Food 3D printing control system and printing method based on artificial intelligence

Abstract

The application discloses a kind of food 3D printing control system and printing method based on artificial intelligence, belong to food 3D printing technical field.The system includes AI instruction generation module, dynamic precision prediction module and real-time error correction and compensation module.AI instruction generation module is used to parse the three-dimensional model of user input, combines material characteristic database, generates printing instruction set by machine learning algorithm;Dynamic precision prediction module predicts geometric accuracy, interlayer bonding strength and structure deformation risk by fusing material physical property, historical printing data, real-time printing state data;Real-time error correction and compensation module detects defects by real-time monitoring printing state using lightweight model, and realizes millisecond-level dynamic compensation control according to hierarchical strategy library.The application realizes the full-process intelligent closed-loop control of food 3D printing from model analysis to finished product output, significantly improves printing precision, success rate and adaptability, and reduces dependence on human experience.

Description

Technical Field

[0001] This invention relates to the field of food 3D printing technology, specifically to an artificial intelligence-based food 3D printing control system and printing method. Background Technology

[0002] As an emerging additive manufacturing technology, 3D printing has demonstrated unique advantages in the food industry. By driving the layer-by-layer deposition of food ingredients through digital models, it can achieve customized food with complex shapes (such as biomimetic meat texture replication and nutrient gradient distribution structures), precise nutrient control, and zero-waste production, providing a new path for the traditional food industry to transform towards personalization and intelligentization. However, the industrial application of 3D printing in food still faces core technological bottlenecks. Printing parameters for different materials require repeated exploration and adjustment, and there is a lack of technologies such as centralized material characteristic control, dynamic control of the printing process, and multimodal data fusion.

[0003] The rheological properties of food materials (such as viscosity, elastic modulus, and thixotropy) are strongly coupled with printing accuracy. For example, high-viscosity materials (such as fruit and vegetable purees containing dietary fiber) are prone to clogging of the extrusion nozzle, while low-viscosity materials (such as jam) may cause structural collapse due to insufficient interlayer support. Simultaneously, the thermal response characteristics of materials (such as glass transition temperature and crystallization rate) directly affect the ability to maintain shape during the printing process. For instance, chocolate paste is difficult to mold in a fluid state above 30°C, but can quickly solidify to form a stable structure when cooled to below 20°C. While existing technologies can achieve some material control through preset temperature ranges, they lack the ability to dynamically sense and adaptively adjust the real-time state of the material (such as viscosity fluctuations under changes in shear rate).

[0004] Patent document CN118155120A discloses a 3D printing control method and system based on artificial intelligence. This scheme collects data such as temperature and pressure through an edge computing device and uses a convolutional neural network to identify printing obstacles. However, its defect detection logic, which relies on fixed thresholds, is difficult to adapt to the different characteristics of various food products (such as flour-based and protein-based materials), and it does not construct a dynamic mapping model between material rheological parameters and printing parameters, resulting in high process migration costs for new materials.

[0005] Patent document CN118489922A discloses a 3D printing device and method based on pressure sensing intelligent temperature control. This device monitors the material texture parameters through a mechanical sensor and adjusts the barrel temperature in conjunction with a constant temperature circulation controller. However, it only focuses on temperature control compensation for a single material and does not solve the interface compatibility problem when printing multi-material composites (such as interlayer cracking caused by the difference in thermal expansion coefficients between cheese sauce and biscuit base). Furthermore, the cooling rate control accuracy is insufficient in low-temperature scenarios (such as ice cream printing), which can easily lead to coarse ice crystals in the material, affecting the taste.

[0006] Patent document CN119610677A discloses an intelligent control method and system for the 3D printing process based on real-time feedback. The system predicts crack risk through image analysis and thermal stress calculation. However, it only trains the model based on historical image data and does not integrate the material viscoelastic constitutive relationship (such as the Maxwell model). It has a large error in predicting the gravitational deformation of complex structures (such as spiral hollow surfaces) and lacks the ability to dynamically adjust the material ratio during the printing process.

[0007] In summary, existing technologies have not yet achieved multi-dimensional closed-loop control of "material properties - printing parameters - environmental variables", especially in terms of multi-material collaborative printing, real-time defect compensation and adaptability to low-temperature scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide an artificial intelligence-based food 3D printing control system to solve the core problems in existing food 3D printing technologies, such as printing accuracy relying on manual adjustment, insufficient multi-material collaborative control, lag in real-time defect response, and weak model generalization ability.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an artificial intelligence-based food 3D printing control system, comprising: AI instruction generation module: used to parse the food 3D model data input by the user, combine the rheological properties, thermal response curves and multi-material compatibility matrix in the food material 3D printing database, and generate an appropriate printing program instruction set through machine learning model; receive data feedback from dynamic accuracy prediction module and real-time error correction and compensation module for model learning and optimization. Dynamic accuracy prediction module: It integrates data from the food material 3D printing database with multimodal data collected by the real-time error correction and compensation module to predict the geometric accuracy, interlayer bonding strength, and structural deformation risk of the printed product; it feeds back the prediction results to the AI ​​instruction generation module and the real-time error correction and compensation module; and it receives data from the real-time error correction and compensation module to update the prediction model. Real-time error correction and compensation module: includes a multimodal sensor unit, a defect detection and classification submodule, and a dynamic compensation engine. The multimodal sensor unit collects state data in real time during the food 3D printing process. The defect detection and classification submodule uses computer vision models and time series analysis models to detect defects. The dynamic compensation engine generates compensation strategies based on the prediction results fed back by the dynamic accuracy prediction module and the analysis results of the defect detection and classification submodule.

[0010] The core architecture of this invention consists of the three major modules mentioned above, which work together to achieve intelligent control of the entire process of 3D printing in food processing. Specifically, through intelligent fusion of multimodal data, machine learning model-driven operation, and closed-loop control throughout the entire process, it achieves automated high-precision program generation, dynamic accuracy prediction and risk warning, and multi-dimensional real-time error correction control.

[0011] In this invention, the AI ​​instruction generation module is used for intelligent analysis of 3D printing models and generation of 3D printing program instruction sets.

[0012] Furthermore, the AI ​​instruction generation module includes a 3D model processing unit, a printing model generation unit, and an intelligent slicing unit.

[0013] The 3D model processing unit is used to receive food 3D model instructions input by the user, identify the geometric features, structural features and material distribution requirements of the model through feature extraction algorithms, and output feature parameters, including: side length, height, radius, etc.; the geometric features include: curvature, wall thickness, connectivity; the structural features include support requirements, layer complexity.

[0014] In this invention, the 3D model processing unit supports multiple input methods such as voice, text, files, or images. The computer-readable storage medium of the 3D model processing unit stores a feature extraction program.

[0015] As a preferred approach, feature extraction algorithms such as PointNet++, MeshCNN, and MSFFN are used.

[0016] Preferably, the 3D model processing unit stores an edge enhancement preprocessing algorithm. Using this algorithm to reduce noise in the input image can improve feature extraction accuracy to over 98%.

[0017] Preferably, the 3D model processing unit stores a multi-scale feature fusion network extraction algorithm. For complex curved surface models, the multi-scale feature fusion network extracts cross-scale geometric features, ensuring the complete recognition of detailed features such as hollow structures and thin-walled regions.

[0018] The printing model generation unit generates a multi-material structure model adapted for 3D printing based on the parameters output by the 3D model processing unit and combined with the food material 3D printing database. This model includes material zoning planning, support structure design, and printing path optimization.

[0019] The food material 3D printing database stores data such as the rheological properties, thermal response curves, and multi-material compatibility matrix of food materials. Specifically, the rheological properties include consistency coefficient, flow index, shear viscosity, elastic modulus, and yield stress, which are used to determine the printability of inks and optimize the extrusion process. The thermal response curves include phase transition temperature, specific heat capacity, thermal conductivity, and temperature-viscosity curves, which are used to precisely control the printing temperature, manage interlayer bonding, and predict and prevent thermal damage. The multi-material compatibility matrix includes interfacial bonding strength, diffusion coefficient, and coefficient of thermal expansion, which are used to achieve complex multi-material printing, avoid incompatibility, and optimize interface performance.

[0020] The material zoning planning refers to allocating materials based on the model's functional requirements. Based on these requirements (such as taste, nutrition, and structural support) and material properties, the most suitable printing materials are assigned to different geometric regions of the 3D model. The feasibility of interface bonding is verified using a multi-material compatibility matrix in a database to achieve the printing of complex structures and functionally graded materials. For example, a biomimetic meat steak is divided into muscle fiber regions and fat marbling regions, while a filled chocolate is divided into an outer shell region and an inner filling region.

[0021] Preferably, the computer-readable storage medium of the printed model generation unit stores a semantic segmentation network model and a material matching algorithm. The semantic segmentation network is used to divide the 3D model into regions, and the optimal material combination is automatically matched by the material matching algorithm. For example, a protein material with an elastic modulus ≥ 500 kPa is used for the outer shell, and a fiber matrix with a viscosity ≤ 2000 Pa·s is used for the internal filling.

[0022] The aforementioned support structure design refers to generating edible supports through a topology optimization algorithm that generates unit storage from printed models.

