AI-driven child customized insole full-link intelligent adaptation method

Abstract

The invention discloses an AI-driven full-link intelligent adaptation method for customized insoles for children, which belongs to the technical field of customized insoles, and comprises a foot multi-modal data acquisition step, an AI foot data processing and specialization analysis step and an AI foot data processing and specialization analysis step, inputting into an AI analysis model pre-trained by a large-scale foot database, and carrying out deep foot health state evaluation, pathological feature recognition and development trend prediction; the method comprises the steps of generating a structured foot health analysis report, customizing an insole scheme AI intelligent generation step, performing insole intelligent production and quality control, and performing wearing feedback and system self-adaptive iteration, so that the design scheme is ensured to be fit in form, the biomechanical index can be effectively improved in function, and the corrective effectiveness and comfort are remarkably improved; the defects that traditional manual design is low in efficiency and poor in consistency, and the evaluation effect is difficult to quantify are overcome.

Description

Technical Field

[0001] This invention relates to the field of customized insole technology, and in particular to an AI-driven end-to-end intelligent adaptation method for children's customized insoles. Background Technology

[0002] Foot health is a key factor affecting the physical development, gait formation, and even spinal health of children and adolescents. Common problems such as flat feet, high arches, and pronation / eversion, if not addressed promptly and correctly, can lead to a series of long-term harms, including gait abnormalities, decreased athletic ability, and chronic joint pain. Customized insoles, as an effective non-surgical intervention, can improve plantar pressure distribution and guide the foot and ankle to a neutral position by providing biomechanical support and correction, thereby alleviating symptoms and correcting deformities. However, existing methods and production technologies for custom insoles have many limitations, making it difficult to meet, especially, the precise, efficient, and large-scale fitting needs of children and adolescents in their rapid growth and development stages. These shortcomings are mainly reflected in the following aspects: First, in the data collection stage, there is a reliance on empirical and singular measurements, lacking multi-dimensional objective data support. Traditional methods often rely on manual measurements (such as calipers and protractors) and visual observation by rehabilitation therapists or insole makers, the results of which are easily influenced by subjective experience and can only obtain limited static dimensions such as foot length and width. While some advanced clinics have introduced 3D scanners, they typically only acquire static foot morphology data, lacking simultaneous collection of crucial biomechanical data such as plantar pressure distribution and joint range of motion during the dynamic gait cycle. This separation of morphological and functional data fails to comprehensively reflect the true health status and root causes of children's foot problems. Secondly, the analysis and design stages heavily rely on manual experience, resulting in low levels of standardization and precision. Currently, the design of customized insoles heavily depends on the personal knowledge and experience of designers or physicians. Designers need to rely on limited measurement data, experience to determine the type and severity of foot problems, and manually draw design sketches or use CAD software for modeling. The entire process is time-consuming, and design quality varies from person to person, making it difficult to guarantee consistency and optimization. For complex biomechanical problems, the lack of quantitative data and algorithmic models makes it difficult to achieve precise and predictable corrective effects. Furthermore, children's feet are constantly developing, and traditional methods struggle to predict developmental trends and dynamically adjust design schemes.

[0003] Existing technologies also suffer from drawbacks such as reliance on human experience and single static measurements for data collection, as well as strong subjectivity and difficulty in standardization in the analysis and design process. Summary of the Invention

[0004] To address the aforementioned problems, this invention aims to solve the issues described above. One objective of this invention is to provide an AI-driven, end-to-end intelligent adaptation method for customized children's insoles, resolving one of the problems described above.

[0005] The solution adopted in this invention is: an AI-driven end-to-end intelligent adaptation method for customized children's insoles, characterized by the following steps: Foot multimodal data acquisition steps: Static three-dimensional geometric morphology data and dynamic biomechanical data of the user's foot are collected simultaneously through non-contact 3D scanning equipment, plantar pressure distribution sensing equipment and gait dynamic capture equipment. AI foot data processing and professional analysis steps: The collected multimodal data is fused and standardized, and then input into an AI analysis model pre-trained with a large-scale foot database to perform in-depth foot health status assessment, pathological feature identification and development trend prediction, and generate a structured foot health analysis report. Customized insole solution AI intelligent generation steps: Based on the structured foot health analysis report, call the parametric design rule library, automatically generate a personalized 3D model of the insole through the AI ​​computing engine, and finely adjust the geometric shape and physical properties of key areas such as support, cushioning, and correction, and output a digital insole solution that can be directly produced; Intelligent production and quality control steps for insoles: The digital insole solution is transmitted to additive manufacturing equipment for production, and a real-time visual inspection and feedback control mechanism is introduced during the production process to ensure that the product meets the design standards; Wearable feedback and system adaptive iteration steps: The smart insole with built-in sensors collects actual wear data from users, and combines it with subjective user feedback to continuously optimize the AI ​​analysis model and design rule base, so as to realize the closed-loop iteration of the system and the continuous improvement of personalized adaptation capabilities.

[0006] The preferred technical solution is that the fusion and standardization processing and AI analysis model evaluation in the AI ​​foot data processing and professional analysis steps specifically include: Spatiotemporal alignment and registration of heterogeneous data from different acquisition devices are performed to ensure that static morphology and dynamic mechanical data are associated in a unified coordinate system. Specifically, the spatial coordinates of the plantar pressure distribution heatmap are accurately mapped to the plantar surface of the 3D foot model, and the joint angle data at each moment in the gait cycle are synchronized with the three-dimensional posture of the foot. High-dimensional feature extraction is performed on the standardized multimodal data stream. Specifically, a convolutional neural network branch is used to extract spatial morphological features from 3D point cloud data of the foot and pressure distribution images, including the arch curvature, metatarsal head distribution, and the shape of the pressure center trajectory. At the same time, a temporal processing network branch is used to model the gait sequence data to capture the dynamic patterns and abnormal patterns in the gait cycle. The extracted spatial and temporal features are fused and input into a classification and regression model under a multi-task learning framework, simultaneously performing the following tasks: multi-classifying common foot problems such as flat feet, high arches, and pronation / vaulting and outputting confidence scores; performing continuous numerical regression prediction on the severity of the identified problems, such as the angle of arch collapse and the degree of calcaneal eversion; and inferring the future trend and risk probability of foot development based on users' historical data through a temporal prediction model. Finally, by integrating all the analysis results, a structured foot health analysis report is generated, which includes quantitative parameters, qualitative conclusions, risk warnings, and visualization charts.

[0007] The preferred technical solution is that the training and optimization process of the AI ​​analysis model specifically includes: To build a large database of foot health data for children and adolescents aged 6 to 18, covering different age groups, genders, regions, and foot health status. The database will have a sample size of no less than 100,000 cases and will be continuously updated. It will include foot disease types and severity rating labels marked by professional physicians, as well as longitudinal tracking data of some users. An improved CNN-Transformer hybrid architecture is adopted as the core model. The CNN part uses a sparse convolutional network to process 3D point clouds to efficiently extract macro and micro morphological features of the foot. The Transformer part uses its self-attention mechanism to model long-range dependencies in gait sequences and accurately capture gait anomalies. In model training, an adversarial training strategy is introduced. By adding imperceptible noise perturbations to the input data, the robustness of the model to real-world interference such as differences in acquisition devices and slight changes in user posture is improved, ensuring the stability of the analysis results. By employing knowledge distillation technology, the model knowledge, which excels in morphological and gait analysis, is extracted into a lightweight unified model. This reduces the computational complexity of the model while ensuring analytical accuracy, enabling it to be deployed on edge computing devices or achieve rapid response on mobile devices.

