Food-grade microcapsule performance time sequence prediction method and system

Abstract

The invention relates to a food-grade microcapsule performance time sequence prediction method and system, and the method comprises the following steps: S1, experimental design and time sequence data set preparation: preparing a microcapsule sample, and forming a data set composed of static feature input and time sequence feature input; s2, a hybrid neural network model is constructed, and the model comprises S21, an attention enhancement capsule network module which sequentially comprises a primary capsule layer for mapping the static characteristic parameters into a plurality of primary capsule vectors; the high-level capsule layer is used for extracting a high-level capsule vector strongly related to the microcapsule performance from the primary capsule vector through a dynamic routing algorithm; a multi-head self-attention layer; s22, a recurrent neural network time sequence prediction module; s3, system training and verification; and S4, performing prediction application and visualization. Compared with the prior art, the method has the advantages of realizing dynamic prediction and full life cycle performance simulation, improving the explanatory and generalization ability of the model, ensuring the robustness and reliability of a prediction result and the like.

Description

Technical Field

[0001] This invention relates to the fields of microcapsule technology, smart food development and deep learning, and in particular to a method and system for predicting the performance of food-grade microcapsules over time. Background Technology

[0002] Current research and development of food-grade microcapsules mainly relies on numerous repetitive in vitro release experiments and sensory evaluations, which has the following drawbacks: First, it depends on empirical experiments, resulting in long experimental cycles, high time consumption, and high costs; second, it is impossible to predict in advance the flavor decay, nutrient degradation, or dynamic release behavior of microcapsules in the continuous oral-gastric-intestinal environment during storage; third, existing models are mostly static regressions or simple kinetic equations, which are difficult to handle nonlinear time-series changes under the coupling of multiple factors such as core-to-wall ratio, cross-linking degree, particle size, environmental pH, temperature, and enzymes; fourth, there is a lack of methods for fusing and modeling multimodal data, resulting in a low success rate of new formulations and seriously restricting the rapid development of functional foods, flavored beverages, and healthy snacks.

[0003] CN202510805198.9 discloses a method and system for optimizing mesoscale convection parameters based on genetic algorithms, relating to the field of meteorological forecasting technology. The method includes: modeling precipitation evolution using neural network differential equations, identifying dynamic system parameters using variational inference, extracting features using spatiotemporal heterogeneous graph neural networks and diffusion probability models, correcting parameter thresholds using physically guided neural networks, and optimizing parameters by constructing an optimization objective function using deep neural networks. This method can improve the accuracy of precipitation forecasts and reduce forecast errors. However, it involves multiple integrated models, resulting in high complexity. Many models in the scheme require training and optimization based on historical data from specific regions, leading to poor generalization ability. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method that enables dynamic prediction and full life cycle performance simulation, high-dimensional feature extraction and deep fusion of multimodal data, improves model interpretability and generalization ability, reduces R&D costs and cycle, and ensures the robustness and reliability of prediction results.

[0005] The objective of this invention can be achieved through the following technical solutions: This invention constructs an attention-enhanced capsule-recurrent neural network architecture. The recurrent neural network acts as the core processor, specifically designed to analyze the temporal dynamics of action molecule structural sequences or taste perception data, uncovering their inherent causal relationships. The capsule network, as a functional enhancement layer of the model, is more advanced than traditional analysis methods because it better understands how a molecule binds to the tongue's receptors in a three-dimensional manner. Simultaneously, by utilizing an attention-focusing mechanism, it ensures that the recurrent neural network focuses only on the most important structural features found by the capsule network.

[0006] The core mechanism of this invention is based on multimodal data and utilizes an attention-enhanced capsule network-RNN fusion architecture to achieve dynamic and accurate time-series prediction of the blood sugar-lowering and sweetening performance of microcapsules. First, multimodal data such as core material ratio, wall material type, and environmental variables are collected and standardized and segmented to construct a time-series dataset. Then, a multi-layer capsule network captures the geometric and functional features of the microcapsules, integrating a self-attention mechanism to assign weights to key features to enhance the focus on core information. Subsequently, the attention-enhanced capsule features are used as the initial state and input into the LSTM module of the hidden unit. Combined with previously measured time-series performance data (such as active substance release rate, flavor intensity, nutrient retention rate, and viable bacteria count), the performance evolution curves at multiple future time points are predicted. Finally, through the Adam optimizer, a composite loss function of MSE and KL divergence, and cross-validation, a closed-loop system of data preprocessing, high-dimensional feature extraction, time-series modeling, and validation output is formed. This overcomes the limitations of traditional static models and efficiently supports the rapid development and formulation optimization of various functional foods such as flavor microcapsules, nutritional microcapsules, probiotic microcapsules, and low-sodium / low-fat microcapsules.