[0023] As a preferred option, the support structure adopts a lattice-type honeycomb structure. A topology optimization algorithm based on the variable density method is used to generate the lattice-type honeycomb support structure with the goal of maximizing structural stiffness. While ensuring support strength, the amount of material used is reduced, with the support volume accounting for ≤12%. Low-melting-point edible materials are selected, such as gelatin with a melting point of 35℃ or chocolate with a melting point of 32℃, which facilitates later peeling.

[0024] The print path optimization refers to generating the shortest print path using a reinforcement learning model stored in the print model generation unit. When constructing the Markov Decision Process (MDP) model for the print path optimization problem, print time, material changeover times, and idle distance are used as optimization objectives, and the Proximal Policy Optimization (PPO) algorithm is employed for policy search.

[0025] As a preferred option, an obstacle avoidance constraint mechanism is introduced, and an S-shaped path transition is adopted in the multi-material interface area to reduce the risk of material mixing.

[0026] As a preferred option, the idle speed is optimized by combining the time cost function, which reduces the overall printing time by more than 25%.

[0027] The intelligent slicing unit is used to slice the multi-material structure model generated by the printing model generation unit into layers, generating a G-code instruction set including spatial coordinates, extrusion amount, speed, temperature, infill rate, and printing height.

[0028] Specifically, the computer-readable storage medium of the intelligent slicing unit stores a Transformer encoder-decoder. The Transformer encoder-decoder architecture is used to slice the 3D model into layers, generating a G-code instruction set, where temperature accuracy is controlled within ±0.1℃; and the fill rate is adjustable within a range of 5%-100%.

[0029] Preferably, the intelligent slicing unit stores an adaptive layer thickness algorithm and a gated attention mechanism. For complex regions with curvature > 60°, the adaptive layer thickness algorithm automatically reduces the layer thickness to 0.15 mm, and the attention mechanism enhances parameter control over detailed features.

[0030] As a preferred option, the intelligent slicing unit stores a gradient filling strategy. For the internal filling structure, a gradient filling strategy is adopted to reduce material usage by 15%-30% while ensuring structural strength.

[0031] Preferably, the execution process of the gradient filling strategy includes: Region division steps: Based on the structural features of the three-dimensional model, its interior is divided into a core load-bearing area, a transition area, and an edge non-load-bearing area; Gradient planning steps: Assign fill rates and appropriate fill patterns to the different regions mentioned above, with a gradient decreasing from high to low. Simulation optimization steps: Based on the rheological property data in the 3D printing database of the food materials, physical simulation and iterative optimization are performed on the filling structure to finally generate a filling scheme that reduces the amount of material used while ensuring structural strength.

[0032] Preferably, the AI ​​instruction generation module automatically switches between generating adversarial networks (GANs), graph neural networks (GNNs), or reinforcement learning models to generate instructions based on user needs (such as printing speed priority or accuracy priority). For example, when the user selects printing speed priority, the module calls a reinforcement learning model to optimize printing efficiency; when the user selects printing accuracy priority, the module calls a generative adversarial network or graph neural network model to generate printing instructions with the fewest defects and the highest fidelity.

[0033] The 3D printer executes the 3D printing task by reading the instructions output by the AI ​​instruction generation module.

[0034] In this invention, the prediction results of the dynamic accuracy prediction module are output to the AI ​​instruction generation module and the real-time error correction and compensation module; this is used to guide the AI ​​instruction generation module to iteratively optimize the printing program instruction set and to provide forward-looking warnings to the real-time error correction and compensation module, so as to achieve prediction-based proactive compensation.

[0035] Furthermore, the dynamic accuracy prediction module includes a data preprocessing unit, a multimodal data fusion subunit, a geometric accuracy prediction subunit, an interlayer bonding strength prediction subunit, and a structural deformation risk assessment subunit.

[0036] The data preprocessing unit is used to perform parallel preprocessing operations on multi-source heterogeneous data, including normalization, feature alignment, and noise filtering. The multi-source heterogeneous data includes: material property parameters and historical printing data stored in the food material 3D printing database, as well as image data and sensor data acquired in real time by the multimodal sensor unit.

[0037] The material properties include consistency coefficient, flow index, and yield stress, which are obtained by measuring a rheometer.

[0038] The historical printing data refers to time-series data of millions of historical tasks, including extrusion volume, temperature, and path.

[0039] The image data is acquired in real time by the camera and 3D scanner of the multimodal sensor unit, including the surface texture image of the printed layer and the point cloud data acquired by the 3D scanner.

[0040] The sensor data is collected in real time by the mechanical sensor, infrared thermal imager, and environmental sensor of the multimodal sensor unit, including the extrusion pressure collected by the mechanical sensor with an accuracy controlled within ±0.01 MPa; the interlayer temperature gradient collected by the infrared thermal imager with an accuracy controlled within ±0.5℃; and the ambient humidity collected by the environmental sensor with an accuracy controlled within ±2%RH.

[0041] The preprocessed multimodal data is fed into the multimodal data fusion subunit to generate a temporal-spatial correlation feature tensor. This feature tensor serves as the common input basis for the geometric accuracy prediction subunit, the interlayer bonding strength prediction subunit, and the structural deformation risk assessment subunit, and is used for subsequent accurate prediction and analysis.

[0042] The computer-readable storage medium of the multimodal data fusion subunit stores a spatiotemporal sequence model or a graph structure analysis model. Using the spatiotemporal sequence model or graph structure analysis model, multi-source heterogeneous data is fused. The fusion is achieved through feature-level fusion or decision-level fusion, mapping data from different modalities to a unified feature space to extract complementary information and eliminate redundancy. The purpose of fusion is to establish spatiotemporal correlations between data and form a unified feature representation for subsequent prediction tasks. The output after fusion is a temporal-spatial correlation feature tensor.

[0043] Preferably, the computer-readable storage medium of the multimodal data fusion subunit stores a spatiotemporal graph convolutional network (ST-GCN) model. Multi-source heterogeneous data is fused by constructing a spatiotemporal graph. The dynamic graph structure serves as input, and the spatiotemporal state fusion feature tensor serves as output. The dynamic graph structure includes, but is not limited to, real-time temperature (from thermal imaging), appearance image features (from a camera), stress / vibration data (from mechanical sensors), and material property parameters planned for that location retrieved from a database. This output tensor deeply fuses the joint information of multimodal data in the spatiotemporal dimension, accurately representing the comprehensive state of the printing process at any time and any spatial location. It serves as the common input feature for the geometric accuracy prediction subunit, interlayer bond strength prediction subunit, and structural deformation risk assessment subunit, providing a data foundation for subsequent high-precision predictions.

[0044] Preferably, the multimodal data fusion subunit stores a federated learning algorithm. The federated learning framework is used to achieve multi-device data fusion, protecting enterprise data privacy.

[0045] Preferably, the multimodal data fusion subunit stores a feature importance ranking algorithm. This algorithm dynamically evaluates and adjusts the fusion weights of each data source based on the structural complexity, material type, and current printing stage of the real-time printing task, improving the model's robustness and adaptability in different scenarios. For example, when printing cantilever structures, the weight of real-time image data is significantly increased; when controlling the printing temperature, the weights of thermal sensor data and material thermal response curves dominate.

[0046] The computer-readable storage medium of the geometric accuracy prediction subunit stores an image segmentation model or a three-dimensional point cloud analysis model. The geometric data of the design model is compared with the real-time acquired printed image or three-dimensional point cloud using the image segmentation model or the three-dimensional point cloud analysis model to calculate the local size deviation and generate a multi-dimensional error heat map.

[0047] Preferably, the image segmentation model employs a neural network model based on an encoder-decoder structure, such as the U-Net++ network. The U-Net++ network achieves pixel-level deviation segmentation, compares the designed model with the real-time image, generates an error heatmap, and locates areas with deviations ≥ 0.3 mm.

[0048] More preferably, a dilated convolution module is introduced in U-Net++ to expand the receptive field and capture geometric deviations at a larger scale.

[0049] Preferably, the 3D point cloud analysis model adopts a 3D point cloud processing model based on a region proposal network, such as the PointRCNN network. The PointRCNN network analyzes 3D point clouds and quantizes spatial offset vectors with an accuracy of ≤0.15mm.

[0050] More preferably, PointRCNN employs multi-scale point cloud feature extraction, while also paying attention to micro-detail deviations (such as surface roughness) and macro-structural deviations (such as overall size deformation).

[0051] The interlayer bond strength prediction subunit includes a prediction model that combines temporal modeling and an attention mechanism. Specifically, the computer-readable storage medium of the interlayer bond strength prediction subunit stores a temporal-attention hybrid model and an attention mechanism. The temporal-attention hybrid model is used to analyze the temperature history curve, cooling rate and extrusion pressure fluctuations during the material stacking process. The attention mechanism captures parameter mutations and outputs a probability score of interlayer bond strength and a warning of weak layer location.

[0052] The interlayer bonding strength prediction subunit is based on a time-attention hybrid model (such as the LSTM-Transformer architecture), taking time-series sensor data collected during the printing process as input. The input includes, but is not limited to, interlayer temperature history curves, cooling rate change sequences, and extrusion pressure fluctuation sequences. The model automatically captures abrupt changes (such as sudden temperature drops and pressure spikes) and long-term dependencies (such as the impact of continuous low temperatures on overall strength) in the above input features through its internal attention mechanism, and finally outputs two predicted values: bonding strength failure probability and weak layer location identifier.