[0008] The preferred technical solution is that the AI ​​foot data processing and professional analysis step further includes a data quality assessment and enhancement sub-step: Before the data is input into the model, a pre-trained quality assessment network is used to score the integrity and reliability of the data, identify low-quality or invalid data segments caused by user non-cooperation, equipment obstruction or signal interference, and automatically trigger re-collection or provide prompts. For cases where data is missing but not severe, generative adversarial networks or variational autoencoders are used to generate reasonable supplementary data that matches the individual characteristics of the user based on existing high-quality data samples, so as to ensure the completeness of subsequent analysis. To address the challenges of collecting data on children's feet, advanced data augmentation techniques are employed, particularly data augmentation in the feature space. For example, by changing the rotation angle of the foot model and applying simulated mechanical loads, the diversity of training data can be expanded without altering the underlying pathological characteristics.

[0009] The preferred technical solution is that the fine-tuning of key areas in the AI ​​intelligent generation step of the customized insole solution specifically includes: Based on the foot health analysis report, the AI ​​computing engine first calculates the apex position of the arch support pad, the radius of curvature of the support surface, and the gradient distribution of the support strength according to the arch type and height, to ensure that while providing sufficient support, it avoids excessive pressure on the plantar fascia. For the abnormal pressure areas revealed in the plantar pressure analysis report, the insole material properties are designed differently: under the metatarsal area where the pressure is too high, a cushioning structure with specific shock absorption performance and crumple zone is designed to disperse the impact force; in areas with insufficient pressure or that need stimulation, the perception and function are improved by adjusting the material hardness or increasing the contact area. Based on the analysis of the calcaneal inversion / valgus angle and the stability of the heel strike phase during dynamic gait, the depth of the heel cup, the inclination angle of the medial or lateral wall, and the encirclement height are refined to provide precise calcaneal wrapping and corrective torque, guiding the foot and ankle back to a neutral position. The parametric modeling engine parameterizes all the above adjustments and automatically generates a complete, watertight, 3D insole mesh model with optimal printing orientation. This model takes into account the constraints of the printing process and supports export in STL and 3MF universal formats.

[0010] A preferred technical solution is that the AI ​​intelligent generation step further includes an optimization loop based on biomechanical simulation feedback: After generating a preliminary 3D model of the insole, it is registered with a standardized foot skeletal muscle model and collected user foot morphology data. In an integrated biomechanical simulation environment, the interaction between the foot and the insole is simulated under typical movement states such as walking and running. The stress-strain distribution of the insole under load, the pressure change in a specific area of ​​the sole, and the effect of the corrective force on the joint torque were calculated by finite element analysis. The AI ​​engine automatically adjusts the design parameters of the insole based on the difference between the simulation results and the ideal biomechanical target, fine-tuning the shape of the support area and changing the distribution density of the cushioning material; After several iterations of simulation, evaluation, and adjustment, a personalized insole design scheme optimized in terms of biomechanical performance is finally output, ensuring that it not only fits the shape but also effectively improves the user's foot health.

[0011] The preferred technical solution is that, in the AI ​​intelligent generation step of the customized insole solution, a human-machine collaborative review mechanism is introduced for complex or severe foot problem cases: The AI-generated preliminary design scheme will be automatically pushed to the professional review interface of qualified physicians or rehabilitation therapists. This interface displays the original foot data, AI analysis report, insole 3D model and key biomechanical simulation results side by side. The physician-side system provides a set of professional interactive fine-tuning tools, allowing physicians to make local modifications to the AI-generated insole solutions based on their clinical experience, such as manually dragging control points to adjust the height and shape of the arch support curve, or adjusting the suggested hardness value of specific areas through sliders. All fine-tuning operations will be recorded by the system and used as new labeled data, which will be fed back into the training database of the AI ​​design model for reinforcement learning. This will enable the AI ​​model to continuously learn and align with the decision-making logic of experts, gradually reducing its reliance on human intervention in the future and achieving continuous evolution of AI design capabilities.

[0012] The preferred technical solution is that the intelligent production and quality control steps for the insoles are as follows: Industrial-grade photopolymer 3D printing technology is used. The AI ​​system automatically selects the most suitable medical-grade polymer material and optimizes printing parameters based on the mechanical properties and comfort required by the final insole design. During the printing process, layer-by-layer or real-time image monitoring is performed using a high-resolution industrial camera. Computer vision algorithms are used to compare the printed layers with the design slices, instantly identifying and locating defects such as interlayer delamination, material shortages, and deformation. Once a defect is identified, the system can automatically pause printing and recommend or directly execute parameter correction strategies based on the defect type, such as cleaning the material tank, adjusting the laser power, or regenerating the support structure, to ensure the consistency between the physical product and the digital model of the final product.

[0013] The preferred technical solution is that the wearable feedback and system adaptive iteration steps specifically involve: the smart insole has a built-in micro-pressure sensor array and inertial measurement unit, which can continuously or periodically collect data on foot pressure distribution, gait symmetry, and activity intensity during daily wear; this data is transmitted to the user's mobile terminal APP via Bluetooth Low Energy and uploaded to the cloud platform after encryption; the AI ​​system quantitatively evaluates the corrective effect and fit of the insole by comparing changes in foot biomechanical data before and after wearing; at the same time, the APP guides the user to provide subjective feedback such as comfort and pain points; the system uses these objective and subjective feedback data to dynamically adjust the feature weights in the AI ​​analysis model and the thresholds and rules in the parameterized design rule base through online learning or periodic batch retraining, so that the system can generate more accurate subsequent products for the user or provide better initial solutions for similar user groups.

[0014] The preferred technical solution is that the end-to-end method constructs a user-centric digital health record, which fully records all information from initial data collection, subsequent AI analysis reports, insole design schemes, production quality data to long-term wear feedback, forming an immutable and traceable data chain.

[0015] Compared with existing technologies, the AI-driven intelligent adaptation method for children's customized insoles of this invention has the following technical advantages: This application presents an AI-driven, end-to-end intelligent adaptation method for children's customized insoles, achieving objective, accurate, and synchronous collection of multimodal data, laying a solid foundation for scientific adaptation. By integrating 3D scanning, pressure sensing, and gait capture devices, this invention can simultaneously acquire complete morphological and functional data of the foot. The data is rich in dimensions and highly accurate, fundamentally overcoming the one-sidedness and subjectivity caused by traditional reliance on manual measurement and single static data, enabling subsequent analysis and decision-making to be based on a comprehensive and objective data foundation.

[0016] Through AI-powered intelligent analysis, this invention achieves standardized, precise, and proactive identification and assessment of foot problems. Utilizing a CNN+Transformer hybrid AI model trained on a large-scale database, the invention can automatically identify foot problems and classify their severity with high accuracy, avoiding the judgment discrepancies and instabilities inherent in traditional methods that rely on physicians' personal experience. Simultaneously, the introduced developmental trend prediction function provides forward-looking intervention guidance for children in their rapid growth phase, realizing a shift from passive treatment to proactive prevention and management.

[0017] Through parametric AI design and biomechanical simulation, the customization process achieves automation, refinement, and optimization. The AI ​​parameter calculation engine can automatically and quickly generate a 3D model of the insole containing multiple key parameters based on the analysis results, and make fine adjustments to the support, cushioning, and correction areas. Combined with an optimization loop based on biomechanical simulation feedback, this ensures that the design not only fits the body shape but also effectively improves biomechanical indicators, significantly enhancing the effectiveness and comfort of the correction. This overcomes the drawbacks of traditional manual design, such as low efficiency, poor consistency, and difficulty in quantifying and evaluating the effects.

[0018] By streamlining the digital chain from design to production and introducing AI-powered quality inspection, production efficiency and product quality have been significantly improved. Seamless integration between the design and 3D printing systems is achieved through API / RESTful interfaces, combined with an AI vision inspection module for real-time monitoring and automatic defect correction, realizing automation and intelligence in the production process. A single machine can produce 30-50 sets per day with a high product qualification rate, effectively solving core pain points of traditional production such as multiple interruptions, reliance on manual labor, large quality fluctuations, and low capacity.

