Color master batch formula optimization design method and system combined with deep learning

Abstract

This invention provides a method and system for optimizing color masterbatch formulations using deep learning. Relating to the field of deep learning technology, the method first obtains a basic color masterbatch material library and a set of color effect requirements for target plastic products. Then, it utilizes a formulation component association mapping model to mine the association mapping relationships between material effect components. Next, based on these association mapping relationships, it determines candidate color masterbatch material combinations and preliminary proportion ranges, generating an initial candidate formulation set. Then, it calls a proportion optimization deep neural network model to model and analyze the component interaction effects of the candidate combinations, generating a target optimized proportion parameter set. Finally, based on this target optimized proportion parameter set, it generates the final optimized color masterbatch formulation and outputs it to the production control system. This invention improves the efficiency and accuracy of formulation design, reduces costs, and enhances the quality of plastic products.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for optimizing the design of color masterbatch formulations by incorporating deep learning. Background Technology

[0002] As a key raw material for imparting color to plastic products, the formulation design of color masterbatch directly affects the color effect and quality of the final product. Traditional color masterbatch formulation design methods mainly rely on manual experience and extensive experimentation. Designers, based on their accumulated experience, select suitable materials from a limited pool of basic color masterbatch materials and repeatedly adjust the proportions to try and achieve the target color effect.

[0003] However, the above methods have many drawbacks. On the one hand, human experience has limitations; the experience levels of different designers vary, making it difficult to guarantee the accuracy and stability of the formulation design. Moreover, when faced with complex and diverse target color effect requirements, it is difficult to quickly find suitable material combinations and ratios based solely on experience. On the other hand, extensive experimentation not only consumes significant time, manpower, and material resources but also prolongs the product development cycle and reduces production efficiency. Furthermore, traditional methods struggle to fully consider the interaction effects between basic masterbatch materials, making it difficult to achieve optimal formulation design and meet the market's demand for diverse colors and high quality in plastic products. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for optimizing the design of color masterbatch formulations using deep learning, the method comprising: Obtain a basic masterbatch material library and a set of color effect requirements for target plastic products; Each basic masterbatch material in the basic masterbatch material library and each target plastic product color effect requirement in the target plastic product color effect requirement set are input into the formulation component association mapping model for component association mining processing, so as to obtain the material effect component association mapping relationship set between the basic masterbatch material library and the target plastic product color effect requirement set; The candidate masterbatch material combinations and their initial proportion ranges that meet the color effect requirements of each target plastic product are determined by the set of material effect component correlation mapping relationships, and an initial formula candidate set is generated. For each candidate masterbatch material combination and its initial ratio range in the initial formulation candidate set, the formulation parameters are optimized collaboratively. The deep neural network model for ratio optimization is called to model and analyze the component interaction effect of the candidate masterbatch material combination, and the target optimized ratio parameter set corresponding to each candidate masterbatch material combination is generated. Based on the target optimized ratio parameter set, the final optimized design formula of the masterbatch is generated, and the final optimized design formula of the masterbatch is output to the masterbatch production control system.

[0005] Furthermore, embodiments of the present invention also provide a color masterbatch formulation optimization design system combining deep learning, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned deep learning-integrated masterbatch formulation optimization design method by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described deep learning-integrated masterbatch formulation optimization design method.

[0007] Based on the above, by acquiring a basic masterbatch material library and a set of target plastic product color effect requirements, and utilizing a formulation component correlation mapping model to mine the correlation mapping relationships between material effect components, the intrinsic connection between basic masterbatch materials and target color effects can be analyzed. Based on the mined correlation mapping relationships, candidate masterbatch material combinations and their initial proportion ranges are determined, generating an initial formulation candidate set. A deep neural network model for proportion optimization is then invoked to model and analyze the component interaction effects of candidate masterbatch material combinations, enabling a deeper understanding of the complex relationships between materials, achieving synergistic optimization of formulation parameters, and generating an accurate set of target optimized proportion parameters. Finally, based on the target optimized proportion parameter set, the final optimized masterbatch design formulation is generated and output to the production control system. This significantly shortens the formulation design cycle, reduces R&D costs, and improves production efficiency, while ensuring that the masterbatch formulation meets the target color effect requirements and improves the quality of plastic products. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the color masterbatch formulation optimization design method combined with deep learning provided in the embodiments of the present invention.

[0009] Figure 2 This is a schematic diagram of exemplary hardware and software components of the masterbatch formulation optimization design system combining deep learning provided in an embodiment of the present invention. Detailed Implementation

[0010] Figure 1This is a flowchart illustrating a method for optimizing masterbatch formulations using deep learning, provided in one embodiment of the present invention. A detailed description follows.

[0011] Step S110: Obtain the basic masterbatch material library and the set of color effect requirements for the target plastic product.

[0012] In this embodiment, we take the production of a batch of automotive interior plastic parts with a specific deep-sea blue effect by a plastic product manufacturing company as a unified scenario. First, we need to establish the basic data source for formula design.

[0013] Step S111: Extract basic information records of all registered basic masterbatch materials from the masterbatch material database. The basic information records include the material name identifier, material type classification label, material physical morphology description information, and basic dispersion performance parameters of each basic masterbatch material.

[0014] For example, basic information records of all registered basic masterbatch materials are extracted from the enterprise's masterbatch material database. For instance, one material might have the material name identifier "Pigment Blue-15:3," the material type classification label "Organic Pigment / Phthalocyanine Blue," the material physical description information "Powder, D50 particle size = 0.8 micrometers, surface wetted," and the basic dispersion performance parameter "Dispersion Index DI = 95 (based on the rate of increase in filtration pressure of 100g standard resin in a specified extruder)." Another material, such as "Titanium Dioxide R-996," has the basic information record containing the material name identifier "Titanium Dioxide R-996," the material type classification label "Inorganic Pigment / Titanium Dioxide," the material physical description information "Powder, D50 particle size = 0.3 micrometers, surface coated with alumina," and the basic dispersion performance parameter "Dispersion Index DI = 98." All extracted records constitute a raw dataset.

[0015] Step S112: Perform material attribute feature encoding on the basic information records of each basic masterbatch material to generate a material attribute feature vector with the material name identifier as the index key value and the material type classification label, material physical morphology description information and basic dispersion performance parameters as attribute values. Associate the material attribute feature vector with the corresponding material name identifier to form a basic masterbatch material library.

[0016] Each extracted basic information record is encoded. Specifically, the material name identifier "Pigment Blue-15:3" is used as the index key. The material type classification label "Organic Pigment / Phthalocyanine Blue" is converted into a uniquely heated category vector through a predefined category-index mapping table. For example, if the vector dimension is 10, it indicates that there are 10 major material categories, with 1 for "Organic Pigment" and 0 for the rest. The "powder" morphology in the material physical morphology description information is represented by a binary variable, such as "1" for powder and "0" for non-powder. The D50 particle size value "0.8 micrometers" is used directly as a continuous numerical feature. The surface treatment information "wetting treatment" is converted into a multi-thermal label vector through a predefined feature dictionary, such as "wetting treatment" and "coating treatment". The basic dispersion performance parameter "DI=95" is used as a continuous numerical feature. All the encoded features, including category-specific thermal vectors, morphology binary variables, particle size values, surface treatment multi-thermal vectors, and dispersion index values, are concatenated end-to-end to form a complete material property feature vector, for example, a vector with 100 dimensions. This vector is then associated with the index key value "pigment blue-15:3" and stored to form a basic masterbatch material library. The same processing is applied to all basic masterbatch materials.