[0053] Preferably, the interlayer bonding strength prediction subunit comprises a prediction model constructed from an LSTM network combined with a self-attention mechanism (Transformer). The sampling interval for analyzing the temperature history curve is ≤100ms; the cooling rate accuracy is controlled within ±0.1℃ / s. This hybrid model achieves a prediction accuracy of ≥92% for interlayer bonding strength and provides an advance warning of ≥5 layers for weak layers.

[0054] As a preferred approach, a time attention mechanism is introduced to focus on parameter changes at key time points (such as the instant of material extrusion and the initial stage of interlayer contact).

[0055] As a preferred approach, combining prior physical knowledge (such as the Arrhenius equation) with model training can enhance the generalization ability for new materials.

[0056] The structural deformation risk assessment subunit constructs physical simulation constraints by combining the viscoelastic constitutive equation of food materials, and uses a physics-inspired neural network (PINN) to fuse real-time sensor data to predict the deformation trend of the printed structure under changes in gravity, thermal stress or environmental humidity.

[0057] Specifically, the structural deformation risk assessment subunit stores a physically inspired neural network embedded with constraints from the Maxwell / Kelvin-Voigt viscoelastic model. Using material viscoelastic parameters, environmental parameters, and structural load parameters as input values, it outputs the deformation probability distribution and risk level (low / medium / high). The model's prediction error is ≤3%.

[0058] As a preferred approach, a multi-scale physical model is adopted, which considers the influence of material molecular structure on deformation at the microscopic level and optimizes the prediction by combining finite element analysis results at the macroscopic level.

[0059] As a preferred approach, uncertainty quantification techniques are used to output deformation confidence intervals, providing a more reliable basis for compensation decisions.

[0060] In this invention, the real-time error correction and compensation module analyzes the status data of the 3D printer during the execution of food 3D printing tasks in real time, and uses computer vision models, time series analysis models and physical simulation models to generate defect detection and compensation strategies.

[0061] The real-time error correction and compensation module includes a multimodal sensor unit (multimodal data acquisition unit), a defect detection and classification submodule, and a dynamic compensation engine.

[0062] The multimodal data acquisition unit integrates a camera or 3D scanner, an infrared thermal imager, a mechanical sensor, and an environmental sensor to construct a three-dimensional monitoring network. It collects status data in real time during the food 3D printing process and transmits the collected data to the data preprocessing unit of the dynamic accuracy prediction module and the defect detection and classification submodule of the real-time error correction and compensation module through communication technology, providing a data foundation for accuracy prediction and real-time error correction.

[0063] The camera or 3D scanner acquires real-time images of the printed layer surface and video streams of material flow, and captures surface defects in real time (resolution ≤ 0.1mm).

[0064] The infrared thermal imager collects the temperature distribution and thermal gradient changes of each layer of the sample in real time during the printing process, with temperature accuracy controlled within ±1℃.

[0065] The mechanical sensors include pressure sensors, strain sensors, and vibration sensors, which collect extrusion pressure (±0.01MP), interlayer stress (±5N), and mechanical vibration data (±0.1Hz) in real time.

[0066] The environmental sensor collects real-time data on humidity, airflow speed, and ambient temperature inside the printing chamber.

[0067] As a preferred approach, a distributed sensor layout is adopted, with multiple sets of sensors deployed on the print head and printing platform respectively, to achieve comprehensive monitoring of the printing process; sensor data are aligned through a time synchronization algorithm to ensure the spatiotemporal consistency of multimodal data.

[0068] As a preferred option, the multimodal data acquisition unit employs an adaptive data sampling strategy: dynamically adjusting the sampling frequency according to the printing stage. For example, during critical stages such as material extrusion and interlayer contact, the sampling rate of the industrial camera is increased to 60fps, the sampling rate of the thermal imager is increased to 25Hz, and the sampling rate of the force sensor is increased to 1kHz to ensure the capture of transient anomalies; during the idle or stable printing stages, the frequency is appropriately reduced to save computing resources.

[0069] The computer-readable storage medium of the defect detection and classification submodule stores a lightweight visual model, a temporal analysis model, and a multimodal fusion model.

[0070] As a preferred option, the lightweight vision model employs a lightweight convolutional neural network model (MobileNetV3). The model takes real-time acquired images of the printed layer surface as input and outputs defect classification labels and location information from the images. It performs real-time analysis of the surface images to identify cracks, bubbles, layer shifts, and material overflow, with a detection speed ≤150ms / frame and an accuracy ≥95%.

[0071] More preferably, during the model training phase, a knowledge distillation technique is used, employing a more complex pre-trained model as the teacher model to train the MobileNetV3 student model to extract its knowledge, thereby compressing the model parameters by 60% while maintaining high accuracy during final deployment.

[0072] Preferably, the time-series analysis model employs a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN) model. The model takes the time-series data of sensors (pressure, vibration, temperature, etc.) as input and the classification label of potential defects or their probability of occurrence as output. It analyzes the sensor time-series data to predict potential defects (such as faults caused by sudden pressure changes), providing an early warning lead time of ≥2 seconds.

[0073] More preferably, when building the model, the time series analysis model introduces residual connections and batch normalization to improve the modeling ability for long series data.

[0074] Preferably, the multimodal fusion model employs a graph neural network (GNN) or a gated attention mechanism model. The model takes preprocessed and extracted multimodal feature vectors as input and a comprehensive defect score and defect classification label as output. It fuses multimodal data to generate a comprehensive defect score (0-100 points) and classification label. The classification accuracy is ≥96%.

[0075] The dynamic compensation engine stores a dynamic compensation strategy library in its computer-readable storage medium. It generates compensation strategies based on the analysis results of the defect detection and classification submodule and drives the actuator to perform printing compensation tasks via a real-time control engine. The compensation strategies in the dynamic compensation strategy library include micro-compensation strategies (response ≤ 10ms), meso-compensation strategies (response ≤ 100ms), and macro-emergency strategies (response ≤ 1s).

[0076] The micro-compensation strategy dynamically adjusts the nozzle movement speed, extrusion pressure, or material flow rate to address minor deviations. These minor deviations include localized temperature fluctuations; for example, if the localized temperature fluctuation is ±2℃, the nozzle temperature is adjusted by ±0.5℃ or the extrusion speed by ±3%. Preferably, the Deep Deterministic Strategy Gradient (DDPG) algorithm is used to learn the optimal compensation parameters online. The agent learns by using the current sensor reading as its state, the parameter adjustment amount as its action, and the degree of improvement in the defect score as its reward.

[0077] The aforementioned mesoscopic compensation strategy targets moderate deviations, triggering local path replanning or fine-tuning of material proportions. These moderate deviations include poor interlayer bonding. Specifically, the material-aware mesoscopic compensation strategy is linked to a material database. When poor interlayer bonding is detected, the system queries the current material's rheological properties (such as thixotropy and yield stress) and generates a material-specific compensation scheme accordingly. For example, for materials with high thixotropy, a strategy of using a "high-frequency micro-amplitude vibration nozzle" is preferentially employed to restore its fluidity, or the support density is increased by 15% or a 3% thickener is injected.

[0078] The macro-level emergency strategy involves pausing printing and generating a repair plan for severe defects, including structural collapse. Preferably, a macro-level emergency plan simulation based on digital twins is used. When a severe defect necessitates triggering the macro-level emergency strategy, the system first performs a rapid simulation within the digital twin. The twin simulates the effects of different solutions, such as cleaning up residual material and reprinting the current layer, and selects the optimal solution (e.g., reprinting at 1.1 times the speed, with an expected reprinting accuracy ≤0.2mm) to send to the actual equipment for execution.

[0079] As a preferred approach, the compensation strategy employs a tiered response mechanism, automatically selecting the optimal compensation combination based on the severity of the defect.

[0080] As a preferred option, the dynamic compensation strategy library includes an interface enhancement compensation algorithm to address multi-material interface defects. This algorithm first queries the multi-material compatibility matrix in the database based on material pairing information to obtain parameters such as the ideal bonding temperature range and extrusion ratio; then, with the goal of improving interface bonding strength, it dynamically calculates and generates an optimized combination of process parameters.

[0081] The dynamic compensation engine also includes a real-time control engine for driving the actuator to perform printing supplement tasks. The real-time control engine achieves low-latency compensation through edge computing nodes and hardware collaboration interfaces.

[0082] Specifically, edge computing nodes: deploy lightweight models (such as Tiny-YOLO, TinyLSTM) locally on the printing device, with a decision latency of ≤80ms.

[0083] Hardware Collaboration Interface: Edge computing nodes establish direct communication links with actuators such as servo motors (response ≤20ms) and intelligent temperature control nozzles (±0.1℃ accuracy) to achieve millisecond-level execution of compensation commands.

[0084] As a preferred option, FPGA acceleration technology is used to optimize model inference speed. Lightweight models such as Tiny-YOLO and TinyLSTM are deployed on the FPGA chip local to the printing device. By utilizing the parallel computing capabilities of the FPGA, hardware acceleration is performed on key inference tasks (such as defect detection), further reducing the latency of key decisions to less than 50ms.

[0085] As a preferred option, redundant control channels are designed so that the backup channel is automatically switched when the main control channel fails, ensuring system reliability.

[0086] In this invention, the real-time error correction and compensation module further includes a feedback learning unit. This feedback learning unit prioritizes optimizing the model in the defect detection and classification submodule through online incremental learning. Simultaneously, the collected incremental dataset can also be used to periodically optimize and update the prediction model in the dynamic accuracy prediction module, and to provide data support for the offline training of the AI ​​instruction generation module, thereby achieving continuous collaborative evolution of the entire system.

