Dynamic proportioning system for functional components of soft sweets based on deep learning

Abstract

The invention relates to the field of dynamic proportioning, and particularly discloses a soft sweet functional component dynamic proportioning system based on deep learning, which comprises the following steps: firstly, acquiring a first component and a dosage thereof and a second component and a dosage thereof, and carrying out embedded coding on the two components to obtain embedded coding vectors of the first component and the second component; and performing normalization processing on each dose to obtain a normalized dose. Next, the embedded coding vector of each component is fused with the normalized dose to generate first and second component dose embedded coding vectors. Then, the vectors are processed through multi-scale progressive feature interaction, and inter-component multi-scale progressive perception fusion coding vectors are obtained; and predicting a synergistic interaction score based on the vector, comparing the synergistic interaction score with a target score, and judging whether the component dynamic ratio needs to be adjusted or not. The process ensures that the synergistic effect among different components is accurately identified and optimized, and the effectiveness and accuracy of the whole formula are improved.

Description

Technical Field

[0001] The present application relates to the field of dynamic proportioning, and more specifically, to a dynamic proportioning system of soft candy functional ingredients based on deep learning. Background Art

[0002] As a popular food carrier, gummies are widely used in functional foods due to their pleasant taste and convenience, delivering various functional ingredients such as vitamins, minerals, and plant extracts. Consumer demand for functional gummies is growing, and consumers are not only interested in the efficacy of a single ingredient, but also in the combined effects of multiple functional ingredients, particularly synergistic effects, where the combination of different ingredients in a specific ratio produces an effect greater than the sum of the individual effects. Achieving the optimal synergistic ratio of functional ingredients in gummies is key to enhancing product efficacy and market competitiveness.

[0003] However, the interactions between functional ingredients are extremely complex, and their synergistic or antagonistic effects depend not only on the type of ingredient but also on the precise dosage and relative ratio of each ingredient. Traditional soft candy formulation development relies heavily on empirical evidence, literature review, and extensive physical experimentation and testing. This trial-and-error approach is time-consuming and costly, and it struggles to systematically explore the vast space of ingredient combinations and dosage ratios. Given the potentially complex interactions among multiple functional ingredients (such as lutein and ginsenosides), relying on manual experience or simple linear models makes it difficult to accurately predict the level of synergy at a specific ratio. Existing formulation optimization methods often focus on adjusting the dosage of a single ingredient or a limited combination of ingredients, lacking an effective approach that can comprehensively understand and model the complex, nonlinear relationships between multiple ingredients and their dosages, allowing for dynamic, intelligent ratio adjustments. In particular, within the specific matrix of functional soft candies, ingredient stability and bioavailability can be affected by the matrix and other excipients, further complicating the prediction problem.

[0004] Therefore, there is an urgent need for an advanced technical solution that can overcome the limitations of traditional methods and utilize data-driven methods, especially the powerful nonlinear modeling capabilities of deep learning, to learn the complex mapping relationship between functional ingredient types, dosages and synergistic enhancements, and guide dynamic and precise proportioning. Summary of the Invention

[0005] In order to solve the above technical problems, this application is proposed.

[0006] According to one aspect of the present application, a deep learning-based dynamic proportioning system for soft candy functional ingredients is provided, which includes: A component dosage acquisition module, configured to acquire a first component and its dosage and a second component and its dosage; a component embedding coding module, configured to perform component embedding coding on the first component and the second component to obtain a first component embedding coding vector and a second component embedding coding vector; a component dose normalization module, configured to normalize the dose of the first component and the dose of the second component to obtain a normalized first component dose and a normalized second component dose; a component dose feature fusion module, configured to fuse the first component embedding code vector and the normalized first component dose and to fuse the second component embedding code vector and the normalized second component dose to obtain a first component dose embedding code vector and a second component dose embedding code vector; a component interaction module, configured to perform multi-scale progressive feature interaction on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain an inter-component multi-scale progressive perceptual fusion coding vector; a synergy prediction module, configured to obtain a predicted synergy score based on the multi-scale progressive perceptual fusion coding vector between the components; The ratio adjustment module is used to compare the predicted synergy score with the target synergy score to determine whether to perform dynamic ratio adjustment of the components.

[0007] Compared with the prior art, the present application provides a dynamic proportioning system for soft candy functional ingredients based on deep learning, which first obtains the first ingredient and its dosage and the second ingredient and its dosage, embeds and encodes the two ingredients to obtain the first and second ingredient embedded coding vectors, and normalizes the respective dosages to obtain the normalized dosage. Then, the embedded coding vectors of each ingredient are fused with the normalized dosage to generate the first and second ingredient dosage embedded coding vectors. These vectors are then interactively processed through multi-scale progressive features to obtain multi-scale progressive perception fusion coding vectors between ingredients. Based on this vector, the synergistic enhancement score is predicted and compared with the target score to determine whether the dynamic proportioning of the ingredients needs to be adjusted. This process ensures the accurate identification and optimization of the synergistic effects between different ingredients, and improves the effectiveness and accuracy of the overall formula. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a block diagram of a dynamic proportioning system of soft candy functional ingredients based on deep learning according to an embodiment of the present application.

[0010] Figure 2 This is a block diagram of the ingredient dosage feature fusion module in the dynamic proportioning system of soft candy functional ingredients based on deep learning according to an embodiment of the present application.

[0011] Figure 3 This is a block diagram of the ingredient interaction module in the dynamic proportioning system of soft candy functional ingredients based on deep learning according to an embodiment of the present application.

[0012] Figure 4 This is a block diagram of a multi-scale fusion unit of ingredients in a dynamic proportioning system of soft candy functional ingredients based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] The following describes embodiments of the present disclosure in more detail with reference to the accompanying drawings. While the drawings illustrate certain embodiments of the present disclosure, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0014] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in a different order and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0015] Figure 1 : is a block diagram of a dynamic proportioning system of soft candy functional ingredients based on deep learning according to an embodiment of the present application. Specifically, Figure 1As shown, according to the embodiment of the present application, the dynamic proportioning system 100 of soft candy functional ingredients based on deep learning includes: an ingredient dosage acquisition module 110, which is used to obtain a first ingredient and its dosage and a second ingredient and its dosage; an ingredient embedding coding module 120, which is used to perform ingredient embedding coding on the first ingredient and the second ingredient to obtain a first ingredient embedding coding vector and a second ingredient embedding coding vector; an ingredient dosage normalization module 130, which is used to normalize the dosage of the first ingredient and the dosage of the second ingredient to obtain a normalized first ingredient dosage and a normalized second ingredient dosage; an ingredient dosage feature fusion module 140, which is used to fuse the first ingredient embedding coding vector and the normalized The first component dose is integrated with the second component embedded coding vector and the normalized second component dose to obtain the first component dose embedded coding vector and the second component dose embedded coding vector; a component interaction module 150 is used to perform multi-scale progressive feature interaction on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain an inter-component multi-scale progressive perception fusion coding vector; a synergy prediction module 160 is used to obtain a predicted synergy score based on the inter-component multi-scale progressive perception fusion coding vector; a ratio adjustment module 170 is used to compare the predicted synergy score with the target synergy score to determine whether to perform dynamic component ratio adjustment.