[0017] Step S113: Parse the desired color effect description text for each target plastic product from the plastic product design requirements document. The color effect description text includes color hue tendency words, color saturation degree modifiers, and color brightness range limiting words.

[0018] From the design requirements document for automotive interior components, extract the descriptive text of the desired color effect for the target plastic product. For example, for the front door trim panel, the descriptive text is "deep navy blue, high saturation, low brightness." For the center console, the descriptive text is "bright blue with a pearlescent effect, medium saturation, high brightness." In these texts, "navy blue" and "bright blue" are terms indicating color hue; "high saturation" and "medium saturation" are modifiers of color saturation level; and "low brightness" and "high brightness" are terms limiting the range of color brightness.

[0019] Step S114: Perform semantic component extraction processing on the color effect description text to extract color hue feature words that represent color hue tendency, color saturation feature words that represent color saturation degree modifiers, and color brightness feature words that represent color brightness range limit words. Combine and encode the color hue feature words, color saturation feature words, and color brightness feature words to generate the target plastic product color effect requirements that include color hue feature identifiers, color saturation level identifiers, and color brightness range identifiers.

[0020] Natural language processing techniques, such as attention-based sequence labeling models, are used to extract semantic components from text describing color effects. For the sentence "deep navy blue, high color saturation, low brightness," the model extracts the color hue feature term "navy blue," the color saturation feature term "high," and the color brightness feature term "low." These extracted terms are then mapped to a predefined color feature encoding system. For example, "navy blue" corresponds to a predefined color hue feature identifier "H-NAVY"; "high" corresponds to the color saturation level identifier "S-3" (assuming S-1 to S-3 are defined, with higher levels indicating higher saturation); and "low" corresponds to the color brightness range identifier "L-1" (assuming L-1 to L-3 are defined, with L-1 representing the low brightness range). These three identifiers are then combined and encoded, for example, by connecting them with a specific delimiter, to generate a structured target plastic product color effect requirement "H-NAVY_S-3_L-1."

[0021] Step S115: Index and store the color effect requirements of the target plastic products according to their corresponding plastic product design requirement document numbers to form a set of target plastic product color effect requirements containing multiple target plastic product color effect requirements.

[0022] The generated structured requirement “H-NAVY_S-3_L-1” is associated with the corresponding design requirement document number “DOC-FRONT-DOOR-001” and stored using this number as an index. Similarly, the effect requirement “H-BRIGHTBLUE_S-2_L-3” corresponding to “DOC-CENTER-CONSOLE-002” is also stored. All these stored requirement entries together constitute the set of color effect requirements for the target plastic product.

[0023] Step S116: Perform a vector dimension consistency check on the material property feature vectors of each basic color masterbatch material in the basic color masterbatch material library, and uniformly adjust the number of dimensions of the material property feature vectors to the preset standard number of dimensions of the material property feature vectors to obtain the material property feature vectors with unified dimensions.

[0024] To ensure consistency in input dimensions for subsequent model processing, the dimensionality of all material property feature vectors in the basic masterbatch material library is checked. It is assumed that the standard dimension of the material property feature vectors is 128. If a material's feature vector has 100 dimensions, dimension padding is required. For example, this is achieved by adding 28 placeholder features with values ​​of 0, expanding it to 128 dimensions. If a vector's dimension exceeds 128, principal component analysis or linear dimensionality reduction is performed to compress it to 128 dimensions, preserving the main information. Ultimately, all materials yield 128-dimensional material property feature vectors with unified dimensions.

[0025] Step S117: Standardize the color effect requirements of each target plastic product in the set of target plastic product color effect requirements. Reorganize the color hue feature identifier, color saturation level identifier, and color brightness range identifier of each target plastic product color effect requirement according to the preset requirement format to generate standardized color effect requirements with a unified format sequence structure.

[0026] Each requirement in the set of color effect requirements for the target plastic product is standardized in format. The preset requirement format arrangement order is "hue identifier_lightness identifier_saturation identifier". Therefore, "H-NAVY_S-3_L-1" generated in step S114 needs to be reorganized into "H-NAVY_L-1_S-3". "H-BRIGHTBLUE_S-2_L-3" is reorganized into "H-BRIGHTBLUE_L-3_S-2". After this processing, all requirements have a unified format sequence structure, generating standardized color effect requirements, which facilitates unified embedding encoding of the model.

[0027] Step S120: Input each basic masterbatch material in the basic masterbatch material library and each target plastic product color effect requirement in the target plastic product color effect requirement set into the formulation component association mapping model for component association mining processing, to obtain the material effect component association mapping relationship set between the basic masterbatch material library and the target plastic product color effect requirement set.

[0028] After constructing the input data, we begin to explore the intrinsic relationship between materials and effects.

[0029] Step S121: Use the material property feature vector of each basic masterbatch material in the basic masterbatch material library after unifying the dimensions as the source input data of the formulation component association mapping model, and use the normalized color effect requirements of each target plastic product color effect requirement in the target plastic product color effect requirement set as the target input data of the formulation component association mapping model.

[0030] The material property feature vectors obtained in step S116 with unified dimensions, such as the 128-dimensional vector corresponding to "Pigment Blue-15:3" and the 128-dimensional vector corresponding to "Titanium Dioxide R-996", are all used as the source inputs of the formulation component association mapping model. Simultaneously, the normalized color effect requirements generated in step S117, such as "H-NAVY_L-1_S-3" and "H-BRIGHTBLUE_L-3_S-2", are used as the target inputs of the model. The model inputs appear in pairs, meaning each material vector is paired with each effect requirement for calculation.

[0031] Step S122: The material property feature vector after dimension unification is processed by feature embedding encoding through the feature embedding layer of the formula component association mapping model to generate material embedding feature representation. At the same time, the normalized color effect requirement is processed by requirement embedding encoding to generate effect embedding feature representation.

[0032] The formulation component association mapping model first processes the original 128-dimensional material feature vector through its internal source-end feature embedding layer. This embedding layer is a multilayer perceptron network containing two fully connected layers. The first layer maps the 128 dimensions to 256 dimensions, and the second layer maps the 256 dimensions to the final 128-dimensional embedding space. The activation function is a linear rectified function. Through this nonlinear transformation, the original, potentially noisy features are converted into a more abstract and expressive material embedding feature representation. Simultaneously, the target-end feature embedding layer processes the text identifiers required for normalized color effects. Since the identifiers are structured, the model first uses an embedding lookup table to map discrete identifiers such as "H-NAVY", "L-1", and "S-3" into continuous vectors. For example, each identifier is mapped to a 64-dimensional vector. These three 64-dimensional vectors are then concatenated to form a 192-dimensional vector, which is finally compressed into a 128-dimensional embedded feature representation through a fully connected layer.