[0087] Preferably, the online incremental learning uses the FedAvg algorithm to store error correction data (sensor data, defect labels, compensation effects) into the incremental dataset and update the model every 2 hours, thus shortening the adaptation period for new ingredients from 72 hours in the traditional approach to 6 hours.

[0088] As a preferred approach, an active learning strategy is introduced, prioritizing the labeling of data that is most valuable for improving model performance.

[0089] Preferably, model fusion technology is used to merge multiple local models learned online into a global model, thereby improving system stability. The local models refer to model copies obtained through online incremental learning based on local data from each independently deployed printing device; the global model refers to a unified model with stronger generalization capabilities generated by aggregating parameters from multiple local models on a central server using fusion algorithms such as federated averaging. This process achieves collaborative optimization and knowledge sharing across all terminal devices while protecting data privacy.

[0090] Another object of the present invention is to provide a method for 3D printing food based on the aforementioned control system, the method comprising the following steps: (1) Input and parse 3D model data to generate printing program instruction set: The user inputs food 3D model data, extracts geometric features, structural features and material distribution requirements through AI instruction generation module, and combines the rheological properties, thermal response curves and multi-material compatibility matrix in food 3D printing material database to generate a multi-material structure model adapted to 3D printing. After layering and slicing, a G code instruction set containing parameters such as spatial coordinates, extrusion amount, temperature and filling rate is generated. (2) Multimodal data fusion and accuracy prediction: The dynamic accuracy prediction module integrates material property parameters, historical printing data, real-time images and sensor data. Through the geometric accuracy prediction sub-unit, interlayer bonding strength prediction sub-unit and structural deformation risk assessment sub-unit, the accuracy assessment, strength prediction and risk warning of the printed product are output. (3) Real-time printing and defect detection: The printing status data is acquired in real time through the multimodal data acquisition unit of the real-time error correction and compensation module, and the defects are identified and a comprehensive score is generated by the defect detection and classification submodule; (4) Dynamic compensation and printing control: Based on the type and severity of the defect, the dynamic compensation engine calls the dynamic compensation strategy library to correct the printing deviation in real time.

[0091] This invention achieves automated high-precision program generation, dynamic accuracy prediction and risk warning, and multi-dimensional real-time error correction control through multimodal data intelligent fusion, machine learning model-driven approach, and full-process closed-loop control.

[0092] Preferably, the method includes: step (5), feedback learning and model optimization: storing the sensor data, defect labels, and compensation effects during the error correction process into an incremental dataset, and performing online incremental training on the model in the defect detection and classification submodule through a feedback learning unit. This step enables the system to learn and optimize autonomously.

[0093] Preferably, the method further includes: step (6), quality assessment and post-processing: after printing, the printed product is comprehensively assessed offline through the dynamic accuracy prediction module, and a quality assessment report is generated. If the product has repairable defects, the system calls the post-processing strategy library according to the assessment report and drives the post-processing execution mechanism to perform corresponding repair operations, such as surface trimming and local reinforcement.

[0094] Through the above steps, the AI-based food 3D printing control system can monitor and dynamically adjust in real time during the 3D printing process, and continuously optimize through feedback learning, ultimately achieving the goal of printing perfect food products.

[0095] In this invention, applicable food materials can be, but are not limited to, the following types: edible base ingredients selected from one or more of flour, starch, protein, fat, sugar, cellulose, and their derivatives; functional additives selected from one or more of thickeners, emulsifiers, nutritional fortifiers, and colorings; and functional ingredients selected from one or more of vitamins, probiotics, and minerals. The food materials comply with national food safety standards and can be directly used in 3D printing to form foods with specific shapes, structures, and nutritional functions.

[0096] The beneficial effects of this invention are as follows: This invention addresses the problems of insufficient perception of food characteristics and rigid control strategies in existing systems by constructing an AI-driven multimodal data fusion model, combined with a real-time sensor network and a dynamic compensation engine. It achieves intelligent optimization of the entire process from model analysis to printing completion, significantly improving the accuracy and process versatility of food 3D printing. Attached Figure Description

[0097] Figure 1 This is a framework diagram of an artificial intelligence-based food 3D printing control system according to the present invention.

[0098] Figure 2 This is a schematic diagram of the process of a food 3D printing control method based on artificial intelligence according to the present invention.

[0099] Figure 3 To compare the effects of traditional methods and the artificial intelligence control method of this invention in the production of 3D printed food products, the upper image shows the traditional method and the lower image shows the method of this invention. Detailed Implementation

[0100] This invention provides an artificial intelligence-based food 3D printing control system, such as... Figure 1As shown, the control system includes: an AI instruction generation module, a dynamic accuracy prediction module, and a real-time error correction and compensation module. These three modules work together to achieve intelligent control of the entire process of food 3D printing. The three modules form a closed loop with the food material 3D printing database through multi-directional data interaction and feedback. (1) AI instruction generation module generates printing instructions based on the material and historical data of the food material 3D printing database; at the same time, it receives the accuracy prediction results of the dynamic accuracy prediction module and the real-time status data collected by the real-time error correction and compensation module, and iteratively optimizes the instructions. (2) Dynamic accuracy prediction module, which integrates food material 3D printing database data and multimodal data collected by real-time error correction and compensation module, predicts accuracy / risk and feeds it back to AI instruction generation module (optimize instructions) and real-time error correction and compensation module (pre-formulate compensation strategy). (3) Real-time error correction and compensation module: collects real-time printing data, combines the database to formulate compensation strategies to drive printing execution, and feeds back data such as defects and compensation effects to AI instruction generation module (to help model learning), dynamic accuracy prediction module (to update prediction model) and database (to enrich historical data). (4) Food material 3D printing database, which stores and updates data such as materials, historical printing, and defects of each module, providing data support for the entire process.

[0101] Ultimately, it achieves intelligent closed-loop control throughout the entire process of "instruction generation → accuracy prediction → real-time error correction → data accumulation → instruction optimization".

[0102] The present invention will be further described below with reference to specific embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Any modifications or substitutions made to the methods, steps, or conditions of the present invention without departing from the spirit and essence of the invention are within the scope of the invention.

[0103] The tests and verifications involved in the following embodiments were all conducted based on conventional food 3D printing process conditions and commercially available food materials. The testing methods used were all conventional testing and measurement methods recognized in the art, such as using a 3D scanner or laser scanner for dimensional accuracy measurement, and a texture analyzer or rheometer for mechanical property testing. Printing materials included, but were not limited to, common protein-based, carbohydrate-based, and lipid-based edible materials and their commercial additives (such as thickeners and gelling agents). The training and testing of the involved artificial intelligence models were all implemented based on publicly available algorithm frameworks and conventional computing platforms.

[0104] Example 1: AI Instruction Generation Module The AI ​​instruction generation module is used for intelligent analysis of 3D printing models and generation of 3D printing program instruction sets. It consists of a 3D model processing unit, a printing model generation unit, and an intelligent slicing unit.

[0105] 3D model processing unit: Receives user input instructions for 3D food models and identifies the geometric features, structural features, and material distribution requirements of the model through feature extraction algorithms; Printing model generation unit: Based on the parameters output by the 3D model processing unit, combined with the food material 3D printing database, it generates a multi-material structure model adapted for 3D printing; Intelligent slicing unit: The multi-material structure model is sliced ​​into layers to generate a G-code instruction set including parameters such as spatial coordinates, extrusion amount, speed, temperature, infill rate, and printing height.

[0106] In this embodiment, the AI ​​instruction generation module has been optimized at the algorithm level to improve the intelligence level and processing efficiency of 3D model parsing and printing instruction generation. Specific optimizations include the following aspects: (1) Algorithm optimization of the three-dimensional model processing unit The 3D model processing unit supports multiple input methods, including voice, text, files, and images. To improve the accuracy and robustness of model feature extraction, this embodiment employs the following algorithm optimization measures: Edge enhancement preprocessing algorithm: This algorithm performs noise reduction and edge enhancement on the input image or model file. By combining the Difference of Gaussian (DoG) operator with adaptive threshold segmentation, it effectively suppresses noise interference and enhances the saliency of contour and structural information. This preprocessing algorithm has been verified through multiple printed tests, consistently improving feature extraction accuracy to over 98%.

[0107] Multi-Scale Feature Fusion Network (MSFFN): For 3D models with complex curved surfaces, hollow structures, or thin-walled regions, a multi-scale feature fusion network is constructed. This network extracts local and global features from different receptive fields through parallel convolutional paths and uses an attention mechanism for weighted fusion to ensure complete recognition of detailed features. Compared with single-scale extraction methods, this network reduces the false negative rate by more than 40% in complex structure recognition tasks.

[0108] Feature extraction algorithm optimization: The system integrates multiple advanced algorithms such as PointNet++, MeshCNN, and Multi-Scale Feature Fusion Network (MSFFN), and can automatically select the optimal extraction strategy based on model complexity. Geometric features include curvature, wall thickness, and connectivity; structural features include support requirements and hierarchical complexity. Output feature parameters include, but are not limited to, key dimensional information such as side length, height, and radius. Their values ​​are calculated using the vertex coordinates and topological relationships of the model mesh, with an error controlled within ±0.1 mm.