[0016] Specifically, the ingredient dosage acquisition module 110 is used to obtain the first ingredient and its dosage and the second ingredient and its dosage. It should be understood that the core of this application is to use a deep learning model to understand and predict the synergistic effect produced by the combination of different functional ingredients and their dosages in soft candies, and to dynamically optimize the ratio on this basis. In order to achieve this goal, the system first needs to know clearly which specific ingredients are being evaluated or proportioned, and what the current or target dosage of each of these ingredients is. Therefore, the step of obtaining the first ingredient and its dosage and the second ingredient and its dosage is the starting point and basis of the entire process, and the reason for its processing is to provide all subsequent calculation and prediction modules with the necessary input data that reflects the current ratio status. In particular, in an example of the present application, the first ingredient and the second ingredient are lutein and ginsenoside, respectively.

[0017] The following is a specific implementation process for obtaining the first component and its dosage and the second component and its dosage: Specifically, in a specific example of the present application, the system provides a graphical user interface (GUI) including an input box for entering a first ingredient and a dosage of the first ingredient, and an input box for entering a second ingredient and a dosage of the second ingredient. For example, when using the system, a researcher or formulator can type lutein in the first ingredient input box and a numerical value, such as 10, in the first ingredient dosage input box (this numerical value typically represents the weight or active unit in a unit of soft candy, such as milligrams). Similarly, type ginsenoside in the second ingredient input box and another numerical value, such as 5, in the second ingredient dosage input box. The system receives these input strings (ingredient name) and numerical values (dosage) and stores them as structured data pairs, such as associating lutein with 10 and ginsenoside with 5. These data are passed to the subsequent ingredient embedding encoding module and ingredient dosage normalization module.

[0018] In another specific example of the present application, the system imports data from a pre-existing database or file. An enterprise may maintain a database containing various functional ingredients and their commonly used dosage ranges, or even historical formula data. The user can select a known formula or ingredient combination through the interface, or specify the two ingredients lutein and ginsenosides, and the system then retrieves the set dosage or historical dosage data associated with these ingredients from the database. For example, the user selects an entry called Eye Protection and Refreshing Gummy Recipe A, and the database records information such as the formula containing lutein (dose 10mg) and ginsenosides (dose 5mg). The system performs a database query operation and extracts lutein and its dosage 10, as well as ginsenosides and its dosage 5. These retrieved data are then received and processed by the ingredient dosage acquisition module within the system.

[0019] Specifically, the component embedding coding module 120 is used to perform component embedding coding on the first component and the second component to obtain a first component embedding coding vector and a second component embedding coding vector. Accordingly, since the deep learning model is essentially a mathematical model for processing numerical data, it is unable to directly understand and process text or category identifiers with specific meanings such as lutein and ginsenosides. In addition, each functional ingredient has a unique chemical structure, biological activity, pharmacological action, and potential interaction mode with other ingredients. These complex intrinsic properties and relationships cannot be simply represented by a discrete identifier. Therefore, in order to learn, identify and utilize the unique information of these ingredients to predict the synergistic effect when they are combined, it is necessary to convert the non-numerical identifiers of these ingredients into a numerical representation that can be understood and calculated by the model.

[0020] Component embedding is the key technology for achieving this transformation. It maps each discrete functional component identifier into a dense vector in a continuous, low-dimensional vector space. In this vector space, components with similar properties, functions, or interaction patterns may have their corresponding embedding vectors closer in space, or they may exhibit similar characteristics along specific dimensions. In this way, component embedding goes beyond simply converting text into digital IDs. More importantly, it can capture and encode the potential semantic, functional, or chemical similarities between components, providing the model with rich information about the intrinsic properties of the components.

[0021] The following is a specific implementation process of performing component embedding coding on the first component and the second component to obtain a first component embedding coding vector and a second component embedding coding vector: Specifically, in a specific example of the present application, it is executed after the system receives the ingredients (for example, "lutein" obtained in the previous step as the first ingredient, and "ginsenosides" as the second ingredient). This process is completed by the ingredient embedding coding module in the system. The ingredient embedding coding module internally contains or accesses an ingredient embedding table (Embedding Table) or an embedding layer (Embedding Layer). This embedding table can be regarded as a lookup table, in which each row corresponds to a functional ingredient known to the system, and the content of the row is a numerical vector of a fixed dimension, that is, the embedding coding vector of the ingredient. This embedding table is usually learned by training a deep learning model on a large amount of data related to the functional ingredients, their properties, biological activities or interactions, or is optimized and learned end-to-end with other modules during the training process of the entire soft candy ratio prediction model.

[0022] When the component embedding encoding module receives the name of the first component, such as lutein, it searches its internal component embedding table for the entry corresponding to lutein. Once it finds the entry, the module extracts the pre-computed or trained numeric vector associated with that entry. For example, if the component embedding table is set to 64 dimensions, the extracted embedding vector for lutein might be an array of 64 floating-point numbers, denoted as Vlutein. This vector is the first component embedding encoding vector described in this step.

[0023] Next, the component embedding encoding module performs a similar operation on the name of the second component, such as ginsenosides. It searches the same component embedding table for the entry corresponding to ginsenosides and extracts the numeric vector associated with that entry. For example, the extracted embedding vector for ginsenosides might also be an array of 64 floating-point numbers, denoted as Vginsenosides. This vector is the second component embedding encoding vector described in this step.

[0024] Finally, the output of the component embedding coding module is the two extracted embedding vectors: V lutein (as the first component embedding coding vector) and V ginsenoside (as the second component embedding coding vector).

[0025] Specifically, the component dose normalization module 130 is configured to normalize the doses of the first component and the second component to obtain normalized first and second component doses. It should be understood that the raw dose values obtained from the component dose acquisition module may vary significantly. For example, the dose of one component may be measured in milligrams (mg), ranging from a few milligrams to tens of milligrams; while the dose of another component may be measured in micrograms (µg), ranging from tens to hundreds of micrograms. Furthermore, even when using the same units, the typical dose ranges for different components may vary significantly. When these raw dose data with significant numerical scale differences are used directly as model input features or fused with the embedding vectors output by the component embedding encoding module, which typically have a specific distribution range, features with larger dose values may dominate the model training process, causing the model to overemphasize these features and ignore equally important features with smaller dose values. This not only leads to unstable model training and slow convergence, but also reduces the model's generalization ability, making it difficult to accurately capture the relative relationships between doses and the complex patterns formed by the fusion of dose and component embedding vectors. Therefore, component doses need to be normalized. It's worth noting that normalization eliminates scale differences between raw dose values by scaling all dose values to a predetermined, uniform range (e.g., [0, 1] or a standard distribution with a mean of 0 and a variance of 1). This makes the dose information of different components comparable when entering subsequent feature fusion and interaction, helping the model better understand the proportional relationship between doses and the nonlinear interaction between dose and component type, leading to more accurate predictions of synergy scores.