[0007] This invention provides a method for time-series prediction of the performance of food-grade microcapsules, comprising the following steps: S1. Experimental Design and Time-Series Dataset Preparation: Microcapsule samples with different core material ratios and wall material types were prepared and tested under different combinations of environmental variables. Performance index data of the microcapsules were continuously collected at fixed time intervals to form a multimodal time-series dataset. The collected data were preprocessed and sequence constructed to form a dataset composed of static feature inputs and time-series feature inputs. Furthermore, the performance index data includes at least one of the following: active substance release curve, flavor intensity, nutrient retention rate, viable bacteria count, saltiness perception, and sweetness perception. The static feature input includes microcapsule preparation parameters, which include the core-to-wall material mass ratio, wall material molecular weight, crosslinking agent concentration, emulsifier HLB value, and core material particle size, totaling 38 dimensions; the temporal feature input is a sequence of performance indexes such as release rate, sweetness residual rate, or simulated blood glucose elevation rate at the first t time points. Specifically, based on the set range of core material ratio and wall material type, a series of experimental groups were designed, and microcapsule samples were prepared using standard microcapsule preparation processes. Then, multimodal time-series data were collected: each group of prepared microcapsule samples was tested under different combinations of environmental variables (e.g., at the same temperature, simulating gastric acid and intestinal environments with different pH values). Under these conditions, the rate of change in glycemic index or the release rate of sweetener in the in vitro experiment was recorded every 5 minutes, continuously for, for example, 300 minutes, continuously collecting performance index data. Then, data preprocessing and sequence construction were performed: the core material ratio, wall material type, and environmental variables of each experimental group were organized into numerical vectors, and the collected release rate and glycemic index change rate were Z-score standardized to eliminate dimensional differences. The collected long-term time-series data was divided into fixed-length time steps, and the final dataset consisted of two parts (static feature input and temporal feature input), and then target prediction was performed (multiple time points could be set: for example, how the sweetness is maintained after 30 minutes, how the glycemic release rate curve changes; and how it remains the same after 50 minutes).

[0008] S2. Construct a hybrid neural network model, which includes: S21, Attention-enhancing capsule network module, used to process the static feature input, comprising: The primary capsule layer maps static feature parameters into multiple primary capsule vectors; for example, setting up 8 capsules (wall material group, core material group, preparation environment group, emulsification process group, etc.), each capsule has 8 dimensions (proportion of wall material, amount of crosslinking agent, concentration, molecular weight, viscosity, degradation rate, etc.).

[0009] In the advanced capsule layer, a dynamic routing algorithm extracts advanced capsule vectors that are strongly correlated with microcapsule performance from the primary capsule vectors. Further, the dynamic routing algorithm determines the contribution of the primary capsule vectors to the advanced capsule vectors through iterative calculation. The dynamic routing algorithm uses univariate experiments, adjusting only the parameters of one primary capsule group at a time while keeping others constant, to observe the tightness of the core-shell bonding. After several iterations, it becomes clear which parameters of the primary capsule groups contribute significantly to the advanced capsule groups.

[0010] Specifically, relevant parameter combinations were selected from the eight capsule groups to form four advanced capsule groups (core-shell interaction group, sustained-release rate group, etc.). Each advanced capsule group contains 16 dimensions (such as the 16 attributes in the core-shell interaction group, including wall thickness, crosslinking agent dosage, core particle size, core solubility, emulsifier type, and emulsion strength). These four advanced capsule layers are combinations of core features extracted from scattered data that are strongly correlated with the actual performance of microcapsules.

[0011] The multi-head self-attention layer weights and focuses the advanced capsule vector to output a refined high-dimensional core feature vector V-capsule. Furthermore, the multi-head self-attention layer assigns weights to key parameter combinations by focusing on different core requirements such as core-shell compactness, slow-release rate, and wall material durability. Specifically, the multi-head self-attention layer focuses on different core requirements (such as core-shell compactness, sustained-release rate, and wall material durability) to highlight key parameter combinations of advanced capsule groups. For example, in the core-shell interaction group, it examines the cross-linking agent dosage and core material particle size. If it finds that a 0.8% glutaraldehyde-5μm core material is the most compact, then this combination is marked as the optimal parameter weight. After passing through the multi-head self-attention layer, a refined high-dimensional core feature vector V-capsule is finally output.

[0012] S22. The recurrent neural network time series prediction module adopts a long short-term memory network LSTM. The LSTM uses the V-capsule vector as the initial hidden state and inputs the time series features to learn the time series pattern of performance changes and output the performance prediction values ​​for multiple future time points. Furthermore, the LSTM module is a two-layer bidirectional LSTM with 128 and 100 hidden units respectively, and a dropout rate of 0.2. The LSTM module outputs predicted values ​​for the next n time points, with each time point corresponding to the output of a neuron. Specifically, the recurrent neural network (RNN) time-series prediction module uses historical data on the release rates of microcapsules to predict their future release rates. The first step involves selecting an LSTM (Long Short-Term Memory) layer. While ordinary RNNs may only remember recent release patterns and forget patterns from earlier periods, LSTMs can simultaneously remember early, mid-term, and recent patterns, resulting in highly accurate predictions. The second step involves the LSTM hidden unit layer, with 100 hidden units. Their core function is to encode the historical time-series data of microcapsule release rates using multi-scale, fine-grained temporal features. Each hidden unit corresponds to a specific temporal pattern capture capability (e.g., the first hidden unit records the release rate increase pattern of the previous 200 minutes, such as an increase of 2% or 1.8% every 10 minutes). The third step is the fusion mechanism: based on the V-capsule feature vector extracted from the attention capsule module, the LSTM analyzes historical release data with clear cognitive and judgment criteria, making predictions less prone to deviation. The fourth step is the input layer: 100 input layers are set up. The LSTM examines the release history of the microcapsules over the past 100 time points (t1-t100) to identify release patterns in order to predict future release rates. The fifth step is the output layer: 10 neurons are set up. After reviewing the historical release data of the microcapsules and combining it with the 100 hidden units, the LSTM provides a predicted release rate for the next 10 time points. For example, neuron 1 predicts a release rate of 62% at t101 (minute 1010). Each neuron corresponds to a predicted value for a future time point.

