic parameter optimization: First, closed-loop control of key parameters, dynamic calibration of core parameters such as creative value, hair roughening, and three-view alignment, a single iteration takes less than 30 seconds, and the cross-model compensation algorithm is enabled. When the "plush texture model" outputs abnormalities, the three-dimensional modeling model is automatically called for parameter compensation; secondly, user-customized risk prevention and control, enabling constraint rules for custom parameters such as size > 100cm and special-shaped facial features, including automatic strengthening of the support structure grid density for unconventional sizes, and mandatory addition of embedded anti-drop buckles and other safety redundant designs for special-shaped parts; 3) Design process collaborative correction: First, manufacturable Real-time detection of sewing feasibility, the process simulation system automatically detects the feasibility of sewing, and automatically generates simplified solutions for designs with high sewing path complexity. The material adaptation engine recommends the optimal fabric combination in real time to avoid the physical property conflict between high-elastic fabrics and hard accessories; secondly, reverse engineering correction, reverse reconstruction is initiated for design solutions with a rework rate of >5%, and physical deviation data is obtained through point cloud scanning to reversely optimize the initial line draft parameters; 4) Iterative upgrade of quality data: first, the construction of a defect feature library, updating 200+ newly collected defect features every week, which are used to retrain the ControlNet model weights, and establishing a design quality KPI dashboard to monitor core indicators such as the first version pass rate and user secondary modification rate in real time; secondly, blockchain evidence storage and traceability, each design solution generates a unique quality ID, associated with design logs, parameter adjustment records and sample quality inspection reports, and the time limit for locating quality problems is shortened to 5 minutes. Through the above measures combined with the MES system to trace abnormal batches, and link the SPC tool to achieve a closed-loop quality of the entire process; when the above problems occur in the trial production of plush toys, take the corresponding corrective measures and return to the above step S56 for re-debugging to ensure the accuracy of the plush toy proofing quality to meet customer requirements.

[0133] S60, the trial production verification module conducts automatic small batch trial production and quality inspection through the plush toy automatic equipment according to the set plush toy design plan to ensure the accuracy and stability of the subsequent mass production quality, and transmits it to the processing center;

[0134] See also Figure 14 As shown, the step S60 includes the following steps:

[0135] S61, the parameter setting unit solidifies and synchronizes the core parameters, performs reverse engineering calibration, and presets the multi-modal quality inspection parameters according to the set process parameters to ensure the standardization of trial production and large-scale flexible production requirements, and transmits them to the cutting and sewing unit;

[0136] For further explanation, first lock the design template according to the program command, and based on the final parameters verified by proofing (stitch overlap 1.5 - 3 mm, filling density 28 - 32 kg / m 3 ), solidify the design file through the blockchain evidence storage system to ensure zero error in the transfer of trial production parameters, disable the dynamic adjustment function, and force the use of standardized AI model weights to avoid the risk of production fluctuations; secondly, the process parameters are directly passed through. Write the optimized sewing path (curvature radius ≥ 5 mm) and stitch density parameter (4.2 stitches / cm) directly into the device control system, support seamless docking of the modular AI production line, and synchronize the fluff texture direction recognition parameter (detection range 0 - 360°) to ensure that the cutting error of the laser cutting machine ≤ 0.5 mm; then perform 3D scan compensation, conduct high-precision scanning (±0.1 mm) on the first version of the trial production sample, reverse-compensate the symmetry deviation of key components, and generate a parameter correction log, and update the filling distribution algorithm (density difference threshold ±2 kg / m 3 ) in real time through the edge computing node; then set the AI vision inspection standard, preset 12 types of defect detection models (such as stitch breakage, missing parts), use the YOLOv8 algorithm to achieve high-speed detection of 60 pieces per minute, the missed detection rate ≤ 0.5%, and link the color difference detector (ΔE ≤ 1.5) and the touch sensor (friction coefficient ≤ 0.3) to establish a dual quality inspection standard for appearance and touch; finally, determine the functional module test parameters, set the voice command response delay < 0.5 seconds, the qualified rate of circuit board soldering ≥ 99.9%, and verify the stability of the AI chip through the pressure test.