[0026] The following is a specific implementation process of normalizing the dose of the first component and the dose of the second component to obtain a normalized first component dose and a normalized second component dose: Specifically, in one specific example of the present application, the component dose normalization module receives the first component dose and the second component dose in raw numerical form from the component dose acquisition module. For example, if the acquired first component dose is 10 (representing 10 mg of lutein) and the acquired second component dose is 5 (representing 5 mg of ginsenosides), the component dose normalization module will receive the numerical values 10 and 5.

[0027] The component dose normalization module has a specific normalization method and its parameters preset internally or learned through training data. Common normalization methods include Min-Max normalization and Z-score normalization. Taking Min-Max normalization as an example, this method linearly scales the value to a fixed range [min_range, max_range], usually [0, 1]. Its formula is: x'=min_range+(x-min)*(max_range-min_range) / (max-min), where x is the original dose value, min and max are the minimum and maximum dose values observed in the training dataset for this type of component. These min and max values need to be counted and fixed during the model training phase so that the same parameters can be used for normalization during the inference phase.

[0028] Suppose that in the training dataset, the lutein dose range is [2mg, 20mg], the ginsenoside dose range is [1mg, 15mg], and the doses are normalized to the range [0, 1]. When the component dose normalization module receives the first component dose of 10mg (lutein), it applies the Min-Max normalization formula for lutein: Normalized first component dose = 0 + (10-2) * (1-0) / (20-2) = 8 / 18 ≈ 0.444. This value, 0.444, is the normalized first component dose.

[0029] Similarly, when receiving a second component dose of 5 mg (ginsenosides), the Min-Ma normalization formula for ginsenosides is applied: Normalized second component dose = 0 + (5-1) * (1-0) / (15-1) = 4 / 14 ≈ 0.286. This value, 0.286, is the normalized second component dose.

[0030] Specifically, the ingredient-dose feature fusion module 140 is configured to fuse the first ingredient embedding vector with the normalized first ingredient dose, and to fuse the second ingredient embedding vector with the normalized second ingredient dose to obtain a first ingredient dose embedding vector and a second ingredient dose embedding vector. Accordingly, considering that the synergistic effect of the functional ingredients of gummy candies depends not only on the type of ingredients themselves (represented by the ingredient embedding vector), but also on the precise dose of each ingredient present (represented by the normalized dose), the ingredient embedding vector captures the intrinsic properties and potential interactions of the ingredients but does not contain information about their quantity; whereas the normalized dose value reflects the quantity of the ingredients but does not reflect the intrinsic properties of the ingredients themselves. Therefore, in order for the deep learning model to accurately predict the synergistic score for a specific combination, it requires a comprehensive representation that simultaneously encodes the ingredient type and its corresponding dose, namely, fusing the ingredient embedding vector with the corresponding normalized dose. These fused vectors contain more comprehensive information, enabling the subsequent ingredient interaction module to more accurately model the complex interaction patterns of different ingredients at a specific combination by considering the comprehensive features of both ingredient type and dose.

[0031] Specifically, in a specific example of this application, Figure 2 FIG is a block diagram of an ingredient dosage feature fusion module in a soft candy functional ingredient dynamic ratio system based on deep learning according to an embodiment of the present application. Specifically, Figure 2 As shown, the component dose feature fusion module 140 includes: a first component dose embedding unit 141, used to add the normalized first component dose to the end of the first component embedded coding vector to obtain the first component dose embedded coding vector; a second component dose embedding unit 142, used to add the normalized second component dose to the end of the second component embedded coding vector to obtain the second component dose embedded coding vector.

[0032] It should be understood that the choice of appending the normalized component dose to the end of the component embedding encoding vector is a common feature fusion method—vector concatenation. This method is simple to operate and retains all the information of the original embedding vector while introducing dose as a new dimension feature. By appending dose as one or more independent dimensions to the end of the embedding vector, subsequent layers of the model can learn how to associate and interact with this dose feature with the various dimensions of the embedding vector, thereby discovering the specific impact of dose changes on component functions and interactions.

[0033] Specifically, in a specific example of the present application, the first component dose embedding unit 141 is implemented as follows: This unit receives two inputs: one is the first component embedding code vector from the component embedding coding module, for example, the D-dimensional embedding vector Vlutein = [v1, v2, ..., vD] of lutein; and the other is the normalized first component dose from the component dose normalization module, for example, the normalized lutein dose value dlutein. The core operation performed by this unit is to treat the normalized scalar dose value dlutein_norm as a single-dimensional vector [dlutein_norm] and then perform vector concatenation with the first component embedding code vector Vlutein. Specifically, the single-dimensional vector [dlutein_norm][dlutein_norm] is added to the end of the vector Vlutein and concatenated. The output is the first component dose embedding code vector Vlutein_dose, which is a (D+1)-dimensional vector of the form [v1, v2, ..., vD, dlutein_norm]. In this way, the intrinsic property information of lutein as a functional ingredient and its normalized dosage information in the current formula are integrated into the same vector representation. Similarly, the implementation process of the second ingredient dosage embedding unit 142 is the same as that of the first ingredient dosage embedding unit.

[0034] Specifically, the component interaction module 150 is configured to perform multi-scale progressive feature interaction on the first component dose embedding vector and the second component dose embedding vector to obtain a multi-scale progressive perceptual fusion encoding vector for each component. Furthermore, considering that the synergistic effect between different functional components in gummy candies (e.g., lutein and ginsenosides) is not simply determined by the independent properties or dosage of each component, but rather stems from their complex interactions at specific dosage combinations. This interaction is multi-layered and nonlinear, potentially occurring at multiple levels, such as chemical, biochemical, and physiological, with interactions at different levels contributing differently to the final synergistic effect. Although the previous step generates dose embedding vectors containing component type and dosage information, these vectors still describe individual components and do not directly model how the two components influence, promote, or inhibit each other. Therefore, it is necessary to comprehensively capture and comprehensively represent the characteristics of this complex "interaction." To this end, the present application performs multi-scale progressive feature interaction on the first component dose embedding vector and the second component dose embedding vector to obtain a multi-scale progressive perceptual fusion encoding vector for each component.

[0035] Figure 3 : is a block diagram of the ingredient interaction module in the dynamic proportioning system of soft candy functional ingredients based on deep learning according to an embodiment of the present application. Specifically, Figure 3As shown, the component interaction module 150 includes: a component multi-level feature extraction unit 151, which is used to perform multi-level implicit feature extraction on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain a first component dose embedded middle-level implicit coding vector, a second component dose embedded middle-level implicit coding vector, a first component dose embedded deep-level implicit coding vector and a second component dose embedded deep-level implicit coding vector; an inter-component multi-level feature fusion unit 152, which is used to perform multi-level feature fusion on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain an inter-component low-level fusion feature coding vector, an inter-component middle-level fusion feature coding vector and an inter-component deep-level fusion feature coding vector; a component multi-scale fusion unit 153, which is used to perform progressive perceptual fusion on the inter-component low-level fusion feature coding vector, the inter-component middle-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector to obtain the inter-component multi-scale progressive perceptual fusion coding vector.