[0033] Step S123: Call the attention interaction module of the formulation component association mapping model to perform bidirectional attention interaction calculation on the material embedding feature representation and the effect embedding feature representation, and calculate the material effect attention weight coefficient between the material embedding feature representation of each basic color masterbatch material and the effect embedding feature representation of each target plastic product color effect requirement.

[0034] The 128-dimensional material embedding feature representation and 128-dimensional effect embedding feature representation generated in step S122 are input into the bidirectional attention interaction module. This module first calculates a similarity matrix; for example, for the embedding vector Mi of material i and the embedding vector Ej of effect j, their dot product is calculated to obtain a similarity value Sij. By calculating the similarity between all materials and all effects, a similarity matrix of shape [total materials N, total effects M] is obtained. Then, Softmax normalization is performed along the material dimension (i.e., each column) to obtain the attention weight matrix A_m2e for materials to effects, where the element A_m2e[i][j] represents the importance weight of the i-th material for the j-th effect. Simultaneously, Softmax normalization is performed along the effect dimension (i.e., each row) to obtain the attention weight matrix A_e2m for effects to materials. These weight coefficients quantify the association strength between each pair of materials and effects.

[0035] Step S124: Based on the material effect attention weight coefficient, select a set of candidate related materials from the basic masterbatch material library whose correlation with the color effect requirements of each target plastic product exceeds a preset correlation threshold. The set of candidate related materials includes multiple candidate related materials and their corresponding material effect attention weight coefficients.

[0036] For each target plastic product color effect requirement, such as "H-NAVY_L-1_S-3", examine the column vector corresponding to the effect calculated in step S123 (i.e., the corresponding column in A_m2e). Each element value in this column vector represents the correlation between the effect and the corresponding material. Set a preset correlation threshold, such as Gamma. Iterate through all N materials and filter out materials with an attention weight coefficient greater than Gamma. The filtered materials, such as "Pigment Blue-15:3" (weight coefficient 0.35) and "Phthalocyanine Blue-15:0" (weight coefficient 0.28), together with their respective weight coefficients, constitute a set of candidate correlated materials for the effect "H-NAVY_L-1_S-3". Perform the same filtering operation for each effect.

[0037] Step S125: Using the component relationship reasoning layer of the formulation component association mapping model, perform component relationship reasoning processing on the dimension-unified material property feature vector of each candidate associated material in the candidate associated material set and the effect embedding feature representation of the corresponding target plastic product color effect requirement, and generate the component contribution parameter between each candidate associated material and the corresponding target plastic product color effect requirement.

[0038] The original 128-dimensional material attribute feature vector of each candidate associated material (e.g., "Pigment Blue-15:3") selected in step S124, and the 128-dimensional effect feature representation of its corresponding target effect ("H-NAVY_L-1_S-3") are embedded for feature fusion. The fusion method can be element-wise multiplication followed by concatenation to form a 256-dimensional combined feature vector. This combined vector is input into the component relationship inference layer, which is a sub-network containing two fully connected layers and a Sigmoid output layer. The first fully connected layer reduces the 256 dimensions to 128 dimensions, and the second fully connected layer reduces the 128 dimensions to 1 dimension. Finally, the Sigmoid function outputs a scalar value between 0 and 1. This scalar value is the component contribution parameter of the material to the effect; for example, "Pigment Blue-15:3" contributes 0.85 to the effect "H-NAVY_L-1_S-3", and "Phthalocyanine Blue-15:0" contributes 0.60. The component contribution parameter quantitatively expresses the potential contribution of the material in achieving the specific color effect.

[0039] Step S126: Based on the component contribution parameters, construct a material effect component association mapping relationship between the color effect requirements of each target plastic product and multiple candidate associated materials. The material effect component association mapping relationship includes the target plastic product color effect requirement identifier, the candidate associated material identifier, and the corresponding component contribution parameters.

[0040] The calculation results from step S125 are stored in a structured manner, forming mapping relationship entries. For example, for the effect "H-NAVY_L-1_S-3", a mapping relationship is constructed as follows: {Effect Identifier: "H-NAVY_L-1_S-3", Candidate Material: [{Material Identifier: "Pigment Blue-15:3", Contribution: 0.85}, {Material Identifier: "Phthalocyanine Blue-15:0", Contribution: 0.60}]}. Each of the above entries represents a material effect component association mapping relationship.

[0041] Step S127: Group and store the constructed multiple material effect component association mapping relationships according to the target plastic product color effect requirement identifier, forming a set of material effect component association mapping relationships with the target plastic product color effect requirement as the index and the candidate associated material identifier and component contribution parameter as the mapping value.

[0042] All constructed mappings are grouped and stored according to their effect identifiers. For example, all mappings related to "H-NAVY_L-1_S-3" are grouped together, and all mappings related to "H-BRIGHTBLUE_L-3_S-2" are grouped together. This creates a key-value pair set indexed by effect, where the key is the effect identifier and the value is a list, with each element in the list being a mapping pair containing a material identifier and its contribution. This set is the material effect component association mapping set.

[0043] Step S128: Redundant mapping relationships are removed from each material effect component association mapping relationship in the set of material effect component association mapping relationships. The mapping relationships between candidate associated materials and the color effect requirements of the target plastic product with component contribution parameters lower than the preset contribution threshold are deleted. The mapping relationships with component contribution parameters reaching the preset contribution threshold are retained as valid material effect component association mapping relationships. The set of material effect component association mapping relationships between the basic color masterbatch material library and the set of color effect requirements of the target plastic product is obtained.

[0044] To streamline the mapping relationships and remove noisy materials with low contributions, a secondary filtering is performed on each mapping relationship entry in the above set. A preset contribution threshold, such as Delta, is set. For the mapping relationship list corresponding to the effect "H-NAVY_L-1_S-3", the contribution parameter of each candidate material is checked. Materials and their mapping relationships with a contribution below Delta (e.g., 0.65) are deleted. In this example, "Phthalocyanine Blue-15:0" has a contribution of 0.60, which is below the threshold, so it is removed from the mapping relationships. Ultimately, the effective mapping relationship for the effect "H-NAVY_L-1_S-3" only contains "Pigment Blue-15:3". After performing this filtering operation on all mapping relationships, the final, streamlined set of material effect component association mapping relationships is obtained. This set clearly reveals which materials are the key components for achieving a specific target color effect.

[0045] Step S130: Determine the candidate masterbatch material combinations and their preliminary proportion ranges that meet the color effect requirements of each target plastic product through the set of material effect component association mapping relationships, and generate an initial formula candidate set.

[0046] Based on the association mapping relationship obtained in the previous step, we begin to construct specific candidate formulation schemes.