[0019] A closed-loop data system has been established, encompassing data collection, design, production, feedback, and iteration, enabling the system's self-evolution and continuous optimization of personalized services. Real-world data is collected through smart sensor insoles and user feedback channels, driving the periodic iteration of AI models and design rule bases. This allows the system to become increasingly intelligent with use, continuously improving its adaptability. This not only provides individual users with long-term, dynamic foot health management solutions but also strengthens the system's technological barriers and service quality over time, forming a core competitiveness that is difficult to replicate.

[0020] The entire process complies with children's data security and privacy protection standards, ensuring the compliance and security of business operations. Through data anonymization, AES-256 encrypted transmission, and distributed secure storage, all children's personal foot information is fully protected during processing and transfer, complying with relevant laws and regulations. This removes legal and privacy obstacles for large-scale application in scenarios such as schools and medical institutions.

[0021] Other features and advantages of the invention will become clear when reading the following description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention. In these drawings, similar reference numerals are used to denote similar elements. The drawings described below are some embodiments of the invention, but not all embodiments. Other drawings will be readily available to those skilled in the art based on these drawings without any inventive effort.

[0023] Figure 1 This is a schematic diagram of the AI-driven intelligent adaptation method for the entire chain of children's customized insoles provided in a specific embodiment of the present invention; Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0025] The following description, in conjunction with the accompanying drawings and embodiments, details the AI-driven end-to-end intelligent adaptation method for children's customized insoles.

[0026] like Figure 1 As shown, an AI-driven intelligent adaptation method for children's customized insoles includes the following steps: Foot multimodal data acquisition step: Through non-contact 3D scanning equipment, foot pressure distribution sensing equipment and gait dynamic capture equipment, the static three-dimensional geometric morphology data and dynamic biomechanical data of the user's feet are collected simultaneously. The working principle of this non-contact 3D scanning device is to project coded light spots or stripes onto the foot, and one or more cameras capture the pattern deformed by the contour of the foot surface from different angles. The three-dimensional coordinates of millions of points are calculated by optical triangulation to form high-precision point cloud data, thereby obtaining complete static morphological parameters including foot length, foot width, arch height, and metatarsal circumference without contact or pressure.

[0027] A plantar pressure distribution sensor is typically a dense array platform consisting of thousands of miniature pressure sensors. When a user stands or walks on it, each sensor converts the pressure it senses into an electrical signal. After calibration and processing by imaging algorithms, a visual plantar pressure distribution map is generated, which intuitively reveals the load concentration points and pressure imbalances in different areas of the foot, such as the heel, forefoot, and arch. This is a key biomechanical indicator for assessing foot dysfunction.

[0028] Gait dynamic capture equipment uses multiple high-speed infrared cameras deployed in a test area to capture the natural optical features of the user's lower limbs and foot surfaces during walking. Through multi-view geometric reconstruction, it calculates the spatial trajectory of the foot and ankle joints, and then analyzes dynamic parameters such as joint angles, gait speed, cadence, and the duration of the stance and swing phases within the gait cycle.

[0029] Synchronous acquisition ensures that the three devices collect data at the same time or for the same specific gait phase, such as the initial heel strike phase, through a unified hardware trigger signal or software timestamp service. This solves the data mismatch problem caused by minute changes in posture during time-division acquisition, laying a solid foundation for the subsequent precise spatial alignment and fusion of static morphology, dynamic pressure, and kinematic data, thereby obtaining a more comprehensive and realistic reflection of the foot's condition.

[0030] AI Foot Data Processing and Professional Analysis Steps: The collected multimodal data is fused and standardized, and then input into an AI analysis model pre-trained with a large-scale foot database for in-depth foot health status assessment, pathological feature identification, and developmental trend prediction, generating a structured foot health analysis report. This step transforms raw data into actionable insights. Fusion and Standardization: Due to the diverse sources, the data first needs to be standardized, for example, normalizing pressure data to the 0-1 range and downsampling and redirecting 3D point clouds to a unified coordinate system. The core of fusion is spatiotemporal registration: using calibration objects or feature points, the pixel coordinate system of the two-dimensional pressure distribution image is uniformly mapped to the surface coordinate system of the three-dimensional foot model, allowing us to accurately know the specific protrusion position on the 3D model corresponding to the pressure peak at the second metatarsal head; simultaneously, the time axis of the gait cycle is aligned with the motion sequence of the 3D model. The AI ​​analysis model is a complex, end-to-end deep learning system. It accepts the fused multimodal data as input and undergoes training and optimization through the process detailed below. The model architecture is typically designed as a multi-branch network to process different types of data in parallel. Deep evaluation, identification, and prediction: The model automatically learns and extracts spatial features such as arch curvature, metatarsal head distribution morphology, and pressure center trajectory from point cloud or mesh data of a 3D foot model and pressure distribution images through convolutional neural network branches. Simultaneously, temporal processing network branches, such as Transformers, model continuous gait joint angle sequences to capture dynamic patterns and abnormalities in gait, such as excessive pronation. Subsequently, these high-dimensional features from different modalities are concatenated or fused through attention mechanisms and input into a multi-task learning framework. This framework simultaneously performs classification tasks to identify flat feet, high arches, and pronation / extension, regression tasks to predict continuous values ​​such as arch collapse angle and calcaneal eversion amplitude, and temporal prediction tasks to infer future foot development trajectories based on user historical data.

[0031] The structured foot health analysis report is not simply a conclusion, but a structured digital document containing both machine-readable and human-readable information. It integrates all analysis results, including quantitative parameters such as arch index and calcaneal eversion angle, qualitative conclusions such as moderate flexible flat feet with functional pronation, risk warnings such as a moderate risk of arch height decline in the next 12 months, and visual aids such as a 3D foot model annotated with problem areas and pressure distribution. This report provides precise and quantitative decision-making support for the next stage of personalized treatment.

[0032] The AI-powered intelligent generation process for customized insoles involves: Based on the structured foot health analysis report, calling upon a parametric design rule library, and using an AI computing engine to automatically generate a personalized 3D model of the insole. This model involves fine-tuning the geometry and physical properties of key areas such as support, cushioning, and correction, outputting a digital insole solution suitable for direct production. The parametric design rule library is a structured collection of expert knowledge stored in a database. It encodes biomechanical principles and clinical experience into conditional-action rules, such as: IF arch type = flat foot AND arch height < threshold T, THEN set arch support height = F (arch height) AND set support curvature radius = R (arch length). These rules cover the design logic for various functional areas, such as the arch, heel, and forefoot, addressing different foot problems.

[0033] AI Computing Engine: This engine parses structured foot health analysis reports, reads key parameters such as arch type, pressure peak location coordinates, and calcaneal eversion angle, and then performs queries, matching, and interpolation calculations in a parametric design rule base to drive a parametric insole base model, automatically calculating dozens of key geometric and physical parameters for various parts of the insole.

[0034] Refined Adjustments: For example, for arch support, the engine generates a saddle-shaped support structure that conforms to the user's arch curve based on the calculated height and curvature; for forefoot cushioning, it achieves a gradient distribution of cushioning performance by changing the internal lattice structure or specifying printing materials of different hardness below the metatarsal areas where pressure is too high, based on the pressure distribution map; for heel correction, it precisely calculates the depth of the heel cup and the inclination angle of the medial or lateral wall to generate a corrective torque that guides the calcaneus back to a neutral position. An optimization loop based on biomechanical simulation feedback is introduced, which combines the generated insole model with the user's foot model in a virtual environment. Finite element analysis is used to simulate the mechanical interactions under walking and other movements. Simulation results include whether the plantar pressure is uniform and whether joint loads are improved. The design parameters are automatically fine-tuned based on the simulation results, and the design with optimal biomechanical performance is obtained after several iterations.

[0035] Human-machine collaboration mechanism: For cases with low AI confidence or very complex scenarios, the proposed solution will be pushed to the doctor's review system. Doctors can use interactive tools to fine-tune the solution on an interface that compares the original data, AI report, and 3D model side-by-side. These fine-tuning operations are recorded by the system and used as high-quality labeled data, which is then fed back to the AI ​​design model for reinforcement learning, thereby enabling the continuous evolution of AI design capabilities.