[0036] In particular, the component multi-level feature extraction unit 151 is used to perform multi-level implicit feature extraction on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain a first component dose embedded middle layer implicit coding vector, a second component dose embedded middle layer implicit coding vector, a first component dose embedded deep layer implicit coding vector, and a second component dose embedded deep layer implicit coding vector. The processing process can be expressed as follows: ;in, is the first component dose embedding coding vector, is the second component dose embedding coding vector, and are the middle-level latent feature weight matrix and the middle-level latent feature bias vector, yes activation function, is the first component dose embedded in the middle layer implicit coding vector, is the second component dose embedded in the middle layer implicit coding vector, and are the deep latent feature weight matrix and the deep latent feature bias vector, yes activation function, is the first component dose embedded deep implicit encoding vector, is the second component dose embedded in the deep implicit encoding vector.

[0037] It should be understood that synergistic effects cannot be described simply by simple linear superposition or direct combination of raw features; they often involve the mutual influence and regulation of components at higher, more abstract levels. For example, the synergistic effect of lutein and ginsenosides may not only be due to their individual doses, but also to their joint effects on specific physiological pathways, cellular targets, or systemic functions. These deeper, more abstract properties and interaction patterns cannot be easily extracted and identified directly from the raw, relatively low-level component dose embedding vector. Therefore, multi-level latent feature extraction is performed on the component dose embedding vector, which contains information about component type and dose. The goal is to leverage the layer-by-layer abstraction capabilities of deep neural networks to decouple the raw dose embedding vector and extract latent features that represent the components at different levels of abstraction at specific doses. For example, a shallow nonlinear transformation of the first component dose embedding vector may extract mid-level latent features related to the physicochemical properties of the component and its dispersibility in the gummy candy matrix. Further transformations may then extract deeper latent features related to more abstract functions of lutein, such as ocular health and antioxidant activity. The same process is applied to the dose embedding encoding vector of the second component, extracting implicit features reflecting its characteristics at different abstract levels in terms of central nervous system, immune regulation, etc. This multi-level feature extraction ensures that the system can comprehensively and meticulously capture the various types of complex nonlinear interactions that may occur between components, thus laying a solid feature foundation for predicting highly complex synergistic effects.

[0038] In particular, in a specific example of the present application, the component multi-level feature fusion unit 152 is used to: The first component dose embedded coding vector and the second component dose embedded coding vector are subjected to MLP-based low-level feature fusion to obtain the inter-component low-level fusion feature coding vector. The processing process can be expressed as follows: ;in, is the first component dose embedding coding vector, is the second component dose embedding coding vector It is added by position point. is a multi-layer perceptron, It is the low-level fusion feature encoding vector between components.

[0039] The first component dose embedded mid-level implicit coding vector and the second component dose embedded mid-level implicit coding vector are subjected to mid-level feature fusion based on attention association to obtain the mid-level fusion feature coding vector between the components. The processing process can be expressed as follows: ;in, is the first component dose embedded in the middle layer implicit coding vector, is the second component dose embedded in the middle layer implicit coding vector, is the vector transpose operation, yes length, yes function, yes and The middle-level attention association weight vector between It is the point product of position. It is the mid-level fusion feature encoding vector between components.

[0040] The first component dose embedded deep implicit coding vector and the second component dose embedded deep implicit coding vector are subjected to deep-level feature fusion based on low-rank projection gating to obtain the inter-component deep-level fusion feature coding vector. The processing process can be expressed as follows: ;in, is the first component dose embedded deep implicit encoding vector, is the second component dose embedded deep implicit coding vector, and is the low-rank projection matrix, yes activation function, is the hidden state deep projection interaction vector, is the deep interaction weight matrix, is the layer normalization operation, It is the deep level fusion feature encoding vector between components.

[0041] Accordingly, the synergistic effect between functional ingredients, at the most basic level, often has a direct or relatively simple nonlinear relationship with the original properties and dosage of the ingredients themselves. Although subsequent deep neural network layers will extract higher-level abstract features, some fine information and direct correspondences of the original input may be lost in the process of layer-by-layer abstraction. For example, the simple additive effect of lutein and ginsenosides at a specific low dose, or the mutual influence of their solubility and stability in the gummy matrix, these basic-level interactions may directly depend on their original chemical properties and their respective dosages. If this low-level interaction pattern can be effectively captured and encoded in the early stages of the model, it will provide a solid foundation that retains the original details for subsequent more complex interaction modeling.

[0042] Therefore, low-level feature fusion is performed directly on the first component dose embedding encoding vector and the second component dose embedding encoding vector, which contain the original component embeddings and normalized doses. The goal is to capture the fundamental interactions between the components at the lowest level of abstraction. By adopting a multi-layer perceptron (MLP)-based fusion method, the system is able to learn the nonlinear combination and mapping between the two original input vectors. As a classic deep learning model, the MLP has powerful nonlinear modeling capabilities. It can perform complex nonlinear transformations and feature combinations on the input vector through hidden layers. In this low-level fusion scenario, the MLP-based fusion can learn how to nonlinearly associate the various dimensions of the first component dose embedding encoding vector (representing component attributes and dose) with the various dimensions of the second component dose embedding encoding vector, thereby capturing the most direct and fundamental interaction pattern between lutein and ginsenosides at the current dose.

[0043] It should be understood that low-level feature fusion captures the original, basic level of association, while the deeper mechanisms of this synergistic effect, such as the ingredients jointly affecting the function of specific organelles, synergistically regulating the key enzyme activity of a metabolic pathway, or forming microstructures in the gummy matrix that are conducive to absorption or stability, are all interactions that occur at a medium level of abstraction. The ingredient dosage generated by the multi-level implicit feature extraction step is embedded in the middle-level implicit encoding vector, which is precisely the characteristic representation of the ingredient at these "component level" or "structure level" abstract levels at a specific dosage. These middle-level features filter out some of the original noise, encode more stable patterns, and are the basis for modeling structural associations and compositional relationships between ingredients. In order to accurately predict synergistic effects, the system needs to specifically capture and encode these middle-level interaction patterns. Specifically, in one example of the present application, a method based on attention association is used for middle-level feature fusion. The fusion based on the attention mechanism allows the model to dynamically learn which middle-level feature dimensions are most important for capturing this interactive relationship. This allows the system to selectively focus on the most relevant features when fusing two mid-level vectors, assigning them higher weights based on their specific content. This allows for more refined and efficient extraction and encoding of the most predictive mid-level interaction information. This biased fusion approach, which dynamically adjusts weights based on input features, is more adaptable to complex and changing component interaction patterns than simple concatenation or summation.