[0137] S62. The cutting and sewing unit performs high-precision laser cutting with an AI-assisted defect pre-inspection, automated sewing execution, and real-time quality monitoring through a laser cutting machine to shorten the development cycle of a single toy, reduce fabric waste, and transfer it to the filling and shaping unit;

[0138] Further explanation, according to the program commands and set process parameters, first, high-precision laser cutting is carried out. An industrial-grade laser cutting machine (cutting error ≤ 0.5 mm) is used to complete the fabric cutting. Combining with an AI vision system to detect the direction of plush texture in real time (detection range of 0 - 360°), ensuring the consistency of the plush direction. And the layout plan is optimized through algorithms, reducing the fabric waste rate to less than 5%, and supporting the quick switching of multiple batches of orders; secondly, AI-assisted defect pre-inspection is carried out. During the cutting process, fabric defects such as color difference and stains are synchronously identified, the abnormal areas are automatically marked, and the unqualified cut pieces are removed. The missed inspection rate ≤ 0.3%; then, automated sewing is executed. A high-speed lockstitch sewing machine (stitch pitch accuracy ± 0.1 mm) is used to perform standardized sewing. For complex structures such as animal ears and joints, modular prefabricated parts are assembled. The overlap amount of the stitches is strictly controlled within 1.5 - 3 mm, and multi-curved surface splicing is completed by a robotic arm. The curvature radius of the sewing path ≥ 5 mm, avoiding wrinkles or deformations caused by traditional manual operations; finally, real-time quality monitoring is carried out. An industrial camera (resolution of 12 million pixels) is set to monitor the stitch density (4.2 stitches / cm). If thread breakage or stitch skipping (continuous 2 abnormal stitches) is detected, the machine will automatically stop and alarm. And a tensile tester is combined to verify the sewing strength (≥ 50 N), and secondary reinforcement is carried out for high-risk areas such as the ear connection points.

[0139] S63. The filling and shaping unit injects materials with high precision, integrates functional modules, optimizes the three-dimensional shape, corrects dynamic shaping, and provides real-time data feedback through an automatic filling machine to improve the filling uniformity and reduce material waste, and transmits it to the collaborative maintenance unit;

[0140] Further explanation, according to the program commands and set process parameters, first, precise material injection is carried out. The injection amount of the filler is dynamically adjusted for the automatic filling machine through a pressure sensor, and the error rate is controlled within ± 2%, ensuring that the unit volume density is stable in the range of 28 - 32 kg / m 3 range, and an X-ray imaging system is used to monitor the filler distribution in real time. If the density difference exceeds ± 2 kg / m 3, The AI automatically triggers the supplementary filling program; secondly, for the function module integration, a pre-installed AI chip (supporting voice interaction and health monitoring) is implanted during the filling process, and the circuit board encapsulation is completed through an automated welding device, with a welding qualification rate ≥ 99.9%; then, for the three-dimensional shape optimization, the surface texture of the plush toy is generated based on the GAN algorithm, and the three-dimensional fluff effect is achieved by combining a hybrid tool (with the number of steps set from 100 to 300 steps), and the "shrinkage and expansion" effect (parameters from -20 to +20) and roughening treatment (detail threshold from 50 to 80) are executed to enhance the softness of the toy's touch; then, for the dynamic shaping correction, the first-piece sample is 3D scanned (with an accuracy of ±0.1 mm), and the symmetry deviation is compensated inversely (for example, if the position error of facial features > 0.3 mm, it is automatically corrected); finally, for the real-time data feedback, the filling data of each batch (including more than 200 monitoring points) is uploaded to the blockchain, and quality problems can be traced back to the specific filling station within 3 minutes. The filling algorithm is updated weekly through edge computing nodes, and the monthly average optimization efficiency is increased by 1.5%.

[0141] S64. The collaborative maintenance unit ensures the stability of the trial production through multi-machine cooperation network, system dynamic scheduling, data-driven maintenance, maintenance task automation, maintenance execution and optimization, and transmits it to the quality control unit;