[0036] The final output digital insole solution is a complete 3D mesh model, with complete geometric shape, internal structure, and accompanying material property information, which can be directly used to drive additive manufacturing equipment. The intelligent production and quality control steps for the insole are as follows: The digital insole solution is transmitted to the additive manufacturing equipment for production, and a real-time visual inspection and feedback control mechanism is introduced during the production process to ensure that the product meets design standards. The additive manufacturing equipment is preferably an industrial-grade photopolymerization 3D printer. Its working principle is as follows: A light-transmitting window is located at the bottom of a tank containing liquid photosensitive resin. An ultraviolet laser beam or a digital light processor projector selectively irradiates the liquid surface according to the 2D cross-sectional contour of the insole model, causing the resin in a specific area to instantly solidify and adhere to the printing platform. The platform then rises, new resin flows and levels, and the next layer is irradiated and cured. This process is repeated layer by layer until a complete solid insole is formed. This method is particularly suitable for manufacturing insoles with complex curved surfaces and customized internal structures.

[0037] Intelligent manufacturing: The AI ​​system automatically selects the most suitable medical-grade polymer material, such as photosensitive resin with a specific Shore hardness, from the material library based on the mechanical properties sought by the insole design, such as the rigidity of the support parts, the elasticity of the cushioning parts, and the biocompatibility requirements. At the same time, it automatically generates the optimal support structure based on the geometric features of the model, such as the overhang structure and thin-walled areas, and sets core process parameters such as layer thickness, exposure time, and printing speed.

[0038] Real-time visual inspection and feedback control mechanism: This is a digital quality inspector embedded in the production process. High-resolution industrial cameras are installed inside the printing chamber to monitor the printing process layer by layer or in real time. Computer vision algorithms, such as image segmentation and template matching, compare the actual image of each layer with the expected CAD slice image at the pixel level, thereby instantly identifying common defects such as interlayer peeling, material shortage, holes, and shape deformation. Once a defect is identified, the control system can immediately trigger an automatic pause and recommend or directly execute corrective strategies based on the type and location of the defect. For example, cleaning the material tank to ensure a flat liquid surface, fine-tuning the laser power to compensate for insufficient curing, or regenerating the local support structure. This online, on-machine inspection-feedback-control closed loop greatly improves the first-piece success rate and the pass rate of mass production.

[0039] Wearable Feedback and System Adaptive Iteration Steps: Smart insoles with built-in sensors collect actual user wear data. Combined with user feedback, this data is used to continuously optimize the AI ​​analysis model and design rule base, achieving closed-loop iteration and continuous improvement in personalized adaptation capabilities. Smart Insoles with Built-in Sensors: This is an IoT terminal embedded in the final product. The insole integrates a micro-pressure sensor array for continuous monitoring of pressure in various areas of the sole and an inertial measurement unit (IMU) to capture the insole's acceleration and angular velocity in three-dimensional space, thereby calculating gait symmetry, cadence, activity intensity, etc.

[0040] Data Collection and Fusion Analysis: The collected data is transmitted to the user's smartphone app via Bluetooth Low Energy, and then uploaded to the cloud platform after encryption. In the cloud, the AI ​​system performs a crucial comparative analysis: comparing the user's long-term dynamic gait data and pressure distribution after wearing the insole with baseline data before wearing it, thereby quantitatively evaluating the corrective effect of the insole, such as whether the peak plantar pressure has decreased, whether the pressure distribution is more balanced, and whether the gait symmetry index has improved. Simultaneously, the app guides users to provide subjective feedback on comfort, pressure sensations in specific areas, and fatigue through structured questionnaires or natural language input interfaces.

[0041] System Adaptive Iteration: Based on collected objective effect data and subjective comfort information, the system dynamically adjusts feature weights in the AI ​​analysis model through online learning algorithms or periodic batch retraining. For example, if a certain pressure feature is found to be crucial for predicting the correction effect, its weight is increased, and thresholds and rules in the parametric design rule base are revised. For instance, based on extensive feedback data, the formula for calculating the support height of a certain type of foot arch is optimized for greater accuracy. This allows the system to generate more precise follow-up or repeat purchase products for the user, while also providing increasingly optimized initial solutions for potential user groups with similar characteristics, thus achieving a spiral increase in the system's personalized adaptation capabilities.

[0042] The integration and standardization of AI foot data processing and professional analysis, as well as the evaluation of AI analysis models, specifically include: spatiotemporal alignment and registration of heterogeneous data from different acquisition devices to ensure that static morphology and dynamic mechanical data are correlated in a unified coordinate system. Specifically, this involves accurately mapping the spatial coordinates of the plantar pressure distribution heatmap to the plantar surface of the 3D foot model, and synchronizing joint angle data at each moment within the gait cycle with the three-dimensional posture of the foot; this is the core technology for achieving effective multimodal data fusion. First, a predefined calibration process is used to establish a transformation matrix between the pressure plate pixel coordinate system and the 3D scanner's world coordinate system. After user data acquisition, this matrix is ​​used to project the position of each pixel on the pressure heatmap onto the corresponding plantar surface of the 3D foot model, thus binding the pressure value to the three-dimensional position. For gait data, the heel strike events recorded by the gait capture system are aligned with the corresponding postures in the 3D dynamic sequence on the time axis to ensure that the corresponding and accurate three-dimensional shape of the foot and joint angles can be obtained at any percentage point of the gait cycle, thereby constructing a spatiotemporally unified and complete digital model of foot movement and force.

[0043] High-dimensional feature extraction is performed on the standardized multimodal data stream. Specifically, a convolutional neural network branch is used to extract spatial morphological features from 3D foot point cloud data and pressure distribution images, including arch curvature, metatarsal head distribution, and the morphology of the pressure center trajectory. Simultaneously, a temporal processing network branch is used to model gait sequence data, capturing dynamic patterns and abnormalities within the gait cycle. This step is crucial for automatically learning complex foot features using deep learning models. For 3D point cloud data, sparse convolutional networks efficiently handle its disorder and sparsity, automatically learning hierarchical geometric features from macroscopic arch shape to microscopic metatarsal prominences. For pressure images, standard 2D convolutional networks can identify specific patterns in pressure distribution, such as the morphological features of the pressure center trajectory. For temporal data like gait sequences, temporal processing networks such as the Transformer, with their self-attention mechanism, can analyze the long-range dependencies between motion features at different moments in the gait cycle, thereby accurately capturing transient but diagnostically significant abnormal dynamic patterns, such as excessive internal pronation at the beginning of the gait.

[0044] The extracted spatial and temporal features are fused and input into a classification and regression model within a multi-task learning framework. This model performs the following tasks: multi-classification identification of common foot problems such as flat feet, high arches, and pronation / valgus deformities, and outputs confidence scores; continuous numerical regression prediction of the severity of identified problems, such as the angle of arch collapse and the degree of calcaneal eversion; and inferring the future trajectory and risk probability of foot development based on users' historical data through a temporal prediction model. The multi-task learning framework allows models to share underlying features and solve multiple related tasks simultaneously, which not only improves data utilization efficiency but also enhances the model's generalization ability due to the correlation constraints between tasks. The classification task outputs a probability distribution, with the category corresponding to the highest probability being the identified foot problem, and the confidence score reflecting the model's degree of certainty in its judgment. The regression task outputs specific, continuous numerical values, providing precise quantitative input for customized design, such as the required correction degree. The temporal prediction task requires building specialized prediction models such as ARIMA, LSTM, or temporal Transformer to analyze the time series of the user's past test data, fit its developmental curve, and thus make probabilistic predictions about the foot's condition at a specific future time point, enabling proactive intervention. Finally, all analysis results are integrated to generate a structured foot health analysis report containing quantitative parameters, qualitative conclusions, risk warnings, and visualizations. This report is a concentrated embodiment and output of AI analysis capabilities. Its structure is reflected in its composition of predefined, machine-resolvable fields and blocks, such as: {"diagnosis": "flatfoot", "severity_angle": 12.5, "risk_level": "medium", "visualization_url": "..."}. This structure facilitates the direct use of key parameters in subsequent AI design steps. Visualizations are generated using computer graphics technology, such as rendering pressure heatmaps semi-transparently on a 3D foot model, or plotting predicted developmental trajectories as trend lines on a chart, making the analysis results clear at a glance and easy for doctors to quickly understand and confirm.