[0047] Step S131: Analyze each valid material effect component association mapping relationship in the set of material effect component association mapping relationships, and extract multiple candidate associated material identifiers and their component contribution parameters corresponding to the color effect requirement identifier of each target plastic product.

[0048] First, the final mapping relationship set obtained in step S128 is parsed. For example, for the target effect identifier "H-NAVY_L-1_S-3", the candidate material identifier associated with it is extracted as "pigment blue-15:3", and its component contribution parameter is 0.85. Suppose another effect "H-BRIGHTBLUE_L-3_S-2" is associated with multiple materials, such as "titanium dioxide R-996" (contribution 0.92) and "ultramarine blue" (contribution 0.78).

[0049] Step S132: Sort multiple candidate associated material identifiers corresponding to the same target plastic product color effect requirement identifier according to their component contribution parameters from high to low, and generate a priority sorting list of candidate associated materials.

[0050] For the effect “H-BRIGHTBLUE_L-3_S-2”, the two materials associated with it are sorted according to their contribution, resulting in a priority list: [“Titanium Dioxide R-996” (0.92), “Ultramarine Blue” (0.78)]. For “H-NAVY_L-1_S-3”, which contains only a single material, the priority list is naturally [“Pigment Blue-15:3” (0.85)].

[0051] Step S133: Select a preset number of candidate associated material identifiers that are ranked at the top of the candidate associated material priority sorting list as the basic candidate material identifier set that meets the color effect requirements of the target plastic product.

[0052] Set a preset quantity K (e.g., K=2). For “H-BRIGHTBLUE_L-3_S-2”, select the first K=2 materials from its sorted list, namely “Titanium Dioxide R-996” and “Ultramarine Blue”, as the basic candidate material identifier set. For “H-NAVY_L-1_S-3”, since there is only 1 material, its basic candidate material identifier set is [“Pigment Blue-15:3”].

[0053] Step S134: Perform combination construction processing on the basic candidate material identifiers in the basic candidate material identifier set to generate a candidate color masterbatch material combination containing multiple candidate material combination methods. The candidate color masterbatch material combination contains multiple basic candidate material identifiers and their combination relationship descriptors.

[0054] The basic candidate material identifier set obtained in step S133 is combined and constructed. For "H-BRIGHTBLUE_L-3_S-2", its basic candidate material set contains 2 elements, and the possible combinations are: a single-component combination using only "Titanium Dioxide R-996", a single-component combination using only "Ultramarine Blue", and a two-component combination using both. For the two-component combination, a combination relationship descriptor needs to be defined, such as "mixing ratio to be determined". In this way, multiple candidate masterbatch material combinations for this effect are generated. For "H-NAVY_L-1_S-3", there is only one candidate combination, namely the single-component "Pigment Blue-15:3".

[0055] Step S135: Calculate the overall contribution comprehensive evaluation value of the candidate masterbatch material combination based on the component contribution parameters corresponding to each basic candidate material identifier in the candidate masterbatch material combination. The overall contribution comprehensive evaluation value is obtained by weighted summation of the component contribution parameters of all basic candidate material identifiers in the combination.

[0056] For each candidate masterbatch material combination, its overall contribution comprehensive evaluation value is calculated. For example, for the two-component combination "H-BRIGHTBLUE_L-3_S-2" consisting of "Titanium Dioxide R-996 + Ultramarine Blue", its overall contribution comprehensive evaluation value is equal to the weighted sum of the contribution of "Titanium Dioxide R-996" (0.92) and the contribution of "Ultramarine Blue" (0.78). The weighting coefficient can be simply set to 1, i.e., directly added together, resulting in an evaluation value of 1.70. For its single-component combination "Titanium Dioxide R-996", the evaluation value is 0.92.

[0057] Step S136: Select the candidate color masterbatch material combination whose overall contribution comprehensive evaluation value reaches the preset overall contribution threshold from all candidate color masterbatch material combinations as the target candidate material combination corresponding to the color effect requirements of the target plastic product.

[0058] A preset overall contribution threshold is set, such as Epsilon. For "H-BRIGHTBLUE_L-3_S-2", the evaluation values ​​of all its candidate combinations are compared. If the evaluation value of the single component "Ultramarine Blue" (0.78) is below the threshold, it is discarded. The evaluation values ​​of the single component "Titanium Dioxide R-996" (0.92) and the two-component combination (1.70) both reach or exceed the threshold, therefore these two combinations are selected as target candidate material combinations. For "H-NAVY_L-1_S-3", its unique combination has an evaluation value of 0.85 that exceeds the threshold, and it is also confirmed as a target candidate material combination.

[0059] Step S137: For each target candidate material combination, extract the historical ratio usage record of each basic candidate material in the historical formula database. The historical ratio usage record contains the ratio value of each basic candidate material in different historical formulas and its corresponding color effect achievement score.

[0060] For each selected candidate material combination, such as the two-component combination "Titanium Dioxide R-996 + Ultramarine Blue," historical formulation records of "Titanium Dioxide R-996" and "Ultramarine Blue" were retrieved from the historical formulation database. The retrieved records include: in historical formulation A, the formulation ratio of "Titanium Dioxide R-996" is P_A_Ti, with a corresponding color effect achievement score of S_A_Ti; in formulation B, the formulation ratio is P_B_Ti, with a score of S_B_Ti. Similarly, the formulation ratios of "Ultramarine Blue" in multiple historical formulations, P_A_Ul and P_B_Ul, along with their corresponding scores S_A_Ul and S_B_Ul, were retrieved. These scores are values ​​between 0 and 1, indicating the degree to which the product produced by the formulation closely approximates the expected color effect.

[0061] Step S138: Based on the distribution range of the ratio values ​​of each basic candidate material in the historical ratio usage record and the corresponding color effect achievement score, determine the preliminary feasible ratio range of each basic candidate material in the target candidate material combination. The lower limit of the preliminary feasible ratio range is the smallest ratio value in the historical ratio values ​​where the color effect achievement score exceeds the preset score threshold, and the upper limit of the preliminary feasible ratio range is the largest ratio value in the historical ratio values ​​where the color effect achievement score exceeds the preset score threshold.

[0062] Set a preset scoring threshold, such as T_score (assumed to be 0.85). For "Titanium Dioxide R-996", filter all historical mixing records to find those with a color effect achievement score greater than 0.85. Among these records, find the corresponding mixing ratio values, take the minimum value as the lower limit L_Ti of the initial feasible mixing range, and take the maximum value as the upper limit U_Ti. This gives the initial feasible mixing range [L_Ti, U_Ti] for "Titanium Dioxide R-996". Similarly, perform the same operation for "Ultramarine Blue" to obtain its initial feasible mixing range [L_Ul, U_Ul].

[0063] Step S139: Link and store each target candidate material combination and its corresponding multiple basic candidate materials in the preliminary feasible range of proportions to form an initial formulation candidate entry containing the target candidate material combination identifier and multiple preliminary feasible ranges of proportions.