[0044] Accordingly, at the highest level of abstraction or semantics, the synergistic effects of functional ingredients often manifest as joint promotion of overall health goals or systemic functions. The ingredient doses generated by the multi-level latent feature extraction steps are embedded into deep latent encoding vectors, representing highly abstract, global characteristics of the ingredients at specific doses. These deep features may encode the effects of the ingredients on overall physiological balance, important systemic functions (such as visual health, antioxidant capacity, and immune regulation), or their core roles in complex biological networks. The most direct manifestation of synergy is often these high-level functional or semantic synergies, such as the combined effects of lutein and ginsenosides in simultaneously promoting eye health and alleviating systemic fatigue. Deep features represent the model's highest-level understanding and summary of the individual ingredient dose inputs, and their fusion is necessary to capture these highest-level, most crucial interactions. To this end, this application employs a low-rank projection-based gating method for deep-level feature fusion. This gating mechanism makes the fusion process more intelligent, selectively highlighting feature interactions that best reflect deep semantic synergies based on the specific ingredient combination and dose. For example, low-rank projection gating may learn that when lutein and ginsenosides are present simultaneously and at high doses, their interaction on deep features related to energy metabolism and neuroprotection is highly “gated” or enhanced, as this interaction is closely related to their synergistic effect in alleviating fatigue.

[0045] In particular, the component multi-scale fusion unit 153 is configured to perform progressive perceptual fusion on the inter-component low-level fusion feature encoding vectors, the inter-component mid-level fusion feature encoding vectors, and the inter-component deep-level fusion feature encoding vectors to obtain the inter-component multi-scale progressive perceptual fusion encoding vectors. It should be understood that synergistic enhancement between functional components involves multi-scale interactions, including low-level physicochemical or direct biochemical levels, mid-level structural associations or pathway levels, and deep-level overall physiological functions or health benefit promotion. Although the interactive information at these three abstract levels is encoded separately, a single level is insufficient to fully describe the synergistic effect. The low-level fusion feature encoding vector provides foundational details, the mid-level captures structural associations, and the deep-level summarizes global semantic synergy. The information at these different levels has different properties and granularity, and is not simply superimposed, but rather complements each other to determine the ultimate effect. To fully characterize the interactions between components, this multi-source, multi-scale information needs to be intelligently integrated to maximize the unique contributions and complementarity of each. This application uses a progressive perceptual fusion approach to integrate information in an orderly and progressively refined manner, explicitly perceiving the complementarity between fused features at each level. Unlike simple concatenation or summation, this approach dynamically weighs and fuses interactive information from different levels through a learning mechanism, achieving "complementary perception." The model can discern how low-level details support mid-level structure and how mid-levels connect to deep-level semantics, generating a joint representation that is both rich and reflects multi-scale interaction patterns.

[0046] Figure 4 : is a block diagram of a multi-scale fusion unit of a soft candy functional ingredient dynamic proportioning system based on deep learning according to an embodiment of the present application. In particular, in a specific example of the present application, Figure 4 As shown, the component multi-scale fusion unit 153 includes: an inter-component low-level middle-level gating fusion sub-unit 1531, which is used to perform gated weight fusion on the inter-component low-level fusion feature coding vector and the inter-component middle-level fusion feature coding vector to obtain the inter-component low-level middle-level fusion feature coding vector; an inter-component multi-scale attention fusion sub-unit 1532, which is used to perform cross-attention fusion on the inter-component low-level middle-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector to obtain the inter-component multi-scale progressive perception fusion coding vector.

[0047] In particular, the inter-component low-level mid-level gating fusion subunit 1531 is used to perform gated weight fusion on the inter-component low-level fusion feature coding vector and the inter-component mid-level fusion feature coding vector to obtain the inter-component low-level mid-level fusion feature coding vector. The processing process can be expressed as follows: ;in, is the low-level fusion feature encoding vector between components, is the mid-level fusion feature encoding vector between components, For cascade operation, is the cascade weight matrix, is the low- and middle-level probabilistic dynamic control parameter, It is the low-level and mid-level fusion feature encoding vector between components.

[0048] Specifically, the interactions between functional ingredients (such as lutein and ginsenosides) at the low-level (raw features / details) and mid-level (structure / components) capture different aspects of synergy, yet they are closely interconnected and complementary. The low-level fusion feature encoding vector reflects the interaction patterns of the ingredients at the most fundamental and detailed level, potentially related to their physicochemical compatibility or direct biochemical reactions; whereas the mid-level fusion feature encoding vector captures interactions at the structural or local biological pathway level, after preliminary abstraction. To form a more comprehensive and hierarchically structured representation of interactions, early integration of information from these adjacent scales creates an intermediate representation that captures both basic details and higher-level structural connections. The gated weight fusion approach employed here further recognizes that the relative importance of low-level and mid-level information in predicting synergy is not fixed but rather depends on the specific ingredient type and dosage. For example, for some ingredient combinations, the synergistic effect may be primarily manifested by direct interactions at low doses (emphasizing low-level interactions); whereas for other combinations, mid-level structural synergies (e.g., co-regulation of a pathway) may be more critical. The gating mechanism allows the model to dynamically learn and generate weights based on the content of the low-level and mid-level fusion feature encoding vectors between the input components, thereby intelligently deciding in what proportion or manner the information of these two levels should be combined.

[0049] In particular, in a specific example of the present application, the inter-component multi-scale attention fusion sub-unit 1532 is used to: process the inter-component low-level mid-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector based on the pre-training weight matrix to obtain an inter-component progressive perception query vector, an inter-component progressive perception key vector and an inter-component progressive perception value vector; based on the inter-component progressive perception query vector, the inter-component progressive perception key vector and the inter-component progressive perception value vector, perform pre-training weight adaptive optimization based on domain balanced alignment on the pre-training weight matrix to obtain an optimized pre-training weight matrix; based on the optimized pre-training weight matrix, perform cross-level attention fusion based on softmax on the inter-component low-level mid-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector to obtain the inter-component multi-scale progressive perception fusion coding vector.

[0050] In particular, the inter-component low-level mid-level fusion feature encoding vector and the inter-component deep-level fusion feature encoding vector are processed based on the pre-trained weight matrix to obtain the inter-component progressive perception query vector, the inter-component progressive perception key vector and the inter-component progressive perception value vector. The processing process can be expressed as follows: ;in, is the deep-level fusion feature encoding vector between components, is the low-level and mid-level fusion feature encoding vector between components, and is the pre-trained weight matrix, and They are the inter-component progressively perceptual query vector, the inter-component progressively perceptual key vector, and the inter-component progressively perceptual value vector.