[0045] The training and optimization process of the AI ​​analysis model specifically includes: constructing a large database of children and adolescents' feet, covering different age groups, genders, regions, and foot health statuses from 6 to 18 years old. This database has a sample size of no less than 100,000 cases and is continuously updated. It includes foot disease types and severity rating labels annotated by professional physicians, as well as longitudinal tracking data from some users. A large-scale, high-quality, and diverse labeled database is the cornerstone of the success of deep learning models. Covering different ages, genders, and regions ensures the model's universality and avoids inaccurate identification of specific groups due to training data bias. The inclusion of a large number of healthy and various pathological samples enables the model to learn the difference between normal variations and true pathological features. Longitudinal tracking data, i.e., multiple test data of the same user at different time points, is crucial for training the developmental trend prediction model, enabling it to learn the dynamic patterns of individual foot development.

[0046] An improved CNN-Transformer hybrid architecture is adopted as the core model. The CNN part uses a sparse convolutional network to process 3D point clouds to efficiently extract macroscopic and microscopic morphological features of the foot. The Transformer part utilizes its self-attention mechanism to model long-range dependencies in gait sequences and accurately capture gait anomalies. Traditional CNNs perform well in processing regular images, but are inefficient in directly processing disordered and unstructured 3D point clouds. Sparse convolution is specifically designed for 3D point clouds, calculating only regions of the point cloud, which greatly improves processing efficiency and memory utilization. Gait sequences are essentially time series, and actions at different times in the sequence may influence each other, such as the starting posture affecting the push-off phase. The Transformer's self-attention mechanism can calculate the relationship weights between any two time points in the sequence, thereby better understanding the contextual information of the entire gait cycle. Compared with traditional RNNs or LSTMs, it is better at capturing long-range dependencies and is more effective in identifying complex, cross-gait phase anomaly patterns.

[0047] In model training, an adversarial training strategy is introduced. By adding imperceptible noise perturbations to the input data, the robustness of the model to real-world interference such as differences in data acquisition devices and slight changes in user posture is improved, ensuring the stability of the analysis results. Adversarial training is a regularization technique designed to enhance the model's immunity to minor changes in the input data. During training, adversarial examples that are very similar to the original training data but with added subtle perturbations are intentionally generated, and the model is required to make correct predictions on these examples. In this way, the model is forced to learn more essential and robust features, rather than over-relying on certain noises or irrelevant details in the training data. This allows the final deployed model to maintain stable analytical performance even when faced with uncertainties in the real world, such as data collected from different brands of devices and users' non-perfect standing postures. Knowledge distillation is employed to extract model knowledge, which excels in morphological and gait analysis, into a lightweight, unified model. This reduces the model's computational complexity while maintaining analytical accuracy, enabling it to be deployed on edge computing devices or achieve rapid response on mobile devices. Knowledge distillation is a model compression technique. Typically, we first train multiple large, complex, and high-performance teacher models, such as an expert model specifically designed for 3D morphology using CNNs and another expert model specifically designed for gait using Transformers. Then, we build a smaller, simpler student model. During the distillation phase, instead of simply using real labels to train the student model, it's more important to have the student model mimic the outputs of the teacher models, including the final predictions and the feature representations of the intermediate layers. In this way, the student model can learn from the powerful teacher models, achieving near-teacher model performance with a significantly reduced number of parameters and computational cost, thus meeting the needs of real-time operation on edge devices or mobile apps with limited computing resources.

[0048] The AI-powered foot data processing and professional analysis process also includes a data quality assessment and enhancement sub-step: Before inputting data into the model, a pre-trained quality assessment network scores the integrity and reliability of the data, identifying low-quality or invalid data segments caused by user non-cooperation, device obstruction, or signal interference, and automatically triggering re-collection or providing a prompt. This quality assessment network itself is a small binary classification or regression model trained to determine whether input data, such as a 3D point cloud or a pressure curve, is of acceptable quality. Its input can be statistical features of the data, such as point cloud density, pressure signal signal-to-noise ratio, or features directly extracted by a small CNN, with the output being a quality score. When the score is below a threshold, the system will not perform unreliable analysis but will instead prompt the operator through the terminal interface that the data quality is poor, requesting the user to maintain a stable posture and rescan, thus controlling input quality from the source.

[0049] For cases where data is missing but not severe, Generative Adversarial Networks (GANs) or Variational Autoencoders (VAEs) can be used to generate reasonable supplementary data that matches the individual user's characteristics based on existing high-quality data samples, ensuring the completeness of subsequent analysis. GANs or VAEs are powerful generative models. A GAN consists of a generator and a discriminator. The generator learns to generate realistic data from random noise, while the discriminator learns to distinguish between real and generated data; the two learn to compete and improve together. A VAE learns to encode input data into a latent space and then decodes and reconstructs the data from that space. When user data has small missing patches, such as point cloud holes caused by reflections during scanning, these generative models, trained on large amounts of complete data, can be used to generate visually reasonable and anatomically plausible data to fill in the missing areas based on the contextual information around them, thus providing a complete input for downstream AI analysis tasks.

[0050] To address the challenges of collecting pediatric foot data, advanced data augmentation techniques are employed, particularly in the feature space. These include altering the rotation angle of the foot model and applying simulated mechanical loads to expand the diversity of training data without altering the underlying pathological characteristics. Data augmentation is an effective means of addressing insufficient training data and improving model generalization ability. Beyond simple rotations, translations, and scaling in the original data space, feature space data augmentation is more sophisticated. It involves manipulating these feature vectors after the data has been mapped to high-dimensional feature vectors by the encoder, a part of the network. For example, slight linear interpolation or the addition of random noise to the feature vectors of the same foot can create new samples that are similar but different at the feature level. Furthermore, based on biomechanical knowledge, simulations can be performed to apply forces of different directions and intensities to the foot model and calculate minute changes in its morphology, generating a series of foot morphology data under different loads. This significantly increases the diversity and coverage of the training data, making the model more robust.

[0051] The AI-powered intelligent generation process for customized insoles involves precise adjustments to key areas, specifically: Based on the foot health analysis report, the AI ​​calculation engine first accurately calculates the apex position of the arch support pad, the radius of curvature of the support surface, and the gradient distribution of support intensity according to the arch type and height. This ensures sufficient support while avoiding excessive pressure on the plantar fascia. Arch support design is the core of the insole's function. The apex position is typically located below the navicular tuberosity, requiring precise calculation of its three-dimensional coordinates based on a 3D foot model. The radius of curvature of the support surface must match the curvature of the user's arch, forming a snug, saddle-shaped contact surface to distribute pressure. The gradient distribution of support intensity refers to the gradual transition in support rigidity from the apex outwards. This can be achieved, for example, by changing the internal filling density or lattice structure of the insole in this area, thereby avoiding noticeable steps or pressure points at the support edges and improving comfort.

[0052] To address the abnormal pressure areas revealed in the plantar pressure analysis report, differentiated design of insole material properties is employed: Below the metatarsal areas with excessive pressure, a cushioning structure with specific shock absorption properties and crumple zone is designed to disperse impact force; in areas with insufficient pressure or requiring stimulation, the material hardness is adjusted or the contact area is increased to improve sensation and function; this represents proactive intervention in the plantar biomechanics environment. For high-pressure areas such as the second and third metatarsal heads, the cushioning structure can be achieved by designing a hollow lattice with a specific shape and wall thickness inside the insole. This structure can undergo controlled elastic deformation and crumple under pressure, absorbing impact energy. Its cushioning performance and crumple zone can be precisely controlled by the type, size, and rod diameter of the lattice units. For low-pressure areas or areas requiring stimulation, such as the medial arch where enhanced sensation is needed in certain corrective strategies, relatively harder materials or denser support structures can be used to reduce crumple, increase contact feedback, and thus guide the foot to produce the desired biomechanical response.