[0142] Further explanation: According to the program commands and set process parameters, first, the swarm brain network architecture is adopted to integrate the cloud inference nodes and local skill nodes, realizing collaborative handling, sorting, and precision assembly tasks of multiple WalkerS robots. Through the perception technology driven by Transformer, the device task load is dynamically allocated to support large-scale flexible production of more than 5,000 pieces per day. Secondly, for system dynamic scheduling, edge computing nodes are set to monitor the operation status of the devices in real time, and the operation parameters such as the filling machine pressure and sewing machine speed are automatically adjusted in combination with the production demand priority. When a device fails, the system automatically switches the task to a standby device, and the downtime response time is ≤ 30 seconds. Thirdly, for the drive maintenance model, vibration / temperature sensors are deployed on key devices such as filling machines and 3D scanners to collect operation data every 5 seconds. The failure probability in the next 72 hours is predicted through the LSTM algorithm, and an equipment health index evaluation system is established. When the index is lower than 80%, a warning is triggered, and the maintenance response time is shortened to within 15 minutes. Fourthly, for maintenance task automation, the robot autonomously performs lubrication and precision calibration, and the maintenance process is synchronously uploaded to the blockchain for evidence storage. The service life of vulnerable parts such as sewing machine needles is counted, and a replacement instruction is automatically pushed when the cumulative use reaches 200,000 times. Then, for the maintenance work order system, AR glasses are used to guide manual maintenance, the equipment disassembly guide and parameter standards are superimposed in real time, and a performance test is automatically carried out after the maintenance is completed. Production can only be restarted if the pass rate ≥ 99.3%. Finally, for the iteration of the maintenance knowledge base, more than 2,000 maintenance work order data are analyzed every week, and the equipment maintenance cycle is optimized through reinforcement learning, reducing the overall failure rate by 0.8% per month on average, and establishing a full life cycle file for the equipment. The cumulative operation data is used to improve the design of the next generation of equipment.

[0143] S65. The quality control unit conducts real-time calibration and error correction of process parameters, AI vision inspection, pressure testing, traceability, and improvement through quality monitoring equipment to achieve digital quality monitoring and meet customer quality requirements.

[0144] Further explanation: Before production, strict supplier screening and raw material pre-inspection are carried out. A dynamic supplier evaluation model is established, and qualified suppliers are screened based on historical delivery pass rates and environmental certifications. An X-ray detector is set to screen for foreign objects in fillers such as PP cotton, synchronously detecting the fabric color difference and tear strength, and encrypting and storing the raw material batch information through blockchain technology to trace the source of abnormal materials within 3 minutes. According to the program commands and set detection control parameters, first, the process parameters are calibrated in real time. IoT sensors are implanted in the sewing process to dynamically monitor the stitch density and stitch overlap. Automatic shutdown occurs when the threshold is exceeded, and the reinforcement learning algorithm is applied to optimize the filling density, and the pressure sensor compensates in real time by ±2 kg / m 3Density deviation; secondly, intelligent error correction is carried out. A virtual production line is constructed based on digital twin technology to predict 80% of the mold tolerance offset risk in advance, reducing the defective rate by 37.6% in the trial production stage; then AI vision inspection is carried out. The YOLOv8 algorithm is used to achieve high-speed detection of 60 pieces per minute, covering 12 types of defects, with a missed detection rate ≤ 0.5%. The symmetry error of facial features and the overall shape deviation are detected by comparing the finished product with the design model through 3D scanning; then pressure testing is carried out. A 0.5-second response delay test is carried out on the voice chip and the circuit board, with a welding qualification rate ≥ 99.9%. The pulling scene of children is simulated to verify the tensile strength of the sewing part and the firmness of the accessories; finally, the full-link data is connected. The QMS and ERP / MES systems are integrated to achieve real-time collection and cross-platform analysis of more than 2,000 quality parameters per second, reducing the quality problem handling cycle from 72 hours to 4.8 hours. The abnormal process is located through the root cause analysis engine, and the AI quality inspection model is updated monthly based on the edge computing node, with the monthly average defect recognition accuracy increasing by 1.2%.

[0145] S70. The processing center compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: if it meets the standard, it notifies that mass production can be carried out formally; if it does not meet the standard, it is transmitted to the alarm and rework or scrapping is notified.