[0013] S3. System Training and Validation: The dataset is divided into training, validation, and test sets. The hybrid neural network model is trained and validated using a composite loss function and optimizer to obtain the optimal model. Specifically, during model training, the total dataset is first randomly divided into 70% training set, 15% validation set, and 15% test set. The Adam optimizer is used with an initial learning rate of 0.001. The composite loss function is a combination of mean squared error (MSE) and KL divergence. The training is planned for 100-500 rounds. During training, the training set data is input into the model, and the network weights are updated through direction propagation of the total loss. At the end of each round, the MAE and RMSE are evaluated using the validation set. If these two metrics no longer decrease or overfitting occurs, training is stopped or hyperparameters are adjusted. Finally, a 5-fold cross-validation is performed on the training set, and the robustness of the model is ensured by averaging the performance of different data subsets.

[0014] S4. Predictive Application and Visualization: For new microcapsule formulations, static parameters and initial time-series performance data are input into the optimal model to predict the performance evolution curves at future time points and then visualize them.

[0015] Specifically, based on the trained optimal model, combined with the input new formulation parameters and initial performance data, the CapsNet and recurrent neural network modules collaborate to achieve time-series prediction of future performance. The steps are as follows: First, the model weights that have been trained and performed best on the test set are used to ensure the reliability of the prediction. Then, static parameters such as the proportion of raw materials in the new formulation, temperature, and pH value (these are fundamental data representing the characteristics of the microcapsules themselves) are provided. Next, initial performance data such as sweetness and release rate at the first 50 time points are provided, which are initial time-series information reflecting the performance change trend. The model processes the input data through the collaborative work of two modules. The CapsNet module specifically processes the static formulation parameters, extracting vectors that characterize the core features of the microcapsules. The recurrent neural network module uses this feature vector as its initial state, combined with the performance data from the first 50 time points, to learn the time-series pattern of performance changes. Based on the learned pattern, the RNN module outputs the performance prediction value for a specific future time period (t=51 to t=60).

[0016] Results visualization: The data from the model's predicted 10 future time points are concatenated with the input data from 50 time points to generate a time-series curve; a sweetness maintenance time curve is plotted, with time on the horizontal axis and predicted sweetness / release rate on the vertical axis; a blood glucose response graph is plotted, with time on the horizontal axis and the predicted rate of change of blood glucose index on the vertical axis; finally, a comparative analysis is performed, plotting the predicted curve and the available real experimental data curve on the same graph to intuitively assess the accuracy of the prediction.

[0017] This invention also provides a time-series prediction system for the performance of food-grade microcapsules, comprising: The data preparation module is used to perform step S1, which involves preparing and preprocessing the time series dataset. The model building and training module is used to perform steps S2 and S3 to build, train, and validate the hybrid neural network model; The prediction application module is used to perform step S4, load the optimal model, and perform performance prediction and visualization output for the new formula.

[0018] Furthermore, the time-series curves generated by the prediction application module include sweetness maintenance time curves and blood glucose response graphs, and the predicted curves are compared and analyzed with the actual experimental data curves. Compared with the prior art, the present invention has the following advantages: (1) Achieving dynamic prediction and full life-cycle performance simulation. Traditional models are mostly static regressions or simple dynamic equations, which cannot effectively handle nonlinear temporal changes under the coupling of multiple factors. This invention achieves dynamic temporal prediction of the core performance of food-grade microcapsules throughout the entire process of processing, storage and digestion by integrating attention-enhanced capsule network (CapsNet) and recurrent neural network (RNN, specifically LSTM). Only the initial parameters of the new formula and a small amount of early experimental data are required to input the model, which can accurately predict the performance evolution curves at multiple future time points, such as the release rate of active substances, flavor intensity, and nutrient retention rate.

[0019] (2) High-dimensional feature extraction and deep fusion of multimodal data enhance model interpretability and generalization ability. Attention-enhanced capsule networks are used to extract high-dimensional geometric and functional features of microcapsule preparation parameters. Through dynamic routing algorithms, capsule networks automatically aggregate scattered process parameters (such as core-to-wall ratio, crosslinking agent concentration, particle size, etc.) into "high-level feature combinations" with clear physical meaning (such as core-shell interaction, sustained-release kinetic control), which can better understand the complex three-dimensional relationships between parameters than traditional methods. The multi-head self-attention mechanism further assigns weights to these key features, enabling the model to focus on the core parameter combinations that affect performance, thus enhancing the interpretability of the model.