[0053] Based on the analysis of calcaneal inversion / valgus angles and stability during heel strike in dynamic gait, the depth, medial or lateral wall inclination angle, and encirclement height of the heel cup are meticulously designed to provide precise calcaneal encirclement and corrective torque, guiding the foot and ankle back to a neutral position. The heel cup is a key component for correcting foot-ankle alignment. For foot eversion, a heel cup with a higher medial wall and a greater inward inclination angle needs to be designed. When the calcaneus falls into it, the cup wall applies an inward pushing force to the lateral side of the calcaneus, generating an inversion corrective torque to counteract its eversion tendency. Conversely, the opposite is also true. The depth and encirclement height together determine the range of calcaneal encirclement and guidance, which needs to be calculated based on the size of the calcaneus and the severity of eversion to ensure that the corrective force is effective without unduly restricting normal activity.

[0054] The parametric modeling engine parameterizes all the above adjustments and automatically generates a complete, watertight 3D insole mesh model with optimal printing orientation. This model comprehensively considers the constraints of the printing process and supports export in common formats such as STL and 3MF. Parametric modeling engines, such as Python-based CadQuery or OpenSCAD scripts, define the overall shape of the insole and the adjustments of each functional area as a series of parameters and rules. The parameter set output by the AI ​​calculation engine drives the script to automatically generate the final 3D geometry. Watertightness means that the generated mesh model is a completely closed solid model without any broken surfaces, which is a basic requirement of 3D printing slicing software. The optimal printing orientation is automatically calculated by an algorithm, typically aiming to maximize the quality of the insole's functional surface (the foot contact surface), minimize the required support structure, and minimize the total printing time. STL and 3MF are standard file formats in the additive manufacturing field, ensuring compatibility with various 3D printers and slicing software.

[0055] The AI-powered intelligent generation process also includes an optimization loop based on biomechanical simulation feedback: after generating a preliminary 3D insole model, it is registered with a standardized foot skeletal muscle model and collected user foot morphology data. In an integrated biomechanical simulation environment, the interaction between the foot and insole during typical movements such as walking and running is simulated; this is an advanced stage that introduces engineering simulation into the design closed loop. First, the user's personalized foot skin surface model needs to be registered with a general, standardized skeletal muscle model containing bones, ligaments, and muscles to obtain a personalized biomechanical model. Then, the designed insole model is placed under this foot model, and the contact properties between them—friction and stiffness—are defined. In the simulation environment, muscle driving forces and ground reaction forces representing walking or running are applied. Through coupled simulation using multibody dynamics and finite element analysis, the dynamic responses of the foot, insole, and lower limbs during the entire movement process are calculated.

[0056] Finite element analysis (FEA) is used to calculate the stress-strain distribution of the insole under load, the pressure changes in specific areas of the sole, and the impact of corrective forces on joint moments. FEA is the core of the simulation. It discretizes the insole and foot soft tissue into countless tiny units. By solving complex physical equations, it can visually output the stress-strain distribution inside the insole, assessing its structural strength and identifying stress concentration areas. Simultaneously, it can output the pressure distribution at the contact surface between the sole and the insole, intuitively showing whether the design effectively reduces peak pressure in high-pressure areas. Furthermore, by analyzing joint moments at the ankle and subtalar joints, the corrective effect of the insole can be quantitatively evaluated, for example, observing whether the moment causing pronation is reduced.

[0057] The AI ​​engine automatically adjusts the design parameters of the insole based on the difference between the simulation results and the ideal biomechanical target, fine-tuning the shape of the support area and changing the distribution density of the cushioning material; This step enables automated simulation-driven design. An AI engine or optimization algorithm, such as gradient descent or genetic algorithms, compares simulation results (e.g., maximum pressure, joint torque) with preset ideal targets (e.g., uniform pressure distribution, minimum joint torque) and calculates the loss value of the objective function. Then, through backpropagation or heuristic search, it automatically and incrementally adjusts the insole design parameters (e.g., increasing arch support height by 0.5mm, reducing forefoot cushioning material hardness by 10%), and initiates a new simulation. This process is repeated.

[0058] After several iterative cycles of simulation, evaluation, and adjustment, a personalized insole design optimized for biomechanical performance is finally output, ensuring not only a proper fit but also effective improvement of the user's foot health. After several such iterative cycles, the design parameters converge to a solution that optimizes the biomechanical objective function. The final design, validated through virtual clinical trials, theoretically outperforms the initial design in improving the plantar biomechanical environment and correcting abnormal joint torques. This significantly enhances the scientific rigor and effectiveness of customized insoles, elevating design from empirical observation to precise, simulation-verified design.

[0059] In the AI-powered generation of customized insole solutions, a human-machine collaborative review mechanism is introduced for complex or severe foot problems. The initial design plan generated by AI is automatically pushed to a professional review interface on the end of a qualified physician's or rehabilitation therapist's device. This interface displays raw foot data, AI analysis reports, a 3D model of the insole, and key biomechanical simulation results side-by-side. This review interface is a highly integrated decision support workbench. Its design aims to allow experts to quickly and comprehensively understand the AI's decision-making process. Through side-by-side display, experts can easily perform cross-validation: for example, comparing the original 3D scans and pressure data to determine the accuracy of the foot problem identified by the AI; observing the 3D model of the insole to assess whether its design conforms to clinical logic; and referring to the biomechanical simulation results to determine whether its expected corrective effect is ideal. This transparent presentation builds expert trust in the AI ​​and provides sufficient contextual information for fine-tuning.

[0060] The physician-side system provides a suite of professional, interactive fine-tuning tools, allowing physicians to make localized modifications to AI-generated insole designs based on their clinical experience. These tools include manually dragging control points to adjust the height and shape of the arch support curve, or adjusting the suggested hardness value for specific areas using sliders. These tools are specifically designed for insoles. They are intuitive and directly linked to the insole's underlying parametric model. When a physician drags a control point to change the arch curve, the system is actually modifying the control parameters that generated that curve. When the hardness slider is moved, the corresponding value in the material allocation table is being modified. All these modifications are non-destructive and parametric, meaning physicians can undo changes or make fine-tuning adjustments at any time without having to start modeling from scratch. This significantly improves the efficiency of specialists, allowing them to focus on core clinical decisions.

[0061] All fine-tuning operations are recorded by the system and used as new labeled data, flowing back into the training database of the AI ​​design model for reinforcement learning. This allows the AI ​​model to continuously learn and align with expert decision-making logic, gradually reducing future reliance on human intervention and enabling the continuous evolution of AI design capabilities. This is the key data flow for achieving system self-evolution. Each manual fine-tuning is essentially a demonstration: the system records how, under certain foot data input and the initial AI plan, the expert ultimately adjusts parameter A to value A'. A massive amount of such (state, action, new state) data pairs constitute a perfect dataset that can be used to train the AI ​​design model through reinforcement learning or imitation learning. In this way, the AI ​​model gradually learns in what situations and how to adjust the plan to better align with expert preferences and clinical standards, thereby continuously improving its design capabilities. In the future, when encountering similar cases, it may directly generate high-quality plans requiring no or minimal modification.