[0146] Further explanation, the quality standard of the plush toy includes: 1) Material safety: The fabric needs to use environmentally friendly materials such as soft and skin-friendly organic cotton and natural cotton. The color fastness needs to reach ΔE ≤ 1.5 to avoid the risk of fading, and there is no peculiar smell and no residue of irritating chemical substances; The abrasion resistance ≥ 5000 times of friction test, and there shall be no obvious pilling or damage on the surface of the fabric; The filling density is 28 - 32 kg / m 3Scope: Use non-toxic materials such as polyester fiber or acrylic cotton. The resilience should recover to its original state within 3 seconds. Foreign object detection is carried out through X-ray screening to ensure no hard impurities or contaminated particles. For small parts such as eyes and noses, embroidery or ultrasonic welding processes are used. The tensile strength should be ≥50N to avoid the risk of detachment. Electronic components such as voice chips need to pass the 0.5-second response delay test, and the welding qualification rate should be ≥99.9%; 2) Safety compliance: Comply with the GB6675 national standard. The edges of the toy should be smooth without sharp corners. The overlap of the stitches should be controlled within 1.5 - 3mm to avoid scratching children. All parts need to pass the tensile test with a tensile force of ≥70N to ensure no loose threads or accessories fall off; The formaldehyde content should be ≤20mg / kg. The detection of harmful substances such as azo dyes and total lead complies with the EN71 and ASTM F963 international standards. The packaging plastic should be printed with warning slogans and punched to prevent the risk of suffocation; 3) Production process: The cutting error should be ≤0.5mm, the layout utilization rate should be ≥95%. The consistency of the plush texture direction needs to be calibrated by AI vision. The sewing stitch density requirement is 9 - 12 stitches per inch (4.2 stitches / cm ± 0.1). Key parts such as joint connections need to be reinforced twice; The first piece needs to be compared by 3D scanning (the morphological deviation should be ≤1.5%). In mass production, the YOLOv8 algorithm is used to achieve high-speed quality inspection of 60 pieces per minute. The finished product needs to pass the drop test in the box (drop 3 times from 36 inches) and the wash resistance test (≥5 machine washes without deformation); 4) Packaging and labeling: The hangtag should indicate the material composition, safety certification, and the suitable age range for over 3 years old. The blockchain traceability information should be complete and traceable. The outer box printing should clearly mark the plush direction, the type of filling, and the production batch number. If the production does not meet the standards, it is notified to return to step S60 for rework. If it meets the standards, it proves that the plush toy design scheme is completely qualified. The plush toys produced according to this scheme have stable quality and meet the customer's requirements, and can be formally mass-produced. In this way, after model training, the plush toy design can achieve intelligent design generation (shortening the development cycle), data-driven optimization (enhancing personalization), flexible production mode (reducing costs), and the multi-scenario expansion ability for future applications (integration of education + entertainment), realizing a comprehensive upgrade of traditional automation technology, and can deeply bind creativity, production, and user needs, promoting the transformation of plush toys from "standardized products" to "emotional service carriers".

[0147] Further explanation: The above steps are displayed in sequence according to the arrows, but these steps do not necessarily need to be executed in the order indicated by the arrows. Unless specifically stated in this document, there is no strict order requirement for the execution of these steps. These steps can also be executed according to other orders, and some of the steps can include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily need to be completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential either, but can be executed alternately or in rotation with at least a part of other steps or sub-steps or stages of other steps.

[0148] Further explanation: The present invention is described in accordance with the content implemented by the software program of a kind of AI-based plush toy design control system described above. For each AI-based plush toy design method, it is divided into several modules or units to implement the software program instructions generated by each step. The software program instructions include a kind of AI-based plush toy design method described above.

[0149] The system of the present invention further includes a computer-readable storage medium containing a memory; the memory stores a computer program, and when each functional module executes the computer program, it realizes the steps of a kind of AI-based plush toy design method described in any one of the above; a computer program is stored on the computer-readable storage medium, and when the computer program is executed by each functional module, it realizes the steps of a kind of AI-based plush toy design method described in any one of the above.

[0150] Further explanation: The computer-readable storage medium includes a memory, which can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the programs and instructions corresponding to a kind of AI-based plush toy design method in the present invention, including the information transfer instructions of each module; the memory executes various functional applications and data processing of each module by running the stored non-volatile software programs and instructions, that is, it realizes a kind of AI-based plush toy design method of the above process embodiment; one or more units are stored in the memory, and when executed by the one or more modules, they execute a kind of AI-based plush toy design method in any of the above process embodiments; the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by one or more modules, it can also be a kind of AI-based plush toy design method in any of the above process embodiments.

[0151] Further, the computer program is divided into multiple modules / units, which are stored in the memory and executed by each of the modules / units to implement the present invention; the multiple modules / units may be a series of computer program instructions capable of performing specific functions, and these instructions are used to describe the execution process of the computer program in each of the modules / units.

[0152] The system of the present invention further includes an AI-based plush toy design control device, which is implemented by using the above-mentioned AI-based plush toy design method.

[0153] Further, the control device is composed of the above-related monitoring devices or apparatuses. According to the production needs of plush toy enterprises, it can be made into a complete set of plush toy automatic production lines together with plush toy production detection equipment, or made into unit devices of each functional module belonging to each step or the entire control device. When installed, wireless connection is used to connect each control device to the plush toy production equipment; the structures of these devices will not be described in detail here.

[0154] For those skilled in the art, it is obvious that this application is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of this application, this application can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of this application is defined by the appended claims rather than the above description. Therefore, it is intended to cover all changes within the meaning and scope of the equivalent elements of the claims in this application.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of this application can be modified or equivalently replaced, and all should be included in the protection scope of this application.