[0051] It should be understood that achieving cross-level attentional fusion between low- and mid-level fusion information and deep-level fusion information first requires transforming these input feature vectors into a compatible space, where they play the roles of query, key, and value in the attention mechanism. The core of the attention mechanism lies in computing the dot-product similarity between the query and key vectors and using this similarity to weight the sum of the value vectors, thereby enabling focused information extraction and integration. This query, key, and value paradigm is a standard and efficient way to compute correlations between input features and perform weighted aggregation based on them. By transforming the fused feature vectors at different levels (the deep-level fused feature encoding vectors and the low- and mid-level fused feature encoding vectors) into inter-component progressively perceptual query vectors, inter-component progressively perceptual key vectors, and inter-component progressively perceptual value vectors, this lays the structural foundation for subsequent attention computation. In particular, when dealing with tasks such as inter-component interactions, where data may be relatively limited or features may be complex, utilizing weight matrices pre-trained on more general or related tasks can provide a better initial starting point for the model. Pre-trained weights may have captured some common feature mappings or conversion patterns, which helps the model converge faster and learn effective and Projection improves model performance and generalization. For example, these weights might come from a Transformer model pre-trained on a large amount of biomedical or chemical data. The deep-level fusion feature encoding vectors and the low- to mid-level fusion feature encoding vectors between components are mapped to the Q, K, and V spaces, respectively, using these pre-trained linear transformations (i.e., multiplication with the pre-trained weight matrix). This makes features from different layers semantically comparable when performing similarity calculations and weighted summations.

[0052] In particular, based on the inter-component progressive perception query vector, the inter-component progressive perception key vector, and the inter-component progressive perception value vector, the pre-trained weight matrix is adaptively optimized based on domain-balanced alignment to obtain an optimized pre-trained weight matrix. The specific processing process of this process is as follows: When the inter-component low-level fusion feature coding vector, the inter-component mid-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector are progressively perceptually fused, the inter-component low-level fusion feature coding vector and the mid-level fusion feature encoding vector between the components First, by generating low- and middle-level probabilistic dynamic control parameters To achieve fusion, the low-middle layer probabilistic dynamic control parameters The low-level and mid-level features can be fused collaboratively based on the binary probability synergy mechanism. And further, the deep-level fusion feature encoding vector between the components Pre-trained weight matrix and Jointly construct the query target state parameter domain and perform self-attention fusion. Therefore, it is necessary to consider when performing query target state projection. and The target state parameter domain represented by it will also be fused to the low-level and middle-level fusion feature encoding vectors between components in the state distribution. and deep-level fusion feature encoding vectors between components alignment, thus causing domain-balanced alignment loss in the self-attention fusion process, affecting the expression effect of the multi-scale progressive perceptual fusion encoding vector between the components.

[0053] Therefore, for the pre-trained weight matrix and , first divide it into the low-level and middle-level fusion feature encoding vectors belonging to the components The pre-trained weight matrix , and the deep-level fusion feature encoding vectors belonging to the components The pre-trained weight matrix and These two groups calculate the partial derivative target state parameter domain response representation of the self-attention fusion, namely: ;in, yes and Length, and and Same length, is the partial derivative calculation, is the query response representation weight matrix.

[0054] That is, for The partial derivative of How does the change of the vector components affect the final result? The partial derivative is a matrix, that is, a vector Each eigenvalue of To affect the inner product scalar value, and then according to the vector The vector field is enlarged or reduced.

[0055] , which is consistent with the above The corresponding ones are the same, where is the key response representation weight matrix; ;in, is the identity matrix with the same dimensions as the weight matrix, is the value response representation weight matrix.

[0056] Then, define the domain balance alignment loss function: ;in, represents the F norm of the matrix, and To scale the hyperparameters, is the domain-balanced alignment loss function.

[0057] This method uses a self-attention-based fusion response representation in the query-based target state parameter domain to relatively accurately measure the response alignment during the self-attention fusion process. This method eliminates the domain-balanced structured loss caused by hierarchical alignment imbalance, which can hinder self-attention convergence. Subsequently, by iterating the convergence of each weight matrix toward the response, the alignment interaction capability of the weight matrices is maintained based on their domain-balanced properties, optimizing the domain-balanced alignment loss and improving the expressive performance of the multi-scale progressive perceptual fusion coding vector between the components.

[0058] In particular, based on the optimized pre-trained weight matrix, the inter-component low-level mid-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector are subjected to cross-level attention fusion based on softmax to obtain the inter-component multi-scale progressive perceptual fusion coding vector. The processing process can be expressed as follows: ;in, and is to optimize the pre-training weight matrix, and They are the inter-component progressively perceptually optimized query vector, the inter-component progressively perceptually optimized key vector, and the inter-component progressively perceptually optimized value vector. yes length, yes function, is the multi-scale progressive perceptual fusion coding vector between components.

[0059] That is, while the low- and mid-level fused feature encoding vectors (combining underlying details and mid-level structural connections) and the deep-level fused feature encoding vectors (capturing top-level semantic synergies) each represent different-scale interactions between components, this information needs to be intelligently integrated to form a comprehensive and most predictive joint representation. Simple concatenation or summation fails to effectively exploit the complementarity between these different-scale information, nor can it dynamically weigh the importance of interactions at different scales based on component combination and dosage. To capture the complex and multifaceted synergistic mechanisms between components like lutein and ginsenosides, a mechanism is needed that understands how low- and mid-level interactions relate to deep-level interactions and selectively highlights those cross-level signals that best reflect synergistic effects. To this end, this application employs a softmax-based cross-level attention fusion method, which is an effective approach to achieve this intelligent integration. The attention mechanism allows the model to generate a weight distribution by calculating the similarity between the query vector and the key vector, and then use these weights to perform a weighted summation of the value vector. This "cross-level" attention enables the model to learn how low- and mid-level interaction information "queries" or "focuses on" the most relevant parts of deep-level interaction information. The Softmax function ensures the non-negativity and normalization of the attention weights, allowing them to be interpreted as the importance distribution of each dimension of deep-level information, making the weighted sum probabilistic, smooth, and differentiable. The resulting multi-scale progressive perceptual fusion encoding vector between components is the model's most comprehensive and robust encoding of the potential synergistic effects of the current specific component and dosage combination, and is directly used in downstream synergistic effect prediction tasks, significantly improving the accuracy and reliability of the prediction.

[0060] Specifically, all of the aforementioned weight matrices and bias vectors are learned as learnable parameters in the deep learning model through an end-to-end training process. During the training process of the entire gummy functional ingredient dynamic ratio system, the goal is to minimize the error between the model's predicted inter-ingredient synergy score and the true or target synergy score. The weight matrices and bias vectors are integral to its computational logic (e.g., used to calculate gating signals and perform linear transformations). During training, the system calculates the predicted scores through forward propagation, then computes a loss function based on the prediction errors. Next, a backpropagation algorithm is used to calculate the gradient of the loss function with respect to all learnable parameters of the model (weight matrices and biases). An optimization algorithm (such as Adam or SGD) updates the weight matrices and bias vectors based on these gradients. This iterative process is repeated on a large amount of ingredient dosage combination data, allowing the model to gradually learn the optimal weights and biases, enabling the gating mechanism to effectively balance and integrate low-level and mid-level fusion features, thereby minimizing the final synergy prediction error. Ultimately, these weight matrices and bias values reflect the model's data-learned strategy for optimally integrating the interaction information at these two scales.