[0020] (3) Reduced R&D costs and time. Only a small number of initial experiments are needed to obtain time-series data, which can predict long-term behavior and avoid lengthy full-cycle testing. The model can quickly evaluate the potential performance of new formulations, screen out the optimal parameter combination, and significantly improve the success rate of new formulations. It can predict in advance the flavor decay of microcapsules during storage or their dynamic release behavior in the gastrointestinal environment.

[0021] (4) Ensuring the robustness and reliability of prediction results. A complete closed-loop system from data preparation and model building to training and validation was constructed to ensure the practicality of the technology. Multimodal time-series data were collected through standardized experimental design and preprocessed to eliminate dimensional differences. Bidirectional LSTM was used to capture richer information, and a composite loss function (MSE+KL divergence) was combined to simultaneously optimize prediction accuracy and distribution consistency. Overfitting was prevented through validation set monitoring and early stopping, and the robustness of the model was evaluated using 5-fold cross-validation to ensure stable and reliable prediction results. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the architecture of a hybrid neural network model; Figure 2 This is a schematic diagram of the microcapsule preparation and multimodal time-series data acquisition process in Example 2; Figure 3 Example 2: Key feature extraction and attention heatmap of attention-enhancing capsule network; Figure 4 This is a schematic diagram illustrating the time-series performance prediction and real-world experimental verification of the new formulation in Example 2. Figure 5 This is a schematic diagram illustrating the prediction of the survival rate of live bacteria in the gastrointestinal environment and the actual in vitro / in vivo verification of probiotic microcapsules in Example 2. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0024] Example 1 This embodiment provides a method for time-series prediction of the performance of food-grade microcapsules, including the following steps: S1. Experimental Design and Time-Series Dataset Preparation: Microcapsule samples with different core material ratios and wall material types were prepared and tested under different combinations of environmental variables. Performance index data of the microcapsules were continuously collected at fixed time intervals to form a multimodal time-series dataset. The collected data were preprocessed and sequence constructed to form a dataset composed of static feature inputs and time-series feature inputs. In a specific implementation, the performance index data includes at least one of the following: active substance release curve, flavor intensity, nutrient retention rate, viable bacteria count, saltiness perception, and sweetness perception. The static feature input includes microcapsule preparation parameters, which include the core-to-wall material mass ratio, wall material molecular weight, crosslinking agent concentration, emulsifier HLB value, and core material particle size, totaling 38 dimensions. The temporal feature input is a sequence of performance indexes such as release rate, sweetness retention rate, or simulated blood glucose elevation rate at the first t time points. Specifically, based on the set range of core material ratio and wall material type, a series of experimental groups were designed, and microcapsule samples were prepared using standard microcapsule preparation processes. Then, multimodal time-series data were collected: each group of prepared microcapsule samples was tested under different combinations of environmental variables (e.g., at the same temperature, simulating gastric acid and intestinal environments with different pH values). Under these conditions, the rate of change in glycemic index or the release rate of sweetener in the in vitro experiment was recorded every 5 minutes, continuously for, for example, 300 minutes, continuously collecting performance index data. Then, data preprocessing and sequence construction were performed: the core material ratio, wall material type, and environmental variables of each experimental group were organized into numerical vectors, and the collected release rate and glycemic index change rate were Z-score standardized to eliminate dimensional differences. The collected long-term time-series data was divided into fixed-length time steps, and the final dataset consisted of two parts (static feature input and temporal feature input), and then target prediction was performed (multiple time points could be set: for example, how the sweetness is maintained after 30 minutes, how the glycemic release rate curve changes; and how it remains the same after 50 minutes).

[0025] S2. Construct a hybrid neural network model, which includes: S21, Attention-enhancing capsule network module, used to process the static feature input, comprising: The primary capsule layer maps static feature parameters into multiple primary capsule vectors; for example, setting up 8 capsules (wall material group, core material group, preparation environment group, emulsification process group, etc.), each capsule has 8 dimensions (proportion of wall material, amount of crosslinking agent, concentration, molecular weight, viscosity, degradation rate, etc.).

[0026] In the advanced capsule layer, a dynamic routing algorithm extracts advanced capsule vectors that are strongly correlated with microcapsule performance from the primary capsule vectors. In a specific implementation, the dynamic routing algorithm determines the contribution of the primary capsule vectors to the advanced capsule vectors through iterative calculation. The dynamic routing algorithm uses univariate experiments, adjusting only the parameters of one primary capsule group at a time while keeping others constant, to observe the tightness of the core-shell bonding. After several iterations, it becomes clear which parameters of the primary capsule groups contribute significantly to the advanced capsule groups.

[0027] Specifically, relevant parameter combinations were selected from the eight capsule groups to form four advanced capsule groups (core-shell interaction group, sustained-release rate group, etc.). Each advanced capsule group contains 16 dimensions (such as the 16 attributes in the core-shell interaction group, including wall thickness, crosslinking agent dosage, core particle size, core solubility, emulsifier type, and emulsion strength). These four advanced capsule layers are combinations of core features extracted from scattered data that are strongly correlated with the actual performance of microcapsules.