Claims

1. An AI-based plush toy design method, characterized in that : The method is applied to an AI-based plush toy design control system, which includes a data acquisition module, a model construction module, a model training module, a model evaluation module, a solution determination module, a trial production verification module, a wireless communication module, a memory, an alarm, a processing center, and an intelligent mobile terminal; the data acquisition module, the model construction module, the model training module, the model evaluation module, the solution determination module, the trial production verification module, the wireless communication module, the memory, and the alarm are respectively connected to the processing center; the intelligent mobile terminal is wirelessly connected to the wireless communication module within the range of a wireless network or the Internet; the wireless communication module is provided with a wireless network unit, which is responsible for the transceiver of wireless signals and automatically forms a network connection with the intelligent mobile terminal within an effective network range; the method includes the following steps: S10. Before trial production, the data acquisition module acquires the historical quality data of similar plush toys in terms of vision, user interaction, market and design preferences, synthesis and enhancement, and function verification, preprocesses and processes it for subsequent model construction, and transmits it to the model construction module; S20. The model construction module models by performing key feature extraction, model design, integration optimization, parameter initialization, dynamic optimization, model fusion and deployment based on the verified data for subsequent model training and verification, and transmits it to the model training module; S30. The model training module trains the model based on the collected data, adjusts the model parameters to optimize the model performance for a predetermined plush toy design solution, ensures the accuracy of subsequent proofing quality, and transmits it to the model evaluation module; S40. The model evaluation module verifies and evaluates the model based on the trained model, and adjusts and optimizes the model according to the verification results to improve the performance and accuracy of the model, and transmits it to the solution determination module; S50. The solution determination module conducts a proofing test based on the plush toy design solution predicted by the trained model to obtain the best plush toy design solution to ensure the accuracy of subsequent trial production quality, and transmits it to the trial production verification module; S60. The trial production verification module conducts automatic small-batch trial production and quality inspection through plush toy automatic equipment according to the set plush toy design solution to ensure the accuracy and stability of subsequent mass production quality, and transmits it to the processing center; S70. The processing center compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: if it meets the standard, it notifies that formal mass production can be carried out; if it does not meet the standard, it transmits it to the alarm and notifies rework or scrapping.

2. The method for designing a plush toy based on AI according to claim 1, wherein : The system further includes: The wireless communication module is provided with a wireless network unit, which is responsible for the transceiver of wireless signals and automatically forms a network connection with the intelligent mobile terminal within the effective network range; The alarm compares the actual evaluation score of the model with the model evaluation score standard stored in the memory. If it does not meet the standard, it will automatically give an audible alarm and notify to continue training; And it compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory. If it does not meet the standard, it will automatically give an audible alarm and notify to adjust the parameters; And it compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory. If it does not meet the standard, it will automatically give an audible alarm and notify rework or scrapping; The memory is responsible for storing the information of the data acquisition module, model construction module, model training module, model evaluation module, scheme determination module, trial production verification module, wireless communication module, and alarm, as well as storing the model evaluation score standard and the plush toy quality standard; The processing center is responsible for the information transfer of the data acquisition module, model construction module, model training module, model evaluation module, scheme determination module, trial production verification module, wireless communication module, alarm, and memory. It is the hub center of the system, and compares the actual evaluation score of the model with the model evaluation score standard stored in the memory: If it meets the standard, it is the designed scheme for the predetermined plush toy. If it does not meet the standard, it is passed to the alarm and notifies to continue training; And it compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: If it meets the standard, it is set as the plush toy design scheme. If it does not meet the standard, it is passed to the alarm and notifies to adjust the parameters; And it compares the quality information of the plush toy with the quality standard of the plush toy stored in the memory: If it meets the standard, it notifies that mass production can be carried out officially. If it does not meet the standard, it is passed to the alarm and notifies rework or scrapping; The data acquisition module includes an information collection unit, a cleaning and repair unit, a segmentation and extraction unit, a standard enhancement unit, a conversion and encoding unit, and a physical verification unit, which is responsible for obtaining the historical quality data of visual, user interaction, market and design preferences, synthesis and enhancement, and functional verification of similar plush toys, preprocessing and processing, and passing it to the model construction module; The model construction module includes a feature extraction unit, a model design unit, an integration and optimization unit, a parameter initialization unit, a dynamic optimization unit, and a network layer unit, which models by performing key feature extraction, model design, integration and optimization, parameter initialization, dynamic optimization, model fusion and deployment based on the verified data, and passes it to the model training module; The model training module includes a convolution input unit, a pooling conversion unit, a fully connected layer unit, a normalization output unit, an output conversion unit, and a backpropagation unit, which trains the model according to the collected data, adjusts the model parameters to optimize the model performance for the predetermined plush toy design scheme, and passes it to the model evaluation module; The model evaluation module includes a classification and marking unit, a learning and training unit, a cost verification unit, and an application evaluation unit, which verifies and evaluates the model according to the trained model, adjusts and optimizes the model according to the verification results, and passes it to the scheme determination module; The solution determination module includes a fabric consumption unit, a filling weight unit, a seam strength unit, a joint torque unit, a cost accounting unit, a proofing verification unit, and an anomaly identification unit. It conducts proofing tests based on the plush toy design solution predicted by the trained model to obtain the optimal plush toy design solution and transmits it to the trial production verification module. The trial production verification module includes a parameter setting unit, a cutting and sewing unit, a filling and shaping unit, and a quality control unit. It conducts automatic small-batch trial production and quality inspection of plush toys through automatic plush toy equipment according to the set plush toy design solution and transmits it to the processing center.