[0062] The intelligent production and quality control steps for insoles are as follows: Utilizing industrial-grade photopolymer 3D printing technology, the AI ​​system automatically selects the most suitable medical-grade polymer material and optimizes printing parameters based on the mechanical properties and comfort requirements of the final insole design. Material selection is a multi-objective decision-making process. The AI ​​system maintains a material library where each material is labeled with its key attributes, such as Shore hardness, tensile strength, elongation at break, and biocompatibility rating. Based on the functional requirements of different areas in the design (e.g., high rigidity for the arch and high elasticity for the forefoot), the system uses an optimization algorithm to select one or more materials from the library to achieve the best combination of properties that meets the overall performance goals in multi-material printing. Simultaneously, the optimization of printing parameters such as layer thickness, exposure time, bottom layer exposure time, and lift-off speed is based on extensive historical printing data and physical models to achieve a balance between printing accuracy, speed, and success rate.

[0063] During the printing process, high-resolution industrial cameras provide layer-by-layer or real-time image monitoring. Computer vision algorithms compare the printed layers with the design slices, instantly identifying and locating defects such as interlayer delamination, material shortages, and deformation. The industrial camera is fixed in a position that clearly captures the printing platform. After each layer is printed and before the next layer begins, the camera takes a top-view image of that layer. The computer vision algorithm first preprocesses the image, such as grayscale conversion, filtering, and binarization. Then, using image differencing or template matching techniques, it compares the captured actual printed layer contour with the ideal layer contour image generated by the slicing software. Any significant differences, such as excess resin residue, missing contour portions, or abnormal warping, are identified as potential defects, and their pixel coordinates are marked.

[0064] Once a defect is detected, the system automatically pauses printing and recommends or directly executes parameter correction strategies based on the defect type, such as cleaning the feed trough, adjusting laser power, or regenerating the support structure, ensuring consistency between the final product and the digital model. This achieves a closed loop from detection to control. The system has a built-in defect-countermeasure knowledge base. For example, if insufficient material is detected, it may recommend increasing the exposure time or pausing printing to clean any solidified residue at the bottom of the feed trough; if interlayer delamination is detected, it may recommend increasing the exposure time of the bottom layer to enhance adhesion or inspecting and regenerating the support structure. For simple parameter adjustments, the system can execute automatically and continue printing; for situations requiring manual intervention, it pauses and issues a clear prompt. This minimizes material and time waste and ensures a high degree of consistency between the final product and the digital solution.

[0065] The wearable feedback and system adaptive iteration steps are as follows: The smart insole incorporates a micro-pressure sensor array and an inertial measurement unit (IMU) to continuously or periodically collect data on foot pressure distribution, gait symmetry, and activity intensity during daily wear; the smart insole serves as the mobile data acquisition terminal in the system. The micro-pressure sensor array is typically fabricated using a flexible FPC circuit board, integrating dozens to hundreds of pressure sensing units distributed across key foot areas. The IMU includes a three-axis accelerometer and a three-axis gyroscope, capable of measuring the insole's motion in three-dimensional space. Through sensor fusion algorithms, gait cycle time parameters, gait symmetry indices such as the left-right stride time difference, and estimated activity intensity such as daily steps and medium-to-high intensity activity time can be extracted from the IMU data.

[0066] This data is transmitted to the user's mobile app via Bluetooth Low Energy (BLE) and then uploaded to the cloud platform after encryption. BLE technology ensures that the insoles have a sufficiently long battery life during charging intervals. The mobile app acts as a data relay station, responsible for receiving, temporarily storing, and preprocessing data, and then securely uploading it to the cloud server via encryption protocols such as HTTPS. The entire data transmission chain is encrypted to ensure that the user's privacy and health information are not leaked.

[0067] The AI ​​system quantifies the corrective effect and fit of the insoles by comparing changes in foot biomechanical data before and after wearing them. Simultaneously, the app guides users to provide subjective feedback such as comfort and pain points. This effect evaluation serves as the basis for iterative optimization. The cloud platform AI system retrieves baseline data from the user's initial testing, such as static arch height and dynamic pressure center trajectory, and compares it with average data collected over a long period after wearing the insoles. For example, it calculates the percentage decrease in average peak pressure or the degree of improvement in gait symmetry index, thereby generating a quantitative effect evaluation report. Subjective feedback is guided through a visual interface within the app, allowing users to click on uncomfortable areas on a foot diagram or rate overall comfort using a slider. The combination of subjective and objective data provides a more comprehensive picture of the effect.

[0068] The system utilizes this objective and subjective feedback data to dynamically adjust the feature weights in the AI ​​analysis model and the thresholds and rules in the parametric design rule base through online learning or periodic batch retraining. This enables the system to generate more accurate subsequent products for the user or provide better initial solutions for similar user groups. This is the ultimate manifestation of closed-loop learning. For online learning, the system can fine-tune in real time. For example, if a large number of users report strong pressure in a certain area of ​​the insole, the system can immediately and appropriately lower the default value of the support hardness in that area in the design rule base. For batch retraining, the system regularly collects all new feedback data, such as quarterly, to form an incremental training set. Then, it comprehensively retrains and optimizes the AI ​​analysis model in claims 2 and 3 and the parametric design rules in claim 5. This process allows the entire system to continuously learn from real-world usage effects, becoming increasingly intelligent, and its output analysis reports and design solutions become increasingly accurate and personalized.

[0069] The end-to-end approach constructs a user-centric digital health record that comprehensively records all information from initial data collection, subsequent AI analysis reports, insole design, production quality data, to long-term wear feedback, forming an immutable and traceable data chain. The digital health record serves as the data bus and platform throughout all steps. It is typically built on a distributed database or data lake technology, creating a unique profile ID for each user. This record adopts a timeline structure, permanently recording end-to-end data for every interaction: original collected data, AI-generated health reports, 3D insole design files, process parameters and quality inspection records during production, and continuous wear feedback data. Its immutability can be achieved through blockchain technology or strict database access control and log auditing. This complete traceability has immense value: for users, it is a complete digital history of their foot health; for doctors, it is a reliable basis for assessing long-term efficacy; for developers, it is a valuable asset for optimizing algorithms and rules; for quality control, it is a traceability tool for locating and resolving problems. This continuously growing record is the core and most enduring value created by this approach.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the article or apparatus that includes that element.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. The present invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-driven end-to-end intelligent adaptation method for children's customized insoles, characterized in that, Includes the following steps: Foot multimodal data acquisition steps: Static three-dimensional geometric morphology data and dynamic biomechanical data of the user's foot are collected simultaneously through non-contact 3D scanning equipment, plantar pressure distribution sensing equipment and gait dynamic capture equipment. AI foot data processing and professional analysis steps: The collected multimodal data is fused and standardized, and then input into an AI analysis model pre-trained with a large-scale foot database to perform in-depth foot health status assessment, pathological feature identification and development trend prediction, and generate a structured foot health analysis report. Customized insole solution AI intelligent generation steps: Based on the structured foot health analysis report, call the parametric design rule library, automatically generate a personalized 3D model of the insole through the AI ​​computing engine, and finely adjust the geometric shape and physical properties of key areas such as support, cushioning, and correction, and output a digital insole solution that can be directly produced; Intelligent production and quality control steps for insoles: The digital insole solution is transmitted to additive manufacturing equipment for production, and a real-time visual inspection and feedback control mechanism is introduced during the production process to ensure that the product meets the design standards; Wearable feedback and system adaptive iteration steps: The smart insole with built-in sensors collects actual wear data from users, and combines it with subjective user feedback to continuously optimize the AI ​​analysis model and design rule base, so as to realize the closed-loop iteration of the system and the continuous improvement of personalized adaptation capabilities.