[0028] A multi-head self-attention layer weights and focuses the advanced capsule vectors to output a refined high-dimensional core feature vector V-capsule. In a specific implementation, the multi-head self-attention layer assigns weights to key parameter combinations by focusing on different core requirements among core-shell compactness, slow-release rate, and wall material durability. Specifically, the multi-head self-attention layer focuses on different core requirements (such as core-shell compactness, sustained-release rate, and wall material durability) to highlight key parameter combinations of advanced capsule groups. For example, in the core-shell interaction group, it examines the cross-linking agent dosage and core material particle size. If it finds that a 0.8% glutaraldehyde-5μm core material is the most compact, then this combination is marked as the optimal parameter weight. After passing through the multi-head self-attention layer, a refined high-dimensional core feature vector V-capsule is finally output.

[0029] S22. The recurrent neural network time series prediction module adopts a long short-term memory network LSTM. The LSTM uses the V-capsule vector as the initial hidden state and inputs the time series features to learn the time series pattern of performance changes and output the performance prediction values ​​for multiple future time points. In a specific implementation, the LSTM module is a two-layer bidirectional LSTM with 128 and 100 hidden units, respectively, and a dropout rate of 0.2. The LSTM module outputs predicted values ​​for the next n time points, with each time point corresponding to the output of a neuron. Specifically, the recurrent neural network (RNN) time-series prediction module uses historical data on the release rates of microcapsules to predict their future release rates. The first step involves selecting an LSTM (Long Short-Term Memory) layer. While ordinary RNNs may only remember recent release patterns and forget patterns from earlier periods, LSTMs can simultaneously remember early, mid-term, and recent patterns, resulting in highly accurate predictions. The second step involves the LSTM hidden unit layer, with 100 hidden units. Their core function is to encode the historical time-series data of microcapsule release rates using multi-scale, fine-grained temporal features. Each hidden unit corresponds to a specific temporal pattern capture capability (e.g., the first hidden unit records the release rate increase pattern of the previous 200 minutes, such as an increase of 2% or 1.8% every 10 minutes). The third step is the fusion mechanism: based on the V-capsule feature vector extracted from the attention capsule module, the LSTM analyzes historical release data with clear cognitive and judgment criteria, making predictions less prone to deviation. The fourth step is the input layer: 100 input layers are set up. The LSTM examines the release history of the microcapsules over the past 100 time points (t1-t100) to identify release patterns in order to predict future release rates. The fifth step is the output layer: 10 neurons are set up. After reviewing the historical release data of the microcapsules and combining it with the 100 hidden units, the LSTM provides a predicted release rate for the next 10 time points. For example, neuron 1 predicts a release rate of 62% at t101 (minute 1010). Each neuron corresponds to a predicted value for a future time point.

[0030] S3. System Training and Validation: The dataset is divided into training, validation, and test sets. The hybrid neural network model is trained and validated using a composite loss function and optimizer to obtain the optimal model. Specifically, during model training, the total dataset is first randomly divided into 70% training set, 15% validation set, and 15% test set. The Adam optimizer is used with an initial learning rate of 0.001. The composite loss function is a combination of mean squared error (MSE) and KL divergence. The training is planned for 100-500 rounds. During training, the training set data is input into the model, and the network weights are updated through direction propagation of the total loss. At the end of each round, the MAE and RMSE are evaluated using the validation set. If these two metrics no longer decrease or overfitting occurs, training is stopped or hyperparameters are adjusted. Finally, a 5-fold cross-validation is performed on the training set, and the robustness of the model is ensured by averaging the performance of different data subsets.

[0031] S4. Predictive Application and Visualization: For new microcapsule formulations, static parameters and initial time-series performance data are input into the optimal model to predict the performance evolution curves at future time points and then visualize them.

[0032] Specifically, based on the trained optimal model, combined with the input new formulation parameters and initial performance data, the CapsNet and recurrent neural network modules collaborate to achieve time-series prediction of future performance. The steps are as follows: First, the model weights that have been trained and performed best on the test set are used to ensure the reliability of the prediction. Then, static parameters such as the proportion of raw materials in the new formulation, temperature, and pH value (these are fundamental data representing the characteristics of the microcapsules themselves) are provided. Next, initial performance data such as sweetness and release rate at the first 50 time points are provided, which are initial time-series information reflecting the performance change trend. The model processes the input data through the collaborative work of two modules. The CapsNet module specifically processes the static formulation parameters, extracting vectors that characterize the core features of the microcapsules. The recurrent neural network module uses this feature vector as its initial state, combined with the performance data from the first 50 time points, to learn the time-series pattern of performance changes. Based on the learned pattern, the RNN module outputs the performance prediction value for a specific future time period (t=51 to t=60).

[0033] Results visualization: The data from the model's predicted 10 future time points are concatenated with the input data from 50 time points to generate a time-series curve; a sweetness maintenance time curve is plotted, with time on the horizontal axis and predicted sweetness / release rate on the vertical axis; a blood glucose response graph is plotted, with time on the horizontal axis and the predicted rate of change of blood glucose index on the vertical axis; finally, a comparative analysis is performed, plotting the predicted curve and the available real experimental data curve on the same graph to intuitively assess the accuracy of the prediction.