[0109] (2) Algorithm optimization for printing model generation unit The printing model generation unit is responsible for generating multi-material structure models based on the 3D model features and the material database, including material zoning planning, support structure design, and printing path optimization. The following optimization strategy is adopted in this embodiment: Material partitioning using a semantic segmentation network: A U-Net-based semantic segmentation network is used to divide the 3D model into functional regions. The network outputs the material category probability for each voxel through forward inference, and then combines this with a material matching algorithm to automatically match the optimal material combination from a pre-set database. For example, protein materials with an elastic modulus ≥ 500 kPa are used for the outer shell structure, and starch matrix with a viscosity ≤ 2000 mPa·s is used for the internal filling. This allocation strategy, verified by mechanical simulation, can significantly improve the overall stability of the printed structure.

[0110] Topology optimization generates edible supports: A topology optimization algorithm based on the variable density method is used to generate a lattice-type honeycomb support structure with the goal of maximizing structural stiffness. Through iterative calculations, the volume ratio of the supports can be controlled below 12% while ensuring support strength. Low-melting-point edible materials (such as gelatin or chocolate) are selected for the supports. The melting point of the materials is determined through thermodynamic simulation and printing experiments to ensure that the support structure is easy to peel off and does not affect the taste of the food.

[0111] Reinforcement learning optimizes printing paths: A Markov Decision Process (MDP) model is constructed for the printing path optimization problem. Printing time, material changeover times, and idle distance are optimized as objectives, and a proximal policy optimization (PPO) algorithm is used for policy search. The trained policy can automatically generate S-shaped transition paths in multi-material interface regions, reducing the risk of material mixing. Compared with traditional greedy algorithms, this reinforcement learning strategy can reduce overall printing time by more than 25% and significantly improve path smoothness.

[0112] (3) Algorithm optimization of intelligent slicing unit The intelligent slicing unit employs a Transformer encoder-decoder architecture to perform layered slicing of the 3D model, generating a high-precision G-code instruction set. This embodiment further optimizes the following: Adaptive Layer Thickness Control: An adaptive layer thickness adjustment algorithm based on curvature feedback was developed. This algorithm calculates the local curvature of the model and sets the layer thickness-curvature response function, automatically reducing the layer thickness to 0.15mm for complex regions with curvature greater than 60°. Simultaneously, a gated attention mechanism is introduced into the Transformer decoder to enhance parameter control over the detailed features of high-curvature regions, improving the accuracy of printing instructions in these areas by approximately 30%.

[0113] Gradient Infill Strategy: For the infill structure inside the model, a gradient density design method based on structural stress distribution is proposed. This method simulates the stress distribution during the printing process through finite element analysis (FEA) and automatically reduces the infill density in low-stress areas.

[0114] The execution process of the gradient filling strategy includes: a region division step: based on the structural features of the 3D model, its interior is divided into a core load-bearing area, a transition area, and an edge non-load-bearing area; a gradient planning step: assigning a decreasing gradient filling rate and a suitable filling pattern to the above different areas; and a simulation optimization step: based on the rheological property data in the food material 3D printing database, performing physical simulation and iterative optimization on the filling structure, and finally generating a filling scheme that reduces material usage while ensuring structural strength.

[0115] Printing experiments have verified that the gradient filling strategy can reduce material usage by 15%-30% while ensuring structural strength, and effectively avoid internal defects caused by uneven filling.

[0116] Through the aforementioned algorithm optimization and system implementation, the AI ​​instruction generation module achieves fully automated generation from multimodal input to a high-precision G-code instruction set. All optimization effects have been verified through the system's built-in simulation module and actual printing tests. It can adapt to various food materials and complex structural models without relying on manual debugging, significantly improving the system's intelligence level and printing efficiency.

[0117] Example 2: Dynamic Accuracy Prediction Module The dynamic accuracy prediction module includes a data preprocessing unit, a multimodal data fusion subunit, a geometric accuracy prediction subunit, an interlayer bonding strength prediction subunit, and a structural deformation risk assessment subunit.

[0118] The data preprocessing unit is used to perform parallel preprocessing operations on multi-source heterogeneous data, including normalization, feature alignment and noise filtering. The multi-source heterogeneous data includes: material property parameters and historical printing data stored in the food material 3D printing database, as well as image data and sensor data collected in real time by the multi-modal sensor unit. Multimodal data fusion subunit: Employing a spatiotemporal sequence model or graph structure analysis model, preprocessed multi-source heterogeneous data is fused, mapping data from different modalities to a unified feature space, establishing spatiotemporal correlations between data, and outputting a feature tensor of temporal-spatial correlation; the feature tensor serves as the input value for the geometric accuracy prediction subunit, the interlayer bonding strength prediction subunit, and the structural deformation risk assessment subunit; Geometric accuracy prediction subunit: The geometric data of the design model is compared with the real-time acquired printed image or 3D point cloud using an image segmentation model or a 3D point cloud analysis model to calculate local size deviations and generate a multi-dimensional error heat map. Interlayer bond strength prediction sub-unit: The time-attention hybrid model is used to analyze the temperature history curve, cooling rate and extrusion pressure fluctuation during the material stacking process. The attention mechanism is used to capture parameter mutations and output the probability score of interlayer bond strength and the early warning of weak layer location. Structural Deformation Risk Assessment Subunit: Combining the viscoelastic constitutive equation of food materials to construct physical simulation constraints, and using a physics-inspired neural network to fuse real-time sensor data, the deformation trend of the printed structure under changes in gravity, thermal stress or environmental humidity is predicted.

[0119] In this embodiment, the dynamic accuracy prediction module has undergone in-depth algorithm optimization, aiming to achieve accurate prediction and risk warning of printing accuracy, interlayer bonding strength, and structural deformation through multimodal data fusion and advanced prediction models. The specific optimization content and implementation methods are as follows: (1) Algorithm optimization of the multimodal data fusion subunit This subunit is responsible for integrating heterogeneous data from multiple sources, including material properties, historical printing data, real-time images, and sensor data. To achieve efficient fusion, this embodiment adopts the following optimization scheme: A data fusion framework based on Spatiotemporal Graph Convolutional Network (ST-GCN) is constructed. A spatiotemporal graph structure is built, treating each monitoring point on the printing platform as a graph node. Node attributes include data such as temperature, pressure, and image features, while edge weights reflect spatial proximity and temporal correlation. ST-GCN simultaneously extracts spatial features and temporal evolution patterns, generating a unified high-dimensional feature representation. Experimental results demonstrate that this framework effectively captures nonlinear correlations between multimodal data, improving the information entropy of the fused features by approximately 40%.

[0120] Data Collaboration Training under a Federated Learning Framework: To protect data privacy across different printing devices, a federated learning framework is used for distributed model training. Each local device trains its ST-GCN sub-model using local data, uploading only the model gradients to a central server for aggregation and updates. Testing showed that this approach reduces the generalization error of the prediction model by approximately 18% while ensuring data security.

[0121] Dynamic Feature Weight Adjustment Mechanism: An attention-based feature importance scoring algorithm is introduced to calculate the contribution of each data source (such as material parameters, image data, and pressure sensor data) to the current prediction task in real time and dynamically adjust their fusion weights. For example, when a large fluctuation in material viscosity is detected, the weight of pressure sensor data is automatically increased; when the printed structure is complex, image data is given higher importance. This mechanism improves the model's prediction stability under varying operating conditions by more than 25%.

[0122] (2) Algorithm optimization of geometric accuracy prediction sub-unit This subunit is responsible for quantifying the geometric deviation between the printed product and the design model. This embodiment improves its predictive performance through the following measures: Integration of U-Net++ network with dilated convolution: Pixel-level deviation segmentation of the printed layer surface image is performed using the U-Net++ network. To further expand the receptive field to capture large-scale deformations, a dilated convolution module (with dilation rates set to 2, 4, and 8) is introduced into the network encoder. Comparative tests show that this optimization improves the model's accuracy in detecting overall deformations greater than 0.5 mm from 88% to 96%, while maintaining high sensitivity to minute defects (such as 0.1 mm edge burrs).

[0123] 3D Point Cloud Spatial Migration Analysis Based on PointRCNN: This study utilizes the PointRCNN network to process high-precision point cloud data (accuracy 0.05 mm) acquired by a 3D scanner. The network first generates candidate point cloud regions through a Region Proposal Network (RPN), then performs fine spatial transformation regression on each region, outputting a 3D migration vector for each point. Experiments show that the overall spatial migration prediction accuracy of this method is ≤0.15 mm, and its prediction consistency for complex curved surfaces is significantly better than the traditional ICP algorithm.

[0124] (3) Algorithm optimization for predicting interlayer bonding strength sub-units Interlayer bonding strength is a key factor affecting the stability of food printing structures. This unit is optimized as follows: The LSTM-Transformer hybrid model is constructed by employing a Long Short-Term Memory (LSTM) network to capture the long-term temporal dependencies of sensor data such as temperature and pressure. Simultaneously, a Transformer encoder is introduced to apply attention weights to key time steps (such as the instant of material extrusion or the initial stage of interlayer contact), accurately identifying abrupt parameter changes (such as sudden temperature drops or pressure peaks). This hybrid model achieves a prediction accuracy of ≥92% for interlayer bonding strength and provides an early warning lead time of ≥5 layers for weak layers.