3. A method for designing a plush toy based on AI according to claim 1, characterized in that : The step S10 includes the following steps: S11. The information collection unit obtains various normal and abnormal quality data, images, and videos of similar plush toys by matching with the industry database according to the customer's requirements for subsequent data cleaning and transmits it to the cleaning and repair unit. S12. The cleaning and repair unit enhances the reliability of commercial design through noise elimination, blurring repair, local defect image repair, 3D model repair, highlight area processing, and texture consistency optimization and transmits it to the segmentation and extraction unit. S13. The segmentation and extraction unit conducts semantic segmentation and main body separation, reflection and highlight suppression, binarization and edge detection, contour repair and filling, and cross-modal semantic alignment to provide high-quality data for 3D modeling and transmits it to the standard enhancement unit. S14. The standard enhancement unit enhances the modeling ability of material diversity and structural complexity through size normalization, color space conversion, histogram equalization, binarization processing, geometric transformation enhancement, and material and lighting perturbation and transmits it to the conversion and coding unit. S15. The conversion and coding unit provides reliable input for intelligent generation and optimization through numerical data standardization, feature coding, data specification and feature extraction, continuous feature binning, and synthetic feature generation and transmits it to the physical verification unit. S16. The physical verification unit can avoid physical contradiction design and process conflicts through material parameter verification, mechanical simulation testing, ergonomic adaptation, and process constraint mapping to ensure that the generated solution has both innovation and feasibility.

4. The design method of a plush toy based on AI according to claim 1, characterized in that : The step S20 includes the following steps: S21. The feature extraction unit improves the model training speed through invalid data filtering, physical attribute extraction, user preference feature coding, key feature screening, dimension compression, and data hierarchical division and transmits it to the model design unit. S22. The model design unit provides a direction for subsequent model training through the application of generative adversarial networks, multi-modal input fusion, generator network design, discriminator network optimization, and transfer learning application and transmits it to the integration and optimization unit. S23. The integration and optimization unit speeds up subsequent model training through physical constraint embedding, regularization optimization, interpretability enhancement, model compression, and quantization-aware training and transmits it to the parameter initialization unit. S24. The parameter initialization unit provides efficiency for subsequent model training through pre-training transfer, word vector pre-training, random initialization, distribution initialization, material parameter constraint, and style feature enhancement and transmits it to the dynamic optimization unit. S25. The dynamic optimization unit provides security performance for subsequent model training through residual connection parameter configuration, regularization fusion, pre-training fine-tuning, module-based differentiation, initialization sensitivity testing, and production environment adaptation, and transmits it to the network layer unit. S26. The network layer unit obtains the convolutional neural network model architecture of the input layer, convolutional layer, activation layer, pooling layer, and fully connected layer based on feature and parameter optimization for subsequent model training.