[0061] Specifically, the synergy prediction module 160 is configured to obtain a predicted synergy score based on the multi-scale progressive perceptual fusion encoding vector between components. In particular, in one specific example of the present application, the synergy prediction module is configured to input the multi-scale progressive perceptual fusion encoding vector between components into an output head based on a fully connected layer to obtain the predicted synergy score. It should be understood that the preceding component interaction module, particularly the progressive perceptual fusion process, has intelligently integrated and compressed the multi-scale, multi-level interactions (including low-level physicochemical connections, mid-level structural synergies, and deep-level global functional impacts) between functional components such as lutein and ginsenosides at specific dosages into a single, high-information "multi-scale progressive perceptual fusion encoding vector between components." This vector is the model's most comprehensive and robust abstract representation of the potential synergy for this specific component combination and dosage ratio. The task of synergy prediction is essentially to map this high-dimensional, abstract interaction representation into a concrete, typically continuous or discrete, synergy score. Therefore, the inter-component multi-scale progressive perceptual fusion encoding vector is the optimal input feature for prediction tasks because it is designed to maximally capture the complex nonlinear relationships that are directly related to synergy.

[0062] Specifically, in one example of the present application, the process of inputting the inter-component multi-scale progressive perceptual fusion encoding vector into a fully-connected layer-based output head to obtain the predicted synergy score is as follows: First, the inter-component multi-scale progressive perceptual fusion encoding vector is input into the first fully-connected layer, also known as the hidden layer. In this layer, each dimension of the input vector is linearly combined with the learned weight matrix of that layer, and then a learned bias vector is added. The result of this linear combination is then transformed through a nonlinear activation function (e.g., a function that allows the model to learn non-linear relationships). The result of this nonlinear transformation forms a hidden layer output vector. The purpose of this process is to map the input fusion features into a new internal representation space that is more suitable for the prediction task. If the output head contains multiple hidden layers, the above linear combination and nonlinear transformation process is repeated sequentially through these hidden layers, with the output of each layer serving as the input to the next layer, gradually extracting and transforming features. Finally, the vector processed by all hidden layers (or, if there is only one layer, the result of the original input after a linear and nonlinear transformation) is input into the last fully-connected layer, also known as the output layer. In the output layer, another linear combination operation is performed, linearly combining the dimensions of the previous layer's output vector with the weight vector learned in this layer, and adding a learned bias term. The final linear combination result is a single numerical value. This single numerical value is the predicted synergy score. In tasks such as predicting synergy scores, which typically require outputting a continuous numerical value, the output layer typically does not use an activation function and directly outputs the linear combination result as the final score.

[0063] Specifically, the ratio adjustment module 170 is used to compare the predicted synergistic effect score with the target synergistic effect score to determine whether to perform dynamic ingredient ratio adjustment. It should be understood that the goal of the entire system is to find a dosage ratio that can enable functional ingredients such as lutein and ginsenosides to achieve a specific target synergistic effect level in soft candies. The synergistic effect prediction module provides a predicted synergistic effect score at the current given ingredient dosage, which represents the synergistic effect that the model believes the current ratio may produce. However, the goal of the system is to achieve a preset "target synergistic effect score." Therefore, comparing the predicted score with the target score is a direct way to evaluate whether the current ingredient dosage ratio has met or is close to the expected synergistic effect. If there is a significant difference between the predicted score and the target score, it indicates that the current ratio scheme is not optimal and needs to be adjusted to approach the target. This comparison provides a clear feedback signal to drive the system into the next round of ratio optimization process until the predicted score fully matches the target score.

[0064] Specifically, the target synergy score is not calculated internally by the system based on a given ingredient dosage ratio; rather, it is a key metric pre-defined externally. The specific process for determining this score typically occurs during the product development and formulation phase, based on the specific functional goals and market positioning of the gummy product. The target synergy score can be derived from a variety of sources. For example, it may be based on existing scientific research data or clinical trial results that have quantified the synergistic effects of lutein and ginsenosides in specific health areas (such as visual health and fatigue reduction) and established the synergy level required to achieve significant results. It may also be derived from the empirical knowledge of product managers or nutritionists, who determine a desired synergy level based on the target consumer group, the claimed efficacy of the product, and the performance of competitive products. Furthermore, market research, regulatory requirements, or internal performance standards may influence the determination of the target score.

[0065] Specifically, in one embodiment of the present application, the predicted synergy score is compared with the target synergy score to determine whether to dynamically adjust the ingredient ratio. The specific implementation process is as follows: First, the system obtains the predicted synergy score output by the synergy prediction module. This is a quantitative value that reflects the model's prediction of the synergistic effect of the current lutein and ginsenoside dosage combination. Simultaneously, the system has preset or received a target synergy score, which is a quantitative representation of the desired level of synergy achieved by the soft candy ratio. Next, the ratio adjustment module compares these two scores. This comparison is typically performed by calculating the difference between the predicted score and the target score. For example, the absolute difference can be calculated, or the relative difference can be considered. To avoid triggering adjustments even for minor differences, an acceptable error range or threshold is typically set. The specific comparison logic determines whether the difference between the predicted score and the target score exceeds a preset threshold. If the absolute value of the difference between the predicted score and the target score exceeds a preset threshold (e.g., an absolute difference > 0.05), the current ratio is deemed to have failed to meet the target requirement and requires ratio adjustment.

[0066] On the contrary, if the absolute value of the difference between the predicted score and the target score is less than or equal to the set threshold (for example, the absolute difference is <= 0.05), it is considered that the current ingredient dosage ratio is already able to achieve the target synergistic efficiency level well. At this time, there may be no need for further dynamic ratio adjustment, or the current ratio is considered to be an acceptable final solution. Through this comparison-based feedback mechanism, the system can efficiently explore different ingredient dosage combination spaces and ultimately find the solution that can maximize the preset synergistic efficiency. Specifically, the threshold is set to balance the convergence speed of the optimization process with the accuracy and practicality of the final formulation. This is typically achieved through extensive experimentation and experience. During the system development and tuning phase, researchers experiment with different threshold settings and run the dynamic formulation adjustment process on a test set containing a variety of component combinations and dosage data. The effectiveness of the threshold is evaluated by observing the system's convergence behavior at different thresholds (e.g., the number of iterations required to reach the target), the resulting formulation, and the actual difference between the corresponding predicted score and the true or expected synergy score. Excessively large thresholds may cause the system to stop prematurely, resulting in insufficiently optimized formulations and a significant deviation between the predicted score and the target score. Excessively small thresholds may lead the system into unnecessary fine-tuning cycles, increasing computational cost and time, and potentially challenging robustness in production. Small dosage variations are difficult to precisely control in production, and the resulting synergy score differences may be smaller than measurement errors or inherent model prediction uncertainties. Therefore, through repeated experimentation and verification, a threshold is sought that ensures that the predicted score is sufficiently close to the target while also ensuring efficient and stable convergence of the optimization process.