[0125] Physical prior knowledge constrains model training: To improve the model's ability to predict the thermodynamic behavior of new materials, the Arrhenius equation is embedded as a physical constraint in the loss function, forcing the model to ensure that the predicted interlayer bonding strength and temperature history satisfy the mathematical relationship of this equation. This measure reduces the model's prediction error on untrained materials by an average of approximately 30%.

[0126] (4) Algorithm optimization of the structural deformation risk assessment subunit This sub-unit aims to predict the deformation risk of the printed structure under the influence of gravity, thermal stress, etc. Optimizations in this embodiment include: The fusion of Physics-Inspired Neural Networks (PINN) and viscoelastic constitutive models: Maxwell or Kelvin-Voigt viscoelastic models are embedded as partial differential equation constraints into the PINN network. The network input consists of real-time sensor data and material parameters. Automatic differentiation techniques are used to solve the constitutive equations and predict the deformation field. This hybrid method achieves a deformation prediction error ≤3% and can output the deformation probability distribution and risk level (low / medium / high).

[0127] Multi-scale prediction and uncertainty quantification: At the microscale, molecular dynamics simulation data is introduced as auxiliary input to consider the influence of material molecular structure (such as polysaccharide chain length and protein crosslinking degree) on macroscopic deformation; at the macroscale, PINN predictions are corrected by combining finite element analysis (FEA) results. Simultaneously, Monte Carlo Dropout technology is used for multiple inferences to output the confidence interval of deformation predictions, providing a reliability assessment for decision-making.

[0128] Through the above algorithm optimization, the dynamic accuracy prediction module achieves accurate and real-time prediction of multi-dimensional quality indicators in the printing process. All optimization effects were verified through offline simulation and online printing tests. The results show that this module can significantly reduce the printing failure rate and provide reliable data support for subsequent real-time error correction.

[0129] Example 3: Real-time Error Correction and Compensation Module The real-time error correction and compensation module analyzes the status data during the food 3D printing process in real time and uses computer vision models, time series analysis models and physical simulation models to generate defect detection and compensation strategies.

[0130] The real-time error correction and compensation module consists of a multimodal sensor unit, a defect detection and classification submodule, and a dynamic compensation engine.

[0131] The multimodal data acquisition unit integrates a camera or 3D scanner (real-time acquisition of surface images of the printed layers and video streams of material flow, with a resolution ≤0.1mm), an infrared thermal imager (real-time acquisition of temperature distribution and thermal gradient changes of each layer of the sample during the printing process, with temperature accuracy controlled within ±1℃), mechanical sensors (including pressure sensors, strain sensors, and vibration sensors), and environmental sensors (real-time acquisition of humidity, airflow speed, and ambient temperature in the printing chamber). This constructs a three-dimensional monitoring network to acquire state data during the food 3D printing process in real time. The acquired data is transmitted to the data preprocessing unit of the dynamic accuracy prediction module and the defect detection and classification submodule of the real-time error correction and compensation module through communication technology, providing a data foundation for accuracy prediction and real-time error correction.

[0132] The defect detection and classification submodule utilizes a lightweight visual model, a temporal analysis model, and a multimodal fusion model for defect detection.

[0133] The dynamic compensation engine generates a compensation strategy based on the analysis results of the defect detection and classification submodule, and drives the actuator to perform the printing compensation task through the real-time control engine.

[0134] In this embodiment, the real-time error correction and compensation module has undergone systematic algorithm optimization, aiming to build a millisecond-level closed-loop control system with rapid response and accurate decision-making, so as to realize the real-time detection, classification and dynamic compensation of various defects in the printing process.

[0135] (1) Synchronization and optimization of multimodal data acquisition units This unit forms the foundation of the system's perception layer, responsible for acquiring multi-dimensional state data with high frequency and high precision. This embodiment incorporates the following optimizations: Distributed sensor deployment and time synchronization algorithm: Multiple sets of sensors (including industrial cameras, infrared thermal imagers, and pressure / strain / vibration sensors) are distributed in key locations such as the printhead, extrusion mechanism, and printing platform. To solve the spatiotemporal consistency problem of multi-source data, a hybrid synchronization algorithm based on hardware triggering and software timestamps is adopted, ensuring that the data acquisition time deviation of all sensors is controlled within ±5ms, providing a solid foundation for subsequent multimodal fusion.

[0136] Adaptive data sampling strategy: The sampling frequency is dynamically adjusted according to the printing stage. In critical stages such as material extrusion and interlayer contact, the sampling rate of the industrial camera is increased to 60fps, the sampling rate of the thermal imager is increased to 25Hz, and the sampling rate of the force sensor is increased to 1kHz to ensure the capture of transient anomalies; in the idle movement or stable printing stages, the frequency is appropriately reduced to save computing resources.

[0137] (2) Algorithm acceleration and accuracy improvement of defect detection and classification submodule This submodule needs to complete defect identification and classification in a very short time. The optimization focus of this embodiment is on balancing accuracy and speed: Lightweight Visual Model Based on Knowledge Distillation: This approach employs knowledge distillation technology, using a large pre-trained model (such as ResNet-50) as the teacher model to guide the training of a lightweight student model (MobileNetV3-small). The final deployed model has 60% fewer parameters, and the inference speed is increased to ≤150ms / frame, while maintaining a detection accuracy of ≥95% for typical defects such as cracks, bubbles, and layer shifts.

[0138] Temporal Convolutional Networks (TCNs) replace LSTMs for sensor signal analysis: To overcome the slow inference speed of LSTMs in processing long sequences, TCN models with causal convolutional structures are used to analyze sensor time-series data (such as pressure and vibration). The parallel structure of TCNs improves inference speed by about 40% compared to LSTMs and can effectively capture long-term dependencies in sensor data, providing an early warning lead time of ≥2 seconds for potential defects such as sudden changes in extrusion pressure.

[0139] Gated Attention Units (GAUs) are used for multimodal fusion: to fuse multimodal data such as images, thermal images, and mechanical data, a more computationally efficient gated attention mechanism is used to replace the standard Transformer. GAU dynamically weights the importance of each modality feature through gated linear and nonlinear transformations, generating a comprehensive defect score (0-100 points) and classification label, with an overall classification accuracy of ≥96%, and the fusion decision time is controlled within 50ms.

[0140] (3) Hierarchical and optimization of dynamic compensation strategy library The compensation strategy library is the "knowledge brain" of the system, and this embodiment has refined and upgraded it with intelligence: Micro-compensation parameter tuning based on reinforcement learning: For micro-compensation (such as adjusting nozzle temperature and extrusion speed), instead of relying solely on fixed rules, the Deep Deterministic Policy Gradient (DDPG) algorithm is used to learn the optimal compensation parameters online. The agent learns by using the current sensor reading as the state, the parameter adjustment amount as the action, and the degree of improvement in the defect score as the reward. This strategy improves the suppression effect of micro-compensation on minor deviations by approximately 30%.

[0141] Mesoscopic compensation strategies based on material perception: Mesoscopic compensation strategies (such as path replanning and material ratio fine-tuning) are linked to the material database. When poor interlayer bonding is detected, the system queries the current material's rheological properties (such as thixotropy and yield stress) and generates a material-specific compensation scheme accordingly. For example, for materials with strong thixotropy, the strategy of "high-frequency micro-amplitude vibration nozzle" is preferred to restore its flowability, rather than simply increasing the extrusion pressure.

[0142] Macro-level emergency response simulation based on digital twins: When faced with a serious defect requiring the triggering of macro-level emergency strategies (such as pausing printing or generating a repair plan), the system first performs a rapid simulation in the digital twin. The twin simulates the effects of different solutions, such as cleaning up residual material and reprinting the current layer, and selects the optimal solution (such as reprinting at 1.1 times the speed with an expected reprint accuracy of ≤0.2mm) to be sent to the real equipment for execution, greatly improving the reliability of emergency decisions.

[0143] (4) Low-latency optimization of the real-time control engine To achieve millisecond-level control, this embodiment has undergone in-depth optimization at the hardware collaboration and edge computing levels: FPGA-accelerated model inference: Lightweight models such as Tiny-YOLO and TinyLSTM are deployed on the local FPGA chip of the printing device. Utilizing the parallel computing capabilities of the FPGA, key inference tasks (such as defect detection) are hardware-accelerated, further reducing decision latency from ~80ms in software operation to ≤50ms.

[0144] High-priority hardware communication channel: A dedicated real-time communication protocol is designed to enable edge computing nodes to establish direct communication links with actuators such as high-precision servo motors and intelligent temperature-controlled nozzles, bypassing the uncertainty delay caused by operating system scheduling and ensuring that the end-to-end latency from the generation to the execution of compensation instructions is stable within 100ms.

[0145] (5) Online optimization of feedback learning units The system has the ability to continuously evolve, and this embodiment optimizes its online learning mechanism: Incremental data selection based on active learning: Instead of storing all error correction data, the feedback learning unit adopts an active learning strategy to prioritize storing the samples that are least uncertain in the model (e.g., classification probability close to 0.5) or most representative into the incremental dataset. This reduces the amount of labeled data required for the model to update every 2 hours by about 50%, significantly improving learning efficiency.

[0146] FedAvg and Model Fusion: Multiple printing devices within the system act as clients of federated learning, updating their local models using local incremental data. The central server periodically aggregates model parameters using the FedAvg algorithm and employs ensemble learning techniques to fuse multiple global model versions, ultimately resulting in a more stable and generalized enhanced model. This shortens the system's adaptation cycle to new ingredients to 1-2 hours.