5. A method for designing a plush toy based on AI according to claim 1, characterized in that : The step S30 includes the following steps: S31. The convolution input unit obtains the convolution input value and activates the function input according to the convolution input calculation formula "z(t) = ∫x(m)y(t - m)dm, where z(t) is the output function, x(t) is the input function, y(t) is the convolution kernel function, and dm is the integral microelement of the variable m", and transmits it to the pooling conversion unit. S32. The pooling conversion unit obtains the maximum pixel value after pooling according to the average pooling calculation formula "Z(I,j) = mean(X[i*Ps(i + 1)*Ps,j*Ps(j + 1)*Ps])), where Z(I,j) is a pixel value in the output feature map after pooling, (i,j) is the coordinate in the output feature map, mean() is to obtain the average value of all pixel values in the input window, X is the input feature map, and Ps is the hyperparameter" to retain important feature information and transmits it to the fully connected layer unit. S33. The fully connected layer unit obtains the output data of the connection layer according to the fully connected layer calculation formula "y = f(∑(Wn*Xn)+b), where y is the output result, f is the activation function, Wn is the weight of the nth input feature, Xn is the n input features, n is the dimension of the input features, and b is the bias" for subsequent output calculation and transmits it to the normalization output unit. S34. The normalization output unit obtains the BN layer output according to the Batch Norm layer calculation formula "h = φ(Bn(Wx + b)), where h is the output of the fully connected layer, φ is the activation function, Bn is the operator of batch normalization, W is the weight parameter, X is the input of the fully connected layer, and b is the bias parameter" for the output of model data and transmits it to the output conversion unit. S35. The output conversion unit obtains the size of the output value after convolution according to the output layer conversion calculation formula "N = (P - F + 2C) / S + 1, where N is the output size after convolution, P is the input size before convolution, F is the convolution kernel size, C is the number of layers adding 0 around the image, and S is the step size" for subsequent model analysis and training and transmits it to the backpropagation unit. S36. The backpropagation unit obtains the backpropagation of the convolution output value according to the backpropagation calculation formula "dz[l] = da[l]*g[l]′(z[l]), where dz[l] is the input gradient, da[1] is the output gradient, g[l]′(z[l]) is the derivative of the activation function, and z[l] is the input of the activation function" to optimize the performance of the network model.

6. The method for designing a plush toy based on AI according to claim 1, wherein : The step S40 includes the following steps: S41. The classification and marking unit classifies and marks the corresponding data sample categories according to the abnormal categories of the plush toy design quality for subsequent machine learning of the model and transmits them to the learning and training unit; S42. The learning and training unit determines the training model according to the gradient descent data, inputs the plush toy quality data in the training set into the model for simulation training to ensure the accuracy of model recognition, and transmits it to the cost verification unit; S43. The cost verification unit calculates the cost function according to the formula "J(θ) = -1 / m ∑m ∑K [y k (i) log(h θ (x(i))k) + (1 - y k (i)) log(1 - (h θ (x(i))k)] + λ / 2m ∑L-1 ∑sl ∑sl+1 (θj,i(l))2, where J(θ) is the cost function, m is the total number of training samples, K is the total number of categories, y k (i) is the true label value of the sample, h θ (x(i))k is the predicted probability of the sample, λ is the regularization parameter, L is the total number of layers of the neural network, sl is the number of neurons in the l-th layer, and θj,i(l) is the weight parameter between the connected neurons", to obtain the cost function value after model training, ensure the accuracy of model evaluation, and transmit it to the application evaluation unit; S44. The application evaluation unit obtains the comprehensive evaluation score F according to the model evaluation score calculation formula "F = 2(Ac * Re) / (Ac + Re), where F is the training model evaluation score, Ac is the precision rate, and Re is the recall rate", and transmits it to the processing center; S45. The processing center compares the actual evaluation score of the model with the model evaluation score standard stored in the memory: if it meets the standard, it is the predetermined plush toy design plan; if it does not meet the standard, it is transmitted to the alarm and notifies to continue training.