[0067] In summary, a deep learning-based dynamic proportioning system 100 for soft candy functional ingredients based on an embodiment of the present application is illustrated, which first obtains a first ingredient and its dosage and a second ingredient and its dosage, embeds and encodes the two ingredients to obtain first and second ingredient embedding coding vectors, and normalizes each dosage to obtain a normalized dosage. Next, the embedded coding vectors of each ingredient are fused with the normalized dosage to generate first and second ingredient dosage embedding coding vectors. These vectors are then interactively processed through multi-scale progressive features to obtain multi-scale progressive perception fusion coding vectors between ingredients. Based on this vector, a synergistic enhancement score is predicted and compared with the target score to determine whether the dynamic proportioning of ingredients needs to be adjusted. This process ensures the accurate identification and optimization of the synergistic effects between different ingredients, thereby improving the effectiveness and accuracy of the overall formula.

[0068] As described above, the dynamic proportioning system 100 for soft candy functional ingredients based on deep learning according to the embodiment of the present application can be implemented in various wireless terminals, such as a server having a dynamic proportioning algorithm for soft candy functional ingredients based on deep learning. In one possible implementation, the dynamic proportioning system 100 for soft candy functional ingredients based on deep learning according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the dynamic proportioning system 100 for soft candy functional ingredients based on deep learning can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the dynamic proportioning system 100 for soft candy functional ingredients based on deep learning can also be one of the many hardware modules of the wireless terminal.

[0069] Alternatively, in another example, the deep learning-based dynamic proportioning system for soft candy functional ingredients 100 and the wireless terminal may also be separate devices, and the deep learning-based dynamic proportioning system for soft candy functional ingredients 100 may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0070] While various implementations of the present disclosure have been described above, the above description is intended to be illustrative and not exhaustive. The present disclosure is not limited to the disclosed implementations, and numerous modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described implementations.

Claims

1. A dynamic proportioning system of soft candy functional ingredients based on deep learning, characterized in that: include: A component dosage acquisition module, configured to acquire a first component and its dosage and a second component and its dosage; a component embedding coding module, configured to perform component embedding coding on the first component and the second component to obtain a first component embedding coding vector and a second component embedding coding vector; a component dose normalization module, configured to normalize the dose of the first component and the dose of the second component to obtain a normalized first component dose and a normalized second component dose; a component dose feature fusion module, configured to fuse the first component embedding code vector and the normalized first component dose and to fuse the second component embedding code vector and the normalized second component dose to obtain a first component dose embedding code vector and a second component dose embedding code vector; a component interaction module, configured to perform multi-scale progressive feature interaction on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain an inter-component multi-scale progressive perceptual fusion coding vector; a synergy prediction module, configured to obtain a predicted synergy score based on the multi-scale progressive perceptual fusion coding vector between the components; The ratio adjustment module is used to compare the predicted synergy score with the target synergy score to determine whether to perform dynamic ratio adjustment of the components.

2. The dynamic proportioning system of soft candy functional ingredients based on deep learning according to claim 1 is characterized in that: The first component and the second component are lutein and ginsenoside, respectively.

3. The dynamic proportioning system of soft candy functional ingredients based on deep learning according to claim 2 is characterized in that: The component dose feature fusion module includes: a first component dose embedding unit, configured to add the normalized first component dose to the end of the first component embedding code vector to obtain the first component dose embedding code vector; The second component dose embedding unit is used to add the normalized second component dose to the end of the second component embedded coding vector to obtain the second component dose embedded coding vector.

4. The dynamic proportioning system of soft candy functional ingredients based on deep learning according to claim 1, characterized in that: The component interaction module includes: a component multi-level feature extraction unit, configured to perform multi-level implicit feature extraction on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain a first component dose embedded middle-layer implicit coding vector, a second component dose embedded middle-layer implicit coding vector, a first component dose embedded deep-layer implicit coding vector, and a second component dose embedded deep-layer implicit coding vector; an inter-component multi-level feature fusion unit, configured to perform multi-level feature fusion on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain an inter-component low-level fusion feature coding vector, an inter-component mid-level fusion feature coding vector, and an inter-component deep-level fusion feature coding vector; The component multi-scale fusion unit is used to perform progressive perceptual fusion on the inter-component low-level fusion feature coding vector, the inter-component mid-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector to obtain the inter-component multi-scale progressive perceptual fusion coding vector.

5. The dynamic proportioning system of soft candy functional ingredients based on deep learning according to claim 4 is characterized in that: The inter-component multi-level feature fusion unit is used to: Performing MLP-based low-level feature fusion on the first component dose embedded coding vector and the second component dose embedded coding vector to obtain the inter-component low-level fusion feature coding vector; Performing attention-association-based mid-level feature fusion on the mid-level implicit coding vector embedded in the first component dose and the mid-level implicit coding vector embedded in the second component dose to obtain the mid-level fusion feature coding vector between the components; The first component dose embedded deep implicit coding vector and the second component dose embedded deep implicit coding vector are subjected to deep-level feature fusion based on low-rank projection gating to obtain the inter-component deep-level fusion feature coding vector.

6. The deep learning-based dynamic proportioning system for soft candy functional ingredients according to claim 5, characterized in that: The component multi-scale fusion unit includes: An inter-component low-level middle-level gated fusion subunit, configured to perform gated weight fusion on the inter-component low-level fusion feature coding vector and the inter-component middle-level fusion feature coding vector to obtain an inter-component low-level middle-level fusion feature coding vector; The inter-component multi-scale attention fusion sub-unit is used to perform cross-attention fusion on the inter-component low-level and mid-level fusion feature coding vectors and the inter-component deep-level fusion feature coding vectors to obtain the inter-component multi-scale progressive perception fusion coding vectors.

7. The deep learning-based dynamic proportioning system for soft candy functional ingredients according to claim 6, characterized in that: The inter-component multi-scale attention fusion sub-unit is used to: Processing the inter-component low-level mid-level fusion feature encoding vector and the inter-component deep-level fusion feature encoding vector based on a pre-trained weight matrix to obtain an inter-component progressive perception query vector, an inter-component progressive perception key vector, and an inter-component progressive perception value vector; Performing domain-balanced alignment-based pre-training weight adaptive optimization on the pre-training weight matrix based on the inter-component progressive perception query vector, the inter-component progressive perception key vector, and the inter-component progressive perception value vector to obtain an optimized pre-training weight matrix; Based on the optimized pre-trained weight matrix, the inter-component low-level and mid-level fusion feature coding vector and the inter-component deep-level fusion feature coding vector are fused based on softmax to obtain the inter-component multi-scale progressive perception fusion coding vector.

8. The deep learning-based dynamic proportioning system for soft candy functional ingredients according to claim 7, characterized in that: The synergy prediction module is used to: input the inter-component multi-scale progressive perceptual fusion coding vector into an output head based on a fully connected layer to obtain the predicted synergy score.

Images