[0034] This embodiment also provides a time-series prediction system for the performance of food-grade microcapsules, including: The data preparation module is used to perform step S1, which involves preparing and preprocessing the time series dataset. The model building and training module is used to perform steps S2 and S3 to build, train, and validate the hybrid neural network model; The prediction application module is used to perform step S4, load the optimal model, and perform performance prediction and visualization output for the new formula.

[0035] In a specific implementation, the time-series curves generated by the prediction application module include sweetness maintenance time curves and blood glucose response graphs, and the predicted curves are compared and analyzed with the actual experimental data curves. Example 2 This embodiment provides a method for time-series prediction of the performance of food-grade microcapsules, including the following steps: S1. Experimental Design and Time-Series Dataset Preparation: Microcapsule samples with different core material ratios and wall material types were prepared and tested under different combinations of environmental variables. Performance index data of the microcapsules were continuously collected at fixed time intervals to form a multimodal time-series dataset. The collected data were preprocessed and sequence constructed to form a dataset composed of static feature inputs and time-series feature inputs. In a specific implementation, the performance index data includes at least one of the following: active substance release curve, flavor intensity, nutrient retention rate, viable bacteria count, saltiness perception, and sweetness perception. The static feature input includes microcapsule preparation parameters, which include the core-to-wall material mass ratio, wall material molecular weight, crosslinking agent concentration, emulsifier HLB value, and core material particle size, totaling 38 dimensions, which are input in vector form after normalization. The static feature input is then fed into the attention-enhanced capsule network module.

[0036] The time-series feature input is a sequence of performance indicators such as release rate, sweetness residue, or simulated blood glucose rise rate over the first t time points. The time-series feature input is a sequence of performance indicators such as release rate, sweetness residue, and simulated blood glucose rise rate over the first t time points (t=50 or 100 in this embodiment), with one data point collected every 5 minutes.

[0037] like Figure 2 As shown, the data collection was divided into a preparation section on the left and a data acquisition section on the right. The preparation section on the left used a spray-drying-coagulation method to prepare bifunctional microcapsules containing sweeteners and hypoglycemic agents. Sodium alginate-chitosan was used as the wall material, and aspartame + chromium yeast was used as the core material. The core-to-wall ratio was set at four levels: 1:5, 1:10, 1:15, and 1:20. The concentrations of the cross-linking agent glutaraldehyde were 0.5%, 0.8%, 1.2%, and 1.6%, respectively. A total of 4×4=16 orthogonal experiments were conducted, with 100g of microcapsule powder prepared for each group, and the particle size controlled between 50 and 200 μm. In the data acquisition section on the right, the 16 groups of microcapsules were placed in a simulated gastrointestinal dynamic release system: 0–120 min was simulated gastric juice (pH=1.2, 37℃), and 120–300 min was switched to simulated intestinal juice (pH=6.8, containing 0.5% pancreatic enzymes). Samples were taken every 5 minutes, and the concentration of aspartame released into the medium was determined using high-performance liquid chromatography (HPLC). Simultaneously, the sweetness of the medium was measured using an electronic tongue, and the simulated glycemic inhibition rate (obtained through an α-amylase / α-glucosidase inhibition experiment) was recorded. A typical release curve example: In the group with a core-to-wall ratio of 1:15 and 0.8% glutaraldehyde, <8% was released from gastric fluid within 120 minutes, and rapidly increased to 75% within 30 minutes after entering intestinal fluid, with the sweetness peak occurring in the 150–100 μmol / L range.

[0038] S2. Construct a hybrid neural network model, such as Figure 1 As shown, the model includes: S21, Attention-enhancing capsule network module, used to process the static feature input, comprising: The primary capsule layer maps static feature parameters into multiple primary capsule vectors; In the advanced capsule layer, a dynamic routing algorithm extracts advanced capsule vectors that are strongly correlated with microcapsule performance from the primary capsule vectors. In a specific implementation, the dynamic routing algorithm determines the contribution of the primary capsule vectors to the advanced capsule vectors through iterative calculation. The dynamic routing algorithm uses univariate experiments, adjusting only the parameters of one primary capsule group at a time while keeping others constant, to observe the tightness of the core-shell bonding. After several iterations, it becomes clear which parameters of the primary capsule groups contribute significantly to the advanced capsule groups.

[0039] A multi-head self-attention layer weights and focuses the advanced capsule vectors to output a refined high-dimensional core feature vector V-capsule. In a specific implementation, the multi-head self-attention layer assigns weights to key parameter combinations by focusing on different core requirements among core-shell compactness, slow-release rate, and wall material durability. S22. The recurrent neural network time series prediction module adopts a long short-term memory network LSTM. The LSTM uses the V-capsule vector as the initial hidden state and inputs the time series features to learn the time series pattern of performance changes and output the performance prediction values ​​for multiple future time points. The V-capsule, serving as the initial hidden states h0 and c0, is fed into a two-layer LSTM network along with temporal features. The first LSTM layer has 128 hidden units, and the second layer has 100, with dropout=0.2. A bidirectional structure is used to capture contextual information. The LSTM output, after concatenating the last hidden state with the V-capsule, is then connected to a fully connected layer. Finally, the output layer consists of 10 neurons, predicting the sweetener release rate or glycemic inhibition rate at time points 51–60 (or 101–110). Figure 1 The loss function is given below: L = MSE + 0.1×KL divergence, which is used to simultaneously ensure prediction accuracy and distribution consistency.