2. The method according to claim 1, characterized in that, The integration and standardization of AI foot data processing and professional analysis steps, as well as the evaluation of the AI ​​analysis model, specifically include: Spatiotemporal alignment and registration of heterogeneous data from different acquisition devices are performed to ensure that static morphology and dynamic mechanical data are associated in a unified coordinate system. Specifically, the spatial coordinates of the plantar pressure distribution heatmap are accurately mapped to the plantar surface of the 3D foot model, and the joint angle data at each moment in the gait cycle are synchronized with the three-dimensional posture of the foot. High-dimensional feature extraction is performed on the standardized multimodal data stream. Specifically, a convolutional neural network branch is used to extract spatial morphological features from 3D point cloud data of the foot and pressure distribution images, including the arch curvature, metatarsal head distribution, and the shape of the pressure center trajectory. At the same time, a temporal processing network branch is used to model the gait sequence data to capture the dynamic patterns and abnormal patterns in the gait cycle. The extracted spatial and temporal features are fused and input into a classification and regression model under a multi-task learning framework, simultaneously performing the following tasks: multi-classifying common foot problems such as flat feet, high arches, and pronation / vaulting and outputting confidence scores; performing continuous numerical regression prediction on the severity of the identified problems, such as the angle of arch collapse and the degree of calcaneal eversion; and inferring the future trend and risk probability of foot development based on users' historical data through a temporal prediction model. Finally, by integrating all the analysis results, a structured foot health analysis report is generated, which includes quantitative parameters, qualitative conclusions, risk warnings, and visualization charts.

3. The method according to claim 2, characterized in that, The training and optimization process of the AI ​​analysis model specifically includes: To build a large database of foot health data for children and adolescents aged 6 to 18, covering different age groups, genders, regions, and foot health status. The database will have a sample size of no less than 100,000 cases and will be continuously updated. It will include foot disease types and severity rating labels marked by professional physicians, as well as longitudinal tracking data of some users. An improved CNN-Transformer hybrid architecture is adopted as the core model. The CNN part uses a sparse convolutional network to process 3D point clouds to efficiently extract macro and micro morphological features of the foot. The Transformer part uses its self-attention mechanism to model long-range dependencies in gait sequences and accurately capture gait anomalies. In model training, an adversarial training strategy is introduced. By adding imperceptible noise perturbations to the input data, the robustness of the model to real-world interference such as differences in acquisition devices and slight changes in user posture is improved, ensuring the stability of the analysis results. By employing knowledge distillation technology, the model knowledge, which excels in morphological and gait analysis, is extracted into a lightweight unified model. This reduces the computational complexity of the model while ensuring analytical accuracy, enabling it to be deployed on edge computing devices or achieve rapid response on mobile devices.

4. The method according to claim 1, characterized in that, The AI ​​foot data processing and professional analysis steps also include a data quality assessment and enhancement sub-step: Before the data is input into the model, a pre-trained quality assessment network is used to score the integrity and reliability of the data, identify low-quality or invalid data segments caused by user non-cooperation, equipment obstruction or signal interference, and automatically trigger re-collection or provide prompts. For cases where data is missing but not severe, generative adversarial networks or variational autoencoders are used to generate reasonable supplementary data that matches the individual characteristics of the user based on existing high-quality data samples, so as to ensure the completeness of subsequent analysis. To address the challenges of collecting data on children's feet, advanced data augmentation techniques are employed, particularly data augmentation in the feature space. For example, by changing the rotation angle of the foot model and applying simulated mechanical loads, the diversity of training data can be expanded without altering the underlying pathological characteristics.

5. The method according to claim 1, characterized in that, The AI-powered intelligent generation process for the customized insole solution involves fine-tuning key areas, specifically including: Based on the foot health analysis report, the AI ​​computing engine first calculates the apex position of the arch support pad, the radius of curvature of the support surface, and the gradient distribution of the support strength according to the arch type and height, to ensure that while providing sufficient support, it avoids excessive pressure on the plantar fascia. For the abnormal pressure areas revealed in the plantar pressure analysis report, the insole material properties are designed differently: under the metatarsal area where the pressure is too high, a cushioning structure with specific shock absorption performance and crumple zone is designed to disperse the impact force; in areas with insufficient pressure or that need stimulation, the perception and function are improved by adjusting the material hardness or increasing the contact area. Based on the analysis of the calcaneal inversion / valgus angle and the stability of the heel strike phase during dynamic gait, the depth of the heel cup, the inclination angle of the medial or lateral wall, and the encirclement height are refined to provide precise calcaneal wrapping and corrective torque, guiding the foot and ankle back to a neutral position. The parametric modeling engine parameterizes all the above adjustments and automatically generates a complete, watertight, 3D insole mesh model with optimal printing orientation. This model takes into account the constraints of the printing process and supports export in STL and 3MF universal formats.

6. The method according to claim 5, characterized in that, The AI ​​intelligent generation step also includes an optimization loop based on biomechanical simulation feedback: After generating a preliminary 3D model of the insole, it is registered with a standardized foot skeletal muscle model and collected user foot morphology data. In an integrated biomechanical simulation environment, the interaction between the foot and the insole is simulated under typical movement states such as walking and running. The stress-strain distribution of the insole under load, the pressure change in a specific area of ​​the sole, and the effect of the corrective force on the joint torque were calculated by finite element analysis. The AI ​​engine automatically adjusts the design parameters of the insole based on the difference between the simulation results and the ideal biomechanical target, fine-tuning the shape of the support area and changing the distribution density of the cushioning material; After several iterations of simulation, evaluation, and adjustment, a personalized insole design scheme optimized in terms of biomechanical performance is finally output, ensuring that it not only fits the shape but also effectively improves the user's foot health.

7. The method according to claim 1, characterized in that, In the AI-powered intelligent generation step of the customized insole solution, a human-machine collaborative review mechanism is introduced for complex or severe foot problem cases: The preliminary design scheme generated by AI will be automatically pushed to the professional review interface of qualified physicians or rehabilitation therapists. This interface displays the original foot data, AI analysis report, insole 3D model and key biomechanical simulation results side by side. The physician-side system provides a set of professional interactive fine-tuning tools, allowing physicians to make local modifications to the AI-generated insole solutions based on their clinical experience, such as manually dragging control points to adjust the height and shape of the arch support curve, or adjusting the suggested hardness value of specific areas through sliders. All fine-tuning operations will be recorded by the system and used as new labeled data, which will be fed back into the training database of the AI ​​design model for reinforcement learning. This will enable the AI ​​model to continuously learn and align with the decision-making logic of experts, gradually reducing its reliance on human intervention in the future and achieving continuous evolution of AI design capabilities.

8. The method according to claim 1, characterized in that, The specific steps of the intelligent production and quality control of the insoles are as follows: using industrial-grade photopolymer 3D printing technology, the AI ​​system automatically selects the most suitable medical-grade polymer material and optimizes the printing parameters according to the mechanical properties and comfort required by the final design of the insoles. During the printing process, high-resolution industrial cameras are used for layer-by-layer or real-time image monitoring. Computer vision algorithms are used to compare the printed layers with the design slices to instantly identify and locate defects such as interlayer peeling, material shortage, and deformation. Once a defect is detected, the system can automatically pause printing and recommend or directly execute parameter correction strategies based on the defect type, such as cleaning the feed trough, adjusting the laser power, or regenerating the support structure to ensure the consistency between the physical product and the digital model.

9. The method according to claim 1, characterized in that, The wearable feedback and system adaptive iteration steps are as follows: The smart insole has a built-in micro-pressure sensor array and inertial measurement unit, which can continuously or periodically collect data on the distribution of foot pressure, gait symmetry, and activity intensity during daily wear. This data is transmitted to the user's mobile terminal APP via Bluetooth Low Energy and uploaded to the cloud platform after encryption. The AI ​​system quantitatively evaluates the corrective effect and fit of the insole by comparing changes in foot biomechanical data before and after wearing. At the same time, the APP guides the user to provide subjective feedback such as comfort and pain points. The system uses these objective and subjective feedback data to dynamically adjust the feature weights in the AI ​​analysis model and the thresholds and rules in the parameterized design rule base through online learning or periodic batch retraining, so that the system can generate more accurate subsequent products for the user or provide better initial solutions for similar user groups.

10. The method according to claim 1, characterized in that, The described end-to-end approach constructs a user-centric digital health record that comprehensively records all information from initial data collection, subsequent AI analysis reports, insole design schemes, production quality data to long-term wear feedback, forming an immutable and traceable data chain.

Images