[0147] Through the above algorithm optimization and system implementation, the real-time error correction and compensation module constructs an efficient and intelligent closed-loop control system, which can cope with various uncertainties in the food 3D printing process and significantly improve the printing success rate and product quality. All optimization effects have been verified through rigorous hardware-in-the-loop (HIL) simulation and actual printing tests.

[0148] Example 4: System Integration and Multi-Material Collaborative Printing Optimization This embodiment focuses on the integrated application and optimization of the AI-based food 3D printing end-to-end control system, particularly its overall workflow and performance improvement in the complex scenario of multi-material collaborative printing. This embodiment aims to demonstrate how the aforementioned modules work collaboratively and, through system-level algorithm optimization, ultimately achieve high-quality, high-efficiency multi-material food printing.

[0149] (1) System-level workflow integration and optimization This embodiment first defines the end-to-end intelligent workflow of the system from receiving the task to completing the printing: Task Analysis and Intelligent Material Recommendation: After a user uploads a 3D model of the target product, the system does not immediately begin slicing. Instead, it first activates the material compatibility analysis engine. This engine calls upon a multi-material compatibility matrix from the material database, combining the model's geometric features (such as overhanging structures and thin walls) and functional requirements (such as texture and nutritional balance), to recommend the optimal combination of base materials and functional additives using a pre-trained graph neural network (GNN) model. For example, for a pastry model requiring a "crispy crust and soft center," the system recommends a high-viscosity, high-elasticity protein-based material for the outer shell and a low-viscosity, high-flow-rate jam-like material for the filling, automatically avoiding incompatible combinations that could lead to interface separation.

[0150] Full-process digital simulation and parameter pre-optimization: Before actual printing begins, the system utilizes an integrated digital twin engine for full-process virtual printing. The simulation process strictly follows physical laws, including non-Newtonian fluid flow, heat conduction, phase transitions, and multi-material interface interactions. Through simulation, potential problems such as interface delamination and thermal deformation are predicted in advance, and the simulation results are used to pre-optimize and adjust the initial G-code output by the AI ​​instruction generation module (such as fine-tuning the printing speed and temperature gradient in different material regions), minimizing the learning cost that traditional methods require in the face of printing failures. Tests show that this pre-optimization process can improve the first-time success rate of multi-material printing by approximately 50%.

[0151] (2) Enhanced algorithm for multi-material interface combination The core challenge in multi-material printing lies in the interfacial bonding strength. This embodiment employs a specific algorithm to enhance this aspect: Dynamic thermal management strategy for the interface area: The system equips the printhead with multiple nozzles that have independent and precise temperature control. When printing at the interface area of ​​two materials, the dynamic accuracy prediction module calculates and predicts the thermal history of the bonding point in real time. The real-time error correction and compensation module then dynamically adjusts the residence time of the first printed material layer (waiting for it to cool to the optimal bonding temperature) and the temperature of the nozzle of the second printed material (slightly increasing it to promote molecular diffusion), forming a dynamic "thermal handshake" protocol that improves the interface bonding strength by an average of about 35%.

[0152] Generation of micro-mechanical interlocking structures: For certain material combinations with poor compatibility, the AI ​​instruction generation module automatically generates micro-mechanical interlocking structures (such as micrometer-scale dovetail or anchor point structures) in the G-code of its interface area. This design compensates for insufficient chemical bonding by increasing the bonding area and mechanical interlocking. These microstructures are automatically created by the intelligent slicing unit during slicing, and their scale and morphology are optimized by finite element analysis (FEA) to ensure effectiveness without affecting the macroscopic product form.

[0153] (3) System adaptive learning and cross-platform deployment optimization To enable the system to continuously adapt to new materials and be easy to deploy, the following optimizations were made in this embodiment: A Meta-Learning-Based Fast Material Adaptation Framework: When faced with a completely new food material, the system no longer trains its model from scratch. We introduce a meta-learning framework. This framework is pre-trained on a large amount of existing material data, enabling it to learn "how to learn quickly." When encountering a new material, only a few (e.g., 5-10) printing tests are needed for the system to quickly adjust its internal prediction and compensation models, reducing the optimization cycle for printing parameters of new materials from several days to several hours.

[0154] Cloud-Edge-Device Collaborative Inference Architecture: To balance computational load and real-time requirements, the system adopts a collaborative inference architecture. The energy-intensive pre-trained model, digital twin simulation, and meta-learning process run on high-performance servers in the cloud or locally (cloud). The distilled and compressed lightweight inference model is deployed on the edge computing nodes of the printing device itself (edge). Sensors and actuators on the device act as endpoints. All three collaborate efficiently through an optimized communication protocol, ensuring high-level intelligence without sacrificing millisecond-level real-time control capabilities.

[0155] Example 5: Optimization of Intelligent 3D Printing Method for Food This embodiment provides a food printing method in which a 3D printer is controlled by an AI-based food 3D printing control system. This method achieves intelligent control of the entire process from model to finished product through the collaborative work of multiple modules within the system. A flowchart is shown below. Figure 2 As shown, the specific steps include: (1) Input and parse 3D model data to generate printing program instruction set. Users input their food printing needs via voice, text, or images. The system uses a generative AI model to convert unstructured input into a standardized 3D model. Subsequently, the model analysis unit automatically identifies the model's geometric features (including curvature distribution, minimum wall thickness, and overhang structures) and functional zones using deep learning algorithms. Based on rheological properties, thermodynamic parameters, and multi-material compatibility data in the materials knowledge base, it generates material allocation schemes adapted to the characteristics of different regions. Printing path planning employs a reinforcement learning algorithm, optimizing printing efficiency, material utilization, and structural stability. It automatically generates a high-precision G-code instruction set containing parameters such as spatial coordinates, extrusion amount, temperature, and infill rate.

[0156] (2) Multimodal data fusion and accuracy prediction The system integrates multi-source sensor data (including material properties, real-time images, temperature field distribution, pressure changes, and historical printing data) to construct a dynamic accuracy prediction model. This model analyzes real-time point cloud and image data using a geometric accuracy prediction unit to identify dimensional deviations and deformation trends; it integrates thermal history and mechanical data using an interlayer bonding strength prediction unit to assess interlayer bonding quality; and it uses a structural deformation risk unit combined with viscoelastic constitutive equations to predict the instability risk of complex structures during the printing process, outputting quantitative assessment results and early warning signals.

[0157] (3) Real-time printing and defect detection During the printing process, the system continuously monitors the printing status through a multimodal sensing unit, uses a lightweight convolutional network and a temporal model to perform real-time analysis of images and sensor data, identifies common defect types including interlayer delamination, material leakage, and structural warping, and generates a comprehensive risk score based on the defect characteristics.

[0158] (4) Dynamic compensation and printing control Based on the defect detection results, the system uses a dynamic compensation engine to call a pre-set strategy library to achieve hierarchical response control: for local temperature fluctuations or extrusion anomalies, the system adjusts the printhead temperature and extrusion rate in real time; for poor interlayer bonding or regional deformation, the system dynamically inserts auxiliary supports or adjusts the printing path; for severe structural defects, the system initiates a pause-repair-resume printing process to minimize printing failure rate and material waste.

[0159] (5) Feedback learning and model optimization The system stores multimodal data, defect labels, and compensation effects collected during the printing process into an incremental dataset, and iteratively optimizes key algorithms involved in the detection and compensation models based on an online learning mechanism. By continuously integrating printing data from new materials and new working conditions, the system's adaptability and robustness in diverse printing tasks are improved.

[0160] This method achieves closed-loop intelligent control of the entire food 3D printing process through the above steps, significantly reducing the reliance on human experience in the printing process, improving the printing success rate and product consistency, and is suitable for food printing scenarios with multiple materials, multiple structures, and high quality requirements.

[0161] Example 6: Comparison of Overall Performance between Traditional and Intelligent Food 3D Printing To verify the superiority of the intelligent control method of the present invention over the traditional method, a comparative experiment was conducted based on the SHINNOVE-S2 food 3D printer (Hangzhou, China, Shiyin Technology Co., Ltd.).

[0162] Traditional printing methods have fixed process parameters: extrusion temperature 25°C, printing travel speed, filling speed and shrinkage speed are all 30 mm / s, filling density is 80%, and a nozzle diameter of 0.84 mm is used. The whole process relies on preset G-codes and operator experience.

[0163] In contrast, the intelligent system of this invention achieves a fundamental upgrade on the same hardware platform by integrating a multimodal sensor network (including a high-resolution industrial camera, an infrared thermal imager, a force sensor group, and environmental sensors, distributed in key locations) and three intelligent control modules (AI command generation, dynamic accuracy prediction, and real-time error correction and compensation). During testing, the intelligent system can perceive the printing status in real time and dynamically adjust various parameters (such as temperature, speed, and path) instead of adhering to fixed values.

[0164] like Figure 3 As shown, the results indicate that transforming the traditional open-loop control mode, which relies on fixed parameters and manual intervention, into a fully intelligent closed-loop control system driven by multimodal data and capable of real-time perception, prediction, and automatic error correction reduces the reliance on operator skills by 90% due to the high degree of automation of the entire process, significantly lowering labor costs and experience barriers. It also significantly improves the first-time success rate of complex multi-material printing from approximately 65% ​​with traditional methods to 98%, and reduces the average dimensional error from 0.35 mm to less than 0.1 mm.