[0064] The target candidate material combination determined in step S136 (e.g., combination identifier "COMBO-BLUE-001", representing "titanium dioxide R-996 + ultramarine blue") is associated with the preliminary feasible range of proportions for each material ("titanium dioxide R-996", "ultramarine blue") under this combination determined in step S138. This constructs an initial formulation candidate entry: {combination identifier: "COMBO-BLUE-001", material range: [{material: "titanium dioxide R-996", range: [L_Ti, U_Ti]}, {material: "ultramarine blue", range: [L_Ul, U_Ul]}]}.

[0065] Step S1310: Aggregate all initial formula candidate entries corresponding to the color effect requirements of the target plastic products according to the color effect requirement identifier of the target plastic products to generate an initial formula candidate set.

[0066] The initial formula candidate entries generated for all target effects (such as "H-NAVY_L-1_S-3" and "H-BRIGHTBLUE_L-3_S-2") are aggregated according to their respective effect identifiers. For example, all candidate entries related to "H-BRIGHTBLUE_L-3_S-2" (which may include entries for the single component "Titanium Dioxide R-996" and entries for the two component "Titanium Dioxide R-996 + Ultramarine Blue") are aggregated together to form a subset indexed by that effect. Finally, the complete set containing all effects and their corresponding formula candidate entries constitutes the initial formula candidate set.

[0067] Step S140: Perform collaborative optimization of formulation parameters for each candidate masterbatch material combination and its initial ratio range in the initial formulation candidate set, call the ratio optimization deep neural network model to model and analyze the component interaction effect of the candidate masterbatch material combination, and generate the target optimized ratio parameter set corresponding to each candidate masterbatch material combination.

[0068] After obtaining the initial candidate formulation, more precise optimization is needed, taking into account the interactions between components.

[0069] Step S141: Extract the basic candidate material identifiers and the lower and upper limits of the corresponding preliminary feasible range of each candidate masterbatch material combination from the initial formulation candidate set.

[0070] Taking the candidate combination "COMBO-BLUE-001" in step S139 as an example, the list of basic candidate material identifiers included in it is extracted from the initial formulation candidate set: ["Titanium Dioxide R-996", "Ultramarine Blue"]. At the same time, the preliminary feasible range of the ratio corresponding to each material is extracted: for "Titanium Dioxide R-996", its lower limit value L_Ti and upper limit value U_Ti are extracted; for "Ultramarine Blue", its lower limit value L_Ul and upper limit value U_Ul are extracted.

[0071] Step S142: Perform one-hot encoding on the basic candidate material identifiers to generate a material identifier feature vector corresponding to each basic candidate material identifier. At the same time, perform numerical feature encoding on the lower limit and upper limit of the preliminary feasible range of proportions to generate a proportion range feature vector containing the lower limit numerical feature and the upper limit numerical feature.

[0072] The extracted material identifiers are encoded using unique thermal encoding. Assuming the entire material library contains 1000 materials, each material identifier corresponds to a 1000-dimensional unique thermal vector, where only the position corresponding to that material is 1, and the rest are 0. For example, the unique thermal vector for "Titanium Dioxide R-996" is 1 at index i, and the unique thermal vector for "Ultramarine Blue" is 1 at index j. For the initial feasible range of proportions, the lower limit value L_Ti and the upper limit value U_Ti are treated as two separate numerical features and directly concatenated to form a 2-dimensional proportion range feature vector [L_Ti, U_Ti]. A similar 2-dimensional vector [L_Ul, U_Ul] is generated for "Ultramarine Blue".

[0073] Step S143: Perform feature concatenation processing on the material identification feature vector and the ratio range feature vector to generate a combined input feature vector corresponding to each basic candidate material.

[0074] The unique thermal encoding vector (1000-dimensional) of each material is concatenated with its proportion range feature vector (2-dimensional). For "Titanium Dioxide R-996", the concatenation results in a 1002-dimensional combined input feature vector V_Ti. Similarly, for "Ultramarine Blue", a 1002-dimensional combined input feature vector V_Ul is also formed.

[0075] Step S144: Stack the combined input feature vectors corresponding to all basic candidate materials in the candidate masterbatch material combination according to the preset combined feature input order to generate the combined feature input tensor corresponding to the candidate masterbatch material combination.

[0076] For a combination of two materials, "COMBO-BLUE-001", its combination input feature vectors V_Ti and V_Ul are stacked in a preset order (e.g., ascending order by material ID). The stacking can be done by splicing the vectors in a new dimension, forming a two-dimensional tensor of shape [2, 1002]. This tensor is the combination feature input tensor corresponding to the candidate color masterbatch material combination.

[0077] Step S145: Input the combined feature input tensor into the feature extraction layer of the ratio optimization deep neural network model to perform component feature extraction processing, and generate a component feature map of the candidate color masterbatch material combination. The component feature map includes the local feature response of each basic candidate material in the combination and the interactive feature response between multiple basic candidate materials.

[0078] A combined feature input tensor of shape [2, 1002] is fed into the feature extraction layer of a proportioning optimization deep neural network model. This feature extraction layer is a one-dimensional convolutional neural network. This one-dimensional convolutional neural network contains multiple convolutional kernels, each with a size of, for example, 3 and a stride of 1. As the convolutional kernels slide along the 1002-dimensional feature dimension, they can extract local patterns of each material's own features. More importantly, since the first dimension of the input tensor is the material dimension, the convolution operation is actually performed on the same feature dimension of different materials. For example, it can combine the k-th feature value of "Titanium Dioxide R-996" with the k-th feature value of "Ultramarine Blue" to capture the interaction features between materials. After multiple convolution and pooling operations, the model outputs a component feature map with a shape of, for example, [2, F], where F is the feature dimension after convolution extraction. The feature vector of each material in this map has been fused with its own features and the interaction information with other materials learned through the weight sharing of the convolutional kernels.

[0079] Step S146: The component feature map is processed by modeling the component interaction effect through the interaction modeling layer of the ratio optimization deep neural network model, and the material interaction coefficient between any two basic candidate materials in the candidate masterbatch material combination and the synergistic enhancement coefficient between multiple basic candidate materials are calculated.

[0080] The component feature map obtained in step S145 is input into the interaction modeling layer. This interaction modeling layer adopts a graph attention network structure, which treats each material as a node in the graph, and the initial features of the node are the F-dimensional feature vectors of the corresponding material in the map. The graph attention network learns the interaction by calculating the attention coefficients between nodes. Specifically, for materials i and j, it first applies a shared linear transformation to their feature vectors, and then calculates their attention coefficient alpha_ij, which reflects the importance of material j to material i. The updated node features are obtained by calculating alpha_ij for all j and then summing them with weights. In this process, the model can output the learned attention coefficient matrix, where the off-diagonal elements alpha_ij (i ≠ j) can be regarded as the material interaction coefficients between material i and material j. For combinations containing more than two materials, a scalar value can be obtained by aggregating multiple pairwise interaction coefficients or by using higher-order graph convolution operations as the synergistic enhancement coefficient between multiple materials. This synergistic enhancement coefficient quantifies the effect beyond simple summation when multiple materials coexist.