7. A method for designing plush toys based on AI according to claim 1, characterized in that : The step S50 includes the following steps: S51. The fabric consumption unit obtains the fabric consumption of the plush toy according to the fabric consumption calculation formula of the plush toy "So = ∑(Si * Ni) / η, where So is the fabric area of the plush toy (m 2 ), Si is the single-piece area of the i-th cut piece of the plush toy (m 2 ), Ni is the number of the i-th cut piece of the plush toy, η is the fabric utilization rate of the plush toy, and N is the total number of types of cut pieces of the plush toy" to balance cost and efficiency and transmit it to the filling weight unit; S52. The filling weight unit obtains the filling weight of the plush toy according to the formula for calculating the filling weight of the plush toy "Go = V * ρ * (1 + k)", where Go is the filling weight of the plush toy (g), ρ is the density of the filling material of the plush toy (g / m 3 ), V is the volume of the plush toy (m 3 ), and k is the processing loss rate of the plush toy", so as to control the filling cost and transmit it to the seam strength unit; S53. The seam strength unit obtains the seam strength of the plush toy according to the plush toy seam strength calculation formula "F = S * (L / A) * ωo, where F is the seam bearing force of the plush toy (N), S is the single-needle sewing thread strength of the plush toy (N / st), L is the seam length of the plush toy (cm / st), A is the stitch spacing of the plush toy (cm), and ωo is the production safety factor of the plush toy", to avoid toy damage or safety hazards caused by sewing thread breakage, and transmits it to the joint torque unit; S54. The joint torque unit obtains the torque of the movable joints of the plush toy according to the formula for calculating the torque of the movable joints of the plush toy "τ = ρ * V * g * L * ωa, where τ is the joint torque of the plush toy (N·m), ρ is the density of the filling material of the plush toy (g / m 3 ), V is the volume of the plush toy (m 3 ), g is the acceleration due to gravity of the plush toy (m / s 2 ), L is the length of the lever arm of the plush toy (m), and ωa is the safety factor for the use of the plush toy", so as to improve the mobility and durability of the plush toy and transmit it to the cost accounting unit; S55. The cost accounting unit obtains the total cost of the plush toy according to the plush toy cost calculation formula "Ct = [(Pf * So) + (Ps * Go) + Ca] * (1 + μ) + Cb + Cm, where Ct is the total cost of the plush toy (CNY), Pf is the unit price of the plush toy fabric (CNY / ㎡), So is the fabric consumption of the plush toy (㎡), Ps is the unit price of the plush toy filling (CNY / kg), Go is the weight of the plush toy filling (kg), Ca is the auxiliary material cost of the plush toy (CNY), μ is the loss rate of the plush toy, Cb is the labor cost of the plush toy (CNY), and Cm is the mold sharing cost of the plush toy (CNY)" to quickly estimate the standardized production cost of the plush toy and transmits it to the proofing verification unit; S56. The proofing verification unit conducts a proofing test according to the predetermined plush toy design plan and its predetermined process parameters of the model and obtains the quality information of the plush toy to ensure the accuracy of the plush toy design plan and transmits it to the processing center; S57. The processing center compares the quality information of the plush toy sample with the quality standard of the plush toy stored in the memory: if it meets the standard, it is set as the plush toy design plan; if it does not meet the standard, it is transmitted to the alarm and notifies to adjust the parameters. S58. The anomaly recognition unit matches the real-time obtained plush toy production image with the corresponding image in the trained convolutional neural network model and confirms its abnormal category for proofing parameter adjustment.

8. A method for designing plush toys based on AI according to claim 1, characterized in that : The step S60 includes the following steps: S61. The parameter setting unit performs core parameter solidification and synchronization, reverse engineering calibration, and multi-modal quality inspection parameter presetting according to the set process parameters to ensure the standardization of trial production and the requirements of large-scale flexible production, and transmits them to the cutting and sewing unit; S62. The cutting and sewing unit performs high-precision laser cutting with AI-assisted defect pre-inspection, automated sewing execution, and real-time quality monitoring through a laser cutting machine to shorten the development cycle of a single model of plush toy and reduce fabric waste, and transmits them to the filling and shaping unit; S63. The filling and shaping unit performs highly accurate material injection, functional module integration, three-dimensional shape optimization, dynamic shaping correction, and real-time data feedback through an automatic filling machine to improve the filling uniformity and reduce material waste, and transmits them to the collaborative maintenance unit; S64. The collaborative maintenance unit ensures the stability of trial production through a multi-machine cooperation network, system dynamic scheduling, data-driven maintenance, maintenance task automation, maintenance execution and optimization, and transmits them to the quality control unit; S65. The quality control unit performs real-time calibration and error correction of process parameters, AI vision inspection, pressure testing, traceability and improvement through quality monitoring equipment to achieve digital quality monitoring and meet customer quality requirements.

9. A method for designing plush toys based on AI according to claims 1-8, characterized in that : The system further includes a computer-readable storage medium containing a memory; the memory stores a computer program, and when each functional module executes the computer program, the steps of a method for designing a plush toy based on AI according to any one of claims 1-8 above are implemented; a computer program is stored on the computer-readable storage medium, and when the computer program is executed by each functional module, the steps of a method for designing a plush toy based on AI according to any one of claims 1-8 above are implemented.

10. A method for designing plush toys based on AI according to claims 1-8, characterized in that : The system further includes a control device for designing a plush toy based on AI, which is implemented by using a method for designing a plush toy based on AI according to any one of claims 1-8 above.

Images