[0040] Specifically, a set of typical microcapsule preparation parameters (38 dimensions in total, normalized) are input into the attention-enhanced capsule network. These parameters include, but are not limited to: core-to-wall ratio 1:15, sodium alginate molecular weight 180 kDa, chitosan deacetylation degree 92%, glutaraldehyde crosslinking agent concentration 0.8 wt%, average core particle size 5 μm, emulsification speed 9000 rpm, emulsification time 15 min, and spray drying inlet temperature 185℃. First, the 38-dimensional static features are reshaped into 8 primary capsules after passing through two fully connected layers. Each capsule is an 8-dimensional vector, corresponding to 8 process-property subspaces: wall material composition, core material properties, crosslinking process, emulsification process, drying process, particle size control, pH responsiveness, and thermal stability. Then, a 3-round dynamic routing algorithm routes the 8 primary capsules to 4 advanced capsules, named core-shell interaction, sustained-release kinetic control, wall material acid resistance, and particle uniformity, respectively. After routing, the coupling coefficient matrix c_ij (8×4) is as follows: Figure 3 The heatmap on the left shows the output vector magnitudes (squashed) of the four advanced capsules after route convergence: 0.95, 0.92, 0.88, and 0.79, respectively. The "core-shell interaction" and "slow-release kinetics control" advanced capsules were the most strongly activated (magnitude > 0.92), indicating that the model automatically identified the core performance regulation mechanism of the formulation as focusing on the cross-linking strength and diffusion channel regulation at the core-shell interface. The four 16-dimensional advanced capsules were input into a four-head self-attention module. The attention weight matrices for each head are shown below. Figure 3 As shown on the right: Head 1 primarily captures the synergistic relationship between core-shell interaction and controlled release kinetics; Head 2 and Head 4 focus on the weak correlation between particle uniformity and other features, respectively. Head 3 focuses on the regulatory effect of wall material acid resistance on overall release behavior. After multi-head self-attention weighted fusion, a refined 128-dimensional feature vector V_capsule is output. Subsequently, V_capsule is used as the initial hidden states h0 and c0 of the LSTM for subsequent temporal prediction.

[0041] S3. System Training and Validation: The dataset is divided into training, validation, and test sets. The hybrid neural network model is trained and validated using a composite loss function and optimizer to obtain the optimal model. Specifically, during model training, the total dataset is first randomly divided into 70% training set, 15% validation set, and 15% test set. The Adam optimizer is used with an initial learning rate of 0.001. The composite loss function is a combination of mean squared error (MSE) and KL divergence. The training is planned for 100-500 rounds. During training, the training set data is input into the model, and the network weights are updated through direction propagation of the total loss. At the end of each round, the MAE and RMSE are evaluated using the validation set. If these two metrics no longer decrease or overfitting occurs, training is stopped or hyperparameters are adjusted. Finally, a 5-fold cross-validation is performed on the training set, and the robustness of the model is ensured by averaging the performance of different data subsets.

[0042] S4. Predictive Application and Visualization: For new microcapsule formulations, static parameters and initial time-series performance data are input into the optimal model to predict the performance evolution curves at future time points and then visualize them.

[0043] In this embodiment, a pre-trained optimal model is used to predict the time-series performance of a novel formulation, and the results are compared with actual in vitro release experiments. New formulation parameters: core-to-wall ratio 1:12, sodium alginate molecular weight 180 kDa, glutaraldehyde 1.0%, core material particle size 8 μm. The model is input with measured release data from the first 50 time points (0–245 min), and the model predicts the release rate of the active substance from time points 51–60 (250–295 min). Figure 4 As shown in the figure, the solid blue line represents the release rate measured by actual HPLC at the subsequent 10 time points, and the dashed orange line represents the model prediction. The MAE at the 10 time points was only 1.64%, with the largest error occurring at the 58th time point (285 min), an error of 2.8%. The predicted curve almost completely overlaps with the actual curve, especially accurately capturing the inflection point of accelerated release near 290 min caused by local disintegration of the wall material.

[0044] like Figure 5 The diagram shows a comparison between the blood glucose response curve predicted by the model in this embodiment after ingestion of microcapsule-containing gel candy and the actual measurement results of an oral glucose tolerance test (OGTT) in 10 healthy volunteers. The subjects' fasting blood glucose was (4.6 ± 0.3) mmol / L. They ingested gel candy containing 50g of usable carbohydrates, and blood glucose was measured at 0, 15, 30, 45, 60, 90, 120, and 180 minutes. Model input: all preparation parameters of this batch of microcapsules + in vitro release data for the first 30 minutes, predicting the complete blood glucose curve over 180 minutes in vivo (green dashed line). Actual measured average blood glucose curve for 10 participants (red solid line ± standard deviation shading).