[0081] Step S147: Input the material interaction coefficient and synergistic enhancement coefficient into the ratio optimization prediction layer of the ratio optimization deep neural network model, and combine the ratio range feature vector in the combined feature input tensor to perform ratio parameter prediction processing, thereby generating the initial optimized ratio parameter set corresponding to the candidate masterbatch material combination.

[0082] The interaction coefficients (e.g., a 2x2 matrix) and synergy enhancement coefficients (a scalar) calculated in step S146 are flattened or appropriately transformed, and then concatenated with the proportion range feature vector (2D) in the combined feature input tensor from step S144 to form a comprehensive feature vector. This vector is then input into the proportion optimization prediction layer. This proportion optimization prediction layer is a fully connected network with multi-task learning, and the number of nodes in its output layer is equal to the number of materials in the combination. For example, for a combination containing two materials, the output layer has two nodes. The activation function of the network can be a linear function or a constrained sigmoid function (multiplied by a maximum proportion range). For the input candidate combinations, the model outputs a set of specific proportion values, such as "Titanium Dioxide R-996:PTi_opt1, Ultramarine Blue:PUl_opt1". By running the model multiple times (e.g., using different Dropouts or multiple forward propagations), multiple sets of different optimized ratio values ​​can be obtained. These values ​​together constitute the initial optimized ratio parameter set, such as {[PTi_opt1, PUl_opt1], [PTi_opt2, PUl_opt2], ...}.

[0083] Step S148: Verify the rationality constraint of each initial optimized ratio parameter in the initial optimized ratio parameter set, delete the initial optimized ratio parameters that do not meet the preset rationality constraint, and retain the initial optimized ratio parameters that meet the preset rationality constraint as candidate optimized ratio parameters.

[0084] Set a set of reasonable ratio constraints, such as: the ratio of each material must be between 0% and 100%, and the sum of the ratios of all materials must be 100% (or a baseline value). Verify each set of optimized ratio parameters generated in step S147. For example, if the ratio of "Titanium Dioxide R-996" in a certain set of parameters is -5%, or the sum of 30% of "Titanium Dioxide R-996" and 80% of "Ultramarine Blue" is 110%, then the set of parameters is determined to not meet the constraints and is deleted. Only parameter sets where all values ​​are within a reasonable range and the sum meets the requirements, such as [30%, 70%], are retained as candidate optimized ratio parameters.

[0085] Step S149: Perform cluster analysis on the candidate optimized ratio parameters, merge candidate optimized ratio parameters with similar values ​​into the same ratio parameter cluster, and select the candidate optimized ratio parameter corresponding to the cluster center point from each ratio parameter cluster as the representative ratio parameter.

[0086] After the screening in step S148, a large number of candidate optimal ratio parameters may still remain. To reduce redundancy and provide the most representative solution, cluster analysis is performed on the above parameters. Taking a two-component formulation as an example, each parameter [PTi, PUl] is considered as a point on a two-dimensional plane. The K-Means clustering algorithm is used, and the number of clusters K_cluster is set (e.g., K_cluster=3). The algorithm divides all points into 3 clusters. Each cluster has a cluster center (i.e., the mean of all points in the cluster). Then, the candidate optimal ratio parameter point closest to the cluster center is selected from each cluster as the representative ratio parameter of that cluster. For example, three representative ratio parameters are finally obtained: [30%, 70%], [32%, 68%], and [28%, 72%].

[0087] Step S1410: The representative ratio parameters are associated and stored according to their corresponding candidate masterbatch material combination identifiers to generate a target optimized ratio parameter set containing multiple representative ratio parameters for each candidate masterbatch material combination.

[0088] The three representative proportioning parameters obtained in step S149 are associated and stored with the candidate masterbatch material combination identifier “COMBO-BLUE-001” to which they belong. The generated target optimized proportioning parameter set is: {combination identifier: “COMBO-BLUE-001”, optimized proportioning set: [[30%, 70%], [32%, 68%], [28%, 72%]]}. This set provides several optimized, high-quality formulation options that take into account the interactions between components.

[0089] Step S150: Generate the final optimized design formula for color masterbatch based on the target optimized ratio parameter set, and output the final optimized design formula for color masterbatch to the color masterbatch production control system.

[0090] After obtaining the optimized parameter set, the final executable formula is generated based on the actual production situation.

[0091] Step S151: parse each target optimization ratio parameter entry in the target optimization ratio parameter set, and extract the candidate masterbatch material combination identifier and the corresponding representative ratio parameter contained in each target optimization ratio parameter entry.

[0092] The target optimized ratio parameter set obtained in step S1410 is analyzed. The candidate masterbatch material combination identifier "COMBO-BLUE-001" and its corresponding representative ratio parameter set are extracted, which are three two-dimensional arrays.

[0093] Step S152: Based on the candidate masterbatch material combination identifier, retrieve the corresponding basic candidate material identifier list in the initial formulation candidate set, and associate each basic candidate material identifier in the basic candidate material identifier list with the representative proportion parameter to generate a preliminary formulation component correspondence.

[0094] Based on the combination identifier "COMBO-BLUE-001", the process traces back to the initial candidate formulation set (step S139) and retrieves the list of basic candidate material identifiers corresponding to this combination as ["Titanium Dioxide R-996", "Ultramarine Blue"]. Then, these two material identifiers are associated with the first representative proportion parameter [30%, 70%] extracted in step S151 to generate a preliminary formulation component correspondence: {Material: "Titanium Dioxide R-996", Proportion: 30%} and {Material: "Ultramarine Blue", Proportion: 70%}. The same correspondence is generated for the other two sets of representative proportion parameters.

[0095] Step S153: Perform material attribute verification processing on each basic candidate material identifier in the preliminary formulation component correspondence to confirm the available inventory status and material batch information of the material corresponding to each basic candidate material identifier in the current masterbatch production control system.

[0096] For the material identifier "Titanium Dioxide R-996" in the preliminary formulation component correspondence, a query request is sent to the inventory management module of the masterbatch production control system. The system returns the available inventory status of this material, such as "Sufficient Inventory" or "Tight Inventory, 500 kg Remaining". Simultaneously, it returns the batch information of the material currently available for production, such as "Batch Number: Ti-20231021-01". The same query is performed on "Ultramarine Blue" to obtain its inventory status and batch information, such as "Batch Number: Ul-20231015-03".

[0097] Step S154: Extract the actual performance parameters of the material corresponding to each basic candidate material identifier from the material batch database according to the material batch information, calculate the matching degree between the actual performance parameters and the representative proportion parameters, and obtain the matching degree coefficient between the material batch and the proportion parameters.