Claims

1. An artificial intelligence-based food 3D printing control system, characterized in that, include: AI instruction generation module: used to parse the food 3D model data input by the user, combine the rheological properties, thermal response curves and multi-material compatibility matrix in the food material 3D printing database, and generate an appropriate printing program instruction set through machine learning model; receive data feedback from dynamic accuracy prediction module and real-time error correction and compensation module for model learning and optimization. Dynamic accuracy prediction module: It integrates data from the food material 3D printing database and multimodal data collected by the real-time error correction and compensation module to predict the geometric accuracy, interlayer bonding strength and structural deformation risk of the printed product; it feeds back the prediction results to the AI ​​instruction generation module and the real-time error correction and compensation module; and it receives data from the real-time error correction and compensation module to update the prediction model. Real-time error correction and compensation module: includes a multimodal sensor unit, a defect detection and classification submodule and a dynamic compensation engine. The multimodal sensor unit collects state data in real time during the food 3D printing process. The defect detection and classification submodule uses computer vision models and time series analysis models to detect defects. The dynamic compensation engine generates compensation strategies based on the prediction results fed back by the dynamic accuracy prediction module and the analysis results of the defect detection and classification submodule. The dynamic accuracy prediction module includes: The data preprocessing unit is used to perform parallel preprocessing operations on multi-source heterogeneous data, including normalization, feature alignment and noise filtering. The multi-source heterogeneous data includes: material property parameters and historical printing data stored in the food material 3D printing database, as well as image data and sensor data collected in real time by the multi-modal sensor unit. Multimodal data fusion subunit: Employing a spatiotemporal sequence model or graph structure analysis model, preprocessed multi-source heterogeneous data is fused, mapping data from different modalities to a unified feature space, establishing spatiotemporal correlations between data, and outputting a feature tensor of temporal-spatial correlation; the feature tensor serves as the input value for the geometric accuracy prediction subunit, the interlayer bonding strength prediction subunit, and the structural deformation risk assessment subunit; Geometric accuracy prediction subunit: The geometric data of the design model is compared with the real-time acquired printed image or 3D point cloud using an image segmentation model or a 3D point cloud analysis model to calculate local size deviations and generate a multi-dimensional error heat map. Interlayer bond strength prediction sub-unit: The time-attention hybrid model is used to analyze the temperature history curve, cooling rate and extrusion pressure fluctuation during the material stacking process. The attention mechanism is used to capture parameter mutations and output the probability score of interlayer bond strength and the early warning of weak layer location. Structural Deformation Risk Assessment Subunit: Combining the viscoelastic constitutive equation of food materials to construct physical simulation constraints, and using a physics-inspired neural network to fuse real-time sensor data, the deformation trend of the printed structure under changes in gravity, thermal stress or environmental humidity is predicted.

2. The artificial intelligence-based food 3D printing control system as described in claim 1, characterized in that, The AI ​​instruction generation module includes: 3D model processing unit: Receives user-inputted 3D food model instructions, and identifies the model's geometric features, structural features, and material distribution requirements through feature extraction algorithms; the geometric features include curvature, wall thickness, and connectivity; the structural features include support requirements and layer complexity. Printing Model Generation Unit: Based on the parameters output by the 3D model processing unit, and combined with the food material 3D printing database, it generates a multi-material structure model adapted for 3D printing, including material zoning planning, support structure design, and printing path optimization. Intelligent slicing unit: The multi-material structure model is sliced ​​into layers to generate a set of G-code instructions including spatial coordinates, extrusion amount, speed, temperature, fill rate, and printing height parameters.

3. The artificial intelligence-based food 3D printing control system as described in claim 2, characterized in that, The feature extraction algorithm is PointNet++, MeshCNN, or MSFFN; the printing model generation unit stores a semantic segmentation network model, a material adaptation algorithm, a topology optimization algorithm, and a Markov decision process model; the intelligent slicing unit stores a Transformer encoder-decoder.

4. The artificial intelligence-based food 3D printing control system as described in claim 3, characterized in that, The 3D model processing unit stores an edge enhancement preprocessing algorithm and a multi-scale feature fusion network extraction algorithm; the intelligent slicing unit stores an adaptive layer thickness algorithm, a gated attention mechanism, and a gradient filling strategy.

5. The artificial intelligence-based food 3D printing control system as described in claim 1, characterized in that, The multimodal data fusion subunit stores a spatiotemporal graph convolutional network model, a federated learning algorithm, and a feature importance ranking algorithm; the image segmentation model adopts a neural network model based on an encoder-decoder structure; the 3D point cloud analysis model adopts a 3D point cloud processing model based on a region proposal network; the interlayer bonding strength prediction subunit stores a prediction model constructed by combining an LSTM network with a self-attention mechanism; and the structural deformation risk assessment subunit stores a physically inspired neural network and a multi-scale physical model with constraints embedded in the Maxwell / Kelvin-Voigt viscoelastic model.

6. The artificial intelligence-based food 3D printing control system as described in claim 1, characterized in that, The multimodal sensor unit includes: Camera or 3D scanner: Real-time acquisition of images of the printed layer surface and video streams of material flow; Infrared thermal imager: Real-time acquisition of temperature distribution and thermal gradient changes in each layer of the sample during the printing process; Mechanical sensors: including pressure sensors, strain sensors and vibration sensors, to collect extrusion pressure, interlayer stress and mechanical vibration data in real time; Environmental sensors: collect real-time data on humidity, airflow speed, and ambient temperature inside the printing chamber.

7. The artificial intelligence-based food 3D printing control system as described in claim 1, characterized in that, The defect detection and classification submodule includes: Lightweight convolutional neural network model: Real-time analysis of surface images to identify cracks, bubbles, layer shifts and material overflow; Long Short-Term Memory (LSTM) network or temporal convolutional network model: Analyze the temporal characteristics of sensor data and predict potential defects; Graph neural network or gated attention mechanism model: fuse multimodal data to generate comprehensive defect scores and classification labels.

8. The artificial intelligence-based food 3D printing control system as described in claim 1, characterized in that, The dynamic compensation engine includes a dynamic compensation strategy library for generating compensation strategies and a real-time control engine for driving the actuator to perform printing supplementation tasks; the compensation strategies include: Microscopic compensation strategy: For minor deviations, dynamically adjust nozzle movement speed, extrusion pressure, or material flow rate; the minor deviations include local temperature fluctuations; Meso-level compensation strategy: For moderate deviations, trigger local path replanning or fine-tune material ratios; the moderate deviations include poor interlayer bonding; Macro-level emergency strategy: For severe defects, suspend printing and generate a repair plan; the severe defects include structural collapse; The real-time control engine achieves low-latency compensation through edge computing nodes and hardware collaboration interfaces.

9. The artificial intelligence-based food 3D printing control system as described in claim 8, characterized in that, The micro-compensation strategy employs a deep deterministic strategy gradient algorithm to learn the optimal compensation parameters online; the macro-emergency strategy includes a macro-emergency scheme simulation based on digital twins; and the dynamic compensation strategy library stores interface enhancement compensation algorithms.

10. The artificial intelligence-based food 3D printing control system as described in claim 8, characterized in that, The edge computing node is a lightweight model deployed on the FPGA chip local to the printing device; the edge computing node establishes a direct communication link with the printing actuator.

11. The artificial intelligence-based food 3D printing control system as described in claim 1, characterized in that, The real-time error correction and compensation module also includes a feedback learning unit, which optimizes the generalization ability of the model in the defect detection and classification submodule through online incremental learning.

12. A method for 3D printing food based on a control system as described in any one of claims 1-11, characterized in that, The method includes the following steps: (1) Input and parse 3D model data to generate printing program instruction set: The user inputs food 3D model data, extracts geometric features, structural features and material distribution requirements through AI instruction generation module, and combines the rheological properties, thermal response curves and multi-material compatibility matrix in food 3D printing material database to generate a multi-material structure model adapted to 3D printing. After layering and slicing, a G code instruction set containing spatial coordinates, extrusion amount, temperature and filling rate parameters is generated. (2) Multimodal data fusion and accuracy prediction: The dynamic accuracy prediction module integrates material property parameters, historical printing data, real-time images and sensor data. Through the geometric accuracy prediction sub-unit, interlayer bonding strength prediction sub-unit and structural deformation risk assessment sub-unit, the accuracy assessment, strength prediction and risk warning of the printed product are output. (3) Real-time printing and defect detection: The multimodal sensor unit of the real-time error correction and compensation module acquires printing status data in real time, and the defect detection and classification submodule identifies defects and generates a comprehensive score; (4) Dynamic compensation and printing control: Based on the type and severity of the defect, the dynamic compensation engine calls the dynamic compensation strategy library to correct the printing deviation in real time.

13. The method as described in claim 12, characterized in that, The method includes: step (5), feedback learning and model optimization: the sensor data, defect labels and compensation effects during the error correction process are stored in the incremental dataset, and the model in the defect detection and classification submodule is trained online through the feedback learning unit.

14. The method as described in claim 12, characterized in that, The method includes: step (6), quality assessment and post-processing: after printing is completed, the dynamic accuracy prediction module performs offline quality assessment on the printed product and generates a quality assessment report; if the product has repairable defects, the post-processing strategy library is called according to the assessment report, and the post-processing execution mechanism is driven to perform repair operations.

Images