[0045] The results showed that the model predicted a peak blood glucose level of 6.8 mmol / L at 47 minutes; the actual average peak occurred at 7.0 mmol / L at 49 minutes. The predicted blood glucose regression value at 120 minutes was 5.3 mmol / L, while the actual value was 5.5 mmol / L. The predicted area under the curve (AUC) was approximately 15%-20% lower than that of the regular sucrose group, while the actual reduction was approximately 10%-15%.

[0046] This embodiment also provides a time-series prediction system for the performance of food-grade microcapsules, including: The data preparation module is used to perform step S1, which involves preparing and preprocessing the time series dataset. The model building and training module is used to perform steps S2 and S3 to build, train, and validate the hybrid neural network model; The prediction application module is used to perform step S4, load the optimal model, and perform performance prediction and visualization output for the new formula.

[0047] In a specific implementation, the time-series curves generated by the prediction application module include sweetness maintenance time curves and blood glucose response graphs, and the predicted curves are compared and analyzed with the actual experimental data curves. Components not described in detail in this embodiment are all existing components that can be purchased through public channels.

[0048] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A method for time-series prediction of the performance of food-grade microcapsules, characterized in that, Includes the following steps: S1. Experimental Design and Time-Series Dataset Preparation: Microcapsule samples with different core material ratios and wall material types were prepared and tested under different combinations of environmental variables. Performance index data of the microcapsules were continuously collected at fixed time intervals to form a multimodal time-series dataset. The collected data were preprocessed and sequence constructed to form a dataset composed of static feature inputs and time-series feature inputs. S2. Construct a hybrid neural network model, which includes: S21, Attention-enhancing capsule network module, used to process the static feature input, comprising: The primary capsule layer maps static feature parameters into multiple primary capsule vectors; The advanced capsule layer extracts advanced capsule vectors that are strongly correlated with microcapsule performance from the primary capsule vectors using a dynamic routing algorithm; A multi-head self-attention layer weights and focuses the high-level capsule vectors, outputting a refined high-dimensional core feature vector V-capsule; S22. The recurrent neural network time series prediction module adopts a long short-term memory network LSTM. The LSTM uses the V-capsule vector as the initial hidden state and inputs the time series features to learn the time series pattern of performance changes and output the performance prediction values ​​for multiple future time points. S3. System Training and Validation: Divide the dataset into training set, validation set and test set, and use a composite loss function and optimizer to train and validate the hybrid neural network model to obtain the optimal model; S4. Predictive Application and Visualization: For new microcapsule formulations, static parameters and initial time-series performance data are input into the optimal model to predict the performance evolution curves at future time points and then visualize them.

2. The method for time-series prediction of the performance of food-grade microcapsules according to claim 1, characterized in that, In S1, the performance index data includes at least one of the following: active substance release curve, flavor intensity, nutrient retention rate, viable bacteria count, saltiness perception, and sweetness perception.

3. The method for time-series prediction of the performance of food-grade microcapsules according to claim 1, characterized in that, In S1, the static feature input includes microcapsule preparation parameters, which include the mass ratio of core material to wall material, molecular weight of wall material, concentration of crosslinking agent, HLB value of emulsifier, and particle size of core material; the temporal feature input is the release rate, sweetness residual rate, or simulated blood glucose rise rate at the first t time points.

4. The method for time-series prediction of the performance of food-grade microcapsules according to claim 1, characterized in that, In S21, the dynamic routing algorithm determines the contribution of the primary capsule vector to the advanced capsule vector through iterative calculation.

5. The method for time-series prediction of the performance of food-grade microcapsules according to claim 1, characterized in that, In S21, the multi-head self-attention layer assigns weights to key parameter combinations by focusing on different core needs.

6. The method for time-series prediction of the performance of food-grade microcapsules according to claim 1, characterized in that, In S22, the LSTM module is a two-layer bidirectional LSTM with 128 and 100 hidden units respectively, and a dropout rate of 0.

2.

7. The method for time-series prediction of the performance of food-grade microcapsules according to claim 1, characterized in that, In S22, the LSTM module outputs predicted values ​​for the next n time points, with each time point corresponding to the output of a neuron.

8. The method for time-series prediction of the performance of food-grade microcapsules according to claim 1, characterized in that, In S3, the composite loss function is a combination of mean square error (MSE) and KL divergence.

9. A time-series prediction system for the performance of food-grade microcapsules implementing the prediction method as described in any one of claims 1-8, characterized in that, include: The data preparation module is used to perform step S1, which involves preparing and preprocessing the time series dataset. The model building and training module is used to perform steps S2 and S3 to build, train, and validate the hybrid neural network model; The prediction application module is used to perform step S4, load the optimal model, and perform performance prediction and visualization output for the new formula.

10. The time-series performance prediction system for food-grade microcapsules according to claim 9, characterized in that, The prediction application module generates time-series curves including sweetness maintenance time curves and blood glucose response graphs, and compares and analyzes the predicted curves with the actual experimental data curves.

Images