[0098] Based on the batch information "Ti-20231021-01", the actual performance parameters of this batch of materials are extracted from a specialized material batch database. These parameters may include the measured values ​​of the dominant wavelength, tinting strength, and dispersibility of the pigment in this batch. For example, the actual tinting strength of this batch of "Titanium Dioxide R-996" is 98% of the nominal value. The matching degree between the representative proportion parameter 30% and this actual tinting strength 98% is calculated. The calculation method can be to multiply the proportion parameter by a correction factor determined by the actual performance parameter. For example, correction factor = actual tinting strength / 100% = 0.98, then the matching degree coefficient is 0.98. Ideally, the matching degree coefficient is 1, indicating that the performance is completely consistent with the standard value. The calculated matching degree coefficient of this batch of materials with the 30% proportion is 0.98. Similarly, the matching degree coefficient of the "Ultramarine Blue" batch is calculated, assumed to be 1.02.

[0099] Step S155: The basic candidate material identifiers and their corresponding representative ratio parameters that reach the preset matching degree threshold are designated as the formulation component entries to be confirmed, and the basic candidate material identifiers and their corresponding representative ratio parameters that have a matching degree coefficient lower than the preset matching degree threshold are designated as the formulation component entries to be adjusted.

[0100] Set a preset matching threshold, such as 0.95. For the first representative ratio parameter group, the matching degree of "Titanium Dioxide R-996" is 0.98, which is greater than 0.95. Therefore, it and its ratio of 30% are listed as a component to be confirmed in the formulation. Similarly, the matching degree of "Ultramarine Blue" is 1.02, which is also greater than 0.95, and it is also listed as a component to be confirmed. If the matching degree of a material is lower than 0.95, for example, if the performance of a batch of material deteriorates significantly, resulting in a matching degree of only 0.85, then that material and its ratio will be marked as a component to be adjusted in the formulation.

[0101] Step S156: For the basic candidate material identifier marked as the component item to be adjusted, retrieve its list of available alternative materials, select the alternative material identifier with the highest matching degree with the representative ratio parameter from the list of available alternative materials, replace the original basic candidate material identifier with the alternative material identifier, and adjust the representative ratio parameter corresponding to the alternative material identifier to the alternative material suitable ratio parameter.

[0102] Suppose that in another formulation, the batch matching degree of "Ultramarine Blue" is below a threshold and is marked as needing adjustment. In this case, the system retrieves a list of available alternative materials for "Ultramarine Blue," such as "Alternative Blue A" and "Alternative Blue B." It queries the material batch database for the actual performance parameters of the currently available batches of these two alternative materials and calculates their matching degree with the original representative proportion parameter of 70%. Assuming the matching degree of "Alternative Blue A" is 0.98 and that of "Alternative Blue B" is 0.95, "Alternative Blue A," with the highest matching degree, is selected as the alternative material. Then, the material identifier in the original formulation is replaced with "Alternative Blue A." Simultaneously, based on the actual performance parameters of "Alternative Blue A," the original proportion of 70% is adjusted to generate the appropriate proportion parameter for the alternative material. For example, if the tinting strength of "Alternative Blue A" is 105% of the standard value, to achieve the same color effect as the original formulation, the proportion may need to be adjusted to 70%. (100 / 105)≈66.7%. This adjusted value is the appropriate ratio parameter for the alternative materials.

[0103] Step S157: Merge all unconfirmed formulation component entries and adjusted formulation component entries to generate a final formulation component entry list containing multiple formulation component entries. Each formulation component entry includes a material identifier, material batch information, and the corresponding optimized ratio value.

[0104] All confirmed or adjusted component entries are combined. For the successfully matched examples above, the combined list of component entries forms the final formulation: Entry 1: {Material Identifier: "Titanium Dioxide R-996", Batch: "Ti-20231021-01", Proportion: 30%}; Entry 2: {Material Identifier: "Ultramarine Blue", Batch: "Ul-20231015-03", Proportion: 70%}. Any adjusted entries will also be included.

[0105] Step S158: Perform sum normalization processing on the optimized ratio values ​​in the final formula component entry list, adjust the sum of the optimized ratio values ​​to the preset total weight benchmark value of the formula, and generate the normalized optimized ratio values.

[0106] Assume the preset total weight baseline value of the formula is 100 kg. Add all the optimized proportion values ​​in the list of step S157 to get a total of 30% + 70% = 100%. Since the sum is already 100%, the normalized values ​​remain unchanged, still 30 kg and 70 kg. If the sum is not 100%, for example, because after adjustment it becomes 30% and 66.7%, and the sum is 96.7%, then it needs to be scaled proportionally: Normalized value = Original value / Sum Benchmark value. For example, 30% becomes 30 / 96.7. 100 ≈ 31.0 kg, 66.7% becomes 66.7 / 96.7 100 ≈ 69.0 kg.

[0107] Step S159: The material identifier, material batch information, and normalized optimized ratio values ​​are packaged according to a preset formula data format to generate the final optimized design formula for the color masterbatch.

[0108] The above information is encapsulated according to the format required by the production control system. For example, a JSON-formatted formula file is generated, containing the fields "Formula ID" and "Material List". The Material List is an array, and each element is an object containing "Material Code" (material identifier), "Batch Number" (material batch information), and "Weight" (normalized optimized ratio value). This generates the final optimized design formula for the masterbatch, which can be directly used for production.

[0109] Step S1510: The final optimized design formula of the masterbatch is transmitted to the formula receiving module of the masterbatch production control system through the data interface, triggering the masterbatch production control system to perform formula configuration operations such as material weighing, mixing and extrusion granulation according to the optimized ratio value.

[0110] The packaged final formula file is sent to the formula receiving module of the masterbatch production control system via the factory's industrial Ethernet. Upon receiving the formula, the production control system automatically parses the file and sends the material code, batch number, and weight to the automatic weighing system. The weighing system accurately weighs the material according to the instructions and then feeds the weighed material into a high-speed mixer for mixing. The uniformly mixed material is then conveyed to a twin-screw extruder, where it undergoes melting, dispersion, extrusion, and granulation under set process parameters, ultimately producing a masterbatch product that achieves the desired deep-sea blue effect.

[0111] Next, the construction process of the key model involved in the method of this invention will be described in detail.

[0112] Step S210: Construction process of formulation component association mapping model.

[0113] To ensure that the model used in step S120 can accurately uncover the relationship between materials and effects, the model needs to be built and trained in advance.

[0114] Step S211: Collect historical masterbatch formulation design case data, which includes the basic masterbatch material identification used in each historical case, the corresponding target plastic product color effect requirements, and the color effect achievement evaluation score of the actually produced masterbatch products.

[0115] Historical formulation design case data from the past five years were collected from the company's formulation and quality inspection databases. Each case is a record; for example, case number C001 uses basic masterbatch materials identified as "Pigment Blue-15:3" and "Titanium Dioxide R-996," with a target plastic product color effect requirement of "H-NAVY_L-1_S-3." The masterbatch product actually produced in this case, after colorimeter testing, showed a color difference DeltaE of 0.8 compared to the standard color chart. According to predefined mapping rules (e.g., DeltaE less than 1.0 is excellent), its color effect achievement score was marked as 0.95. All collected cases constitute the original dataset for model training.

[0116] Step S212: Extract the material property feature vector corresponding to the basic masterbatch material identifier from each historical masterbatch formulation design case data as the source input sample for model training, extract the standardized color effect requirements corresponding to the target plastic product color effect requirements as the target input sample for model training, and extract the color effect achievement evaluation score as the supervision label data for model training.

[0117] For case C001, based on the material identifiers "Pigment Blue-15:3" and "Titanium Dioxide R-996", the material property feature vectors with unified dimensions are retrieved from the basic masterbatch material library (as constructed in step S112), such as 128-dimensional vectors V_PB and V_Ti. For formulations containing multiple materials, the above vectors need to be aggregated, for example, by averaging or weighted averaging, to form a source input sample vector V_C001_src representing the entire formulation. The target input sample is the representation after normalizing the target effect requirement "H-NAVY_L-1_S-3" (as in step S117), such as a 128-dimensional vector V_C001_tgt transformed by the embedding layer. The supervision label data is the color effect achievement evaluation score of 0.95 for this case. This process is repeated for all historical cases to construct a large number of (source input, target input, supervision label) triples required for model training.

[0118] Step S213: Construct the initial formula component association mapping neural network structure, which includes a source-end feature embedding layer, a target-end feature embedding layer, a bidirectional attention interaction layer, and a component relationship inference output layer.

[0119] Construct a neural network model. The source-side feature embedding layer is a multilayer perceptron that maps the input 128-dimensional aggregate material vector to a new 128-dimensional embedding space. The target-side feature embedding layer is similar, mapping the input 128-dimensional effect vector to the same 128-dimensional embedding space. The bidirectional attention interaction layer, as described in step S123, calculates the attention weights between the two sets of embedding representations. The component relation inference output layer, as described in step S125, is a fully connected network that maps the fused features (e.g., source-side features weighted by attention) to a scalar score, which represents the degree of correlation between the input recipe and the input effect, i.e., the degree of prediction achievement.

[0120] Step S214: Input the source-end input samples of the model training into the source-end feature embedding layer to perform source-end feature embedding learning, and obtain the source-end embedded feature representation. Input the target-end input samples of the model training into the target-end feature embedding layer to perform target-end feature embedding learning, and obtain the target-end embedded feature representation.

[0121] During training, a batch of source input samples V_src_batch is input into the source feature embedding layer, and a batch of source embedded feature representations E_src_batch is calculated through forward propagation. At the same time, a corresponding batch of target input samples V_tgt_batch is input into the target feature embedding layer, and a batch of target embedded feature representations E_tgt_batch is obtained.

[0122] Step S215: Input the source-end embedded feature representation and the target-end embedded feature representation into the bidirectional attention interaction layer to calculate the attention weights, generating the attention weight matrix of the source-end to the target-end and the attention weight matrix of the target-end to the source-end.

[0123] E_src_batch and E_tgt_batch are input into the bidirectional attention interaction layer. This bidirectional attention interaction layer first calculates the similarity matrix S_batch (through dot product or additive attention), and then performs Softmax normalization on the source and target dimensions respectively to obtain the attention weight matrices A_src2tgt_batch and A_tgt2src_batch. These matrices capture the attention of each source sample to each target sample within the batch, and vice versa.

[0124] Step S216: Input the attention weight matrix of the source end to the target end and the attention weight matrix of the target end to the source end into the component relationship inference output layer to perform component relationship inference calculation, and output the predicted correlation score between each source end input sample and each target end input sample.

[0125] The component relationship inference output layer uses an attention weight matrix to weight and aggregate the embedded features. For example, A_src2tgt_batch can be used to weight the target-side embeddings to obtain a context vector for each source sample that incorporates target information. This context vector is concatenated with the source-side embeddings and input into a fully connected network, ultimately outputting a scalar value, which is the predicted correlation score between each source-side input sample (historical recipe) and each target-side input sample (historical effect). For pairwise training data, we are interested in the score corresponding to the same case, such as the score S_pred calculated from the source-side input sample and the target-side input sample for case C001.

[0126] Step S217: Construct a model loss function based on the difference between the predicted correlation score and the color effect achievement evaluation score, and use the backpropagation algorithm to update the network weight parameters of the formula component correlation mapping neural network structure until the model loss function converges, thus obtaining the trained formula component correlation mapping model.

[0127] Construct a loss function, such as the mean squared error loss function, and calculate the difference between the predicted score S_pred and the true supervision label (e.g., 0.95). Using the Adam optimizer, compute the gradient of the loss function with respect to all weight parameters in the network via backpropagation and update these parameters. Iterate the training for dozens of epochs on a dataset containing tens of thousands of historical cases until the model's loss function value on the validation set no longer decreases, i.e., the model converges. At this point, the model parameters are fixed, resulting in the trained recipe component association mapping model.

[0128] Step S220: The process of constructing a deep neural network model with optimized proportions.

[0129] Similarly, the proportioning optimization model used in step S140 will be described in detail.

[0130] Step S221: Collect multiple sets of color masterbatch formulation experimental data. The color masterbatch formulation experimental data includes the basic color masterbatch material combination identifier used in each set of experiments, the actual ratio value of each basic color masterbatch material, and the measured index of the color effect of the color masterbatch products obtained in the experiment.

[0131] A systematic formulation experiment was designed. For example, a series of experiments with different ratios were designed for the combination of "titanium dioxide R-996 + ultramarine blue". Experimental group E001: material combination identifier "COMBO-BLUE-001", actual ratio value is "titanium dioxide R-996 20%, ultramarine blue 80%". The color masterbatch samples obtained in the experiment were tested, and their measured color effect index is a multi-dimensional vector, such as L in CIELab color space. a b The values, along with tinting strength, hiding power, etc., form a measured index vector, for example, 10 dimensions. Hundreds of the above experiments were conducted, covering different material combinations and ratios.

[0132] Step S222: Perform material identification encoding on the basic masterbatch material combination identifiers in each group of masterbatch formulation experimental data to generate a material combination identifier feature vector. Perform ratio normalization on the actual ratio values ​​of each basic masterbatch material to generate a ratio value feature vector.

[0133] For experimental group E001, the material combination identifier "COMBO-BLUE-001" is first encoded. A graph neural network approach can be used, treating the combination as a graph where nodes represent materials, and the initial features of the nodes can be one-heat encodings or attribute vectors of the materials. However, to simplify the input, an embedding representation of the material combination can be pre-trained to obtain a fixed-dimensional material combination identifier feature vector. The actual proportion values ​​"20% and 80%" are normalized to ensure their sum is 1 or 100, generating a 2-dimensional proportion value feature vector.

[0134] Step S223: Perform feature fusion processing on the material combination identification feature vector and the ratio numerical feature vector to generate the experimental input feature tensor corresponding to each set of experimental data, and use the measured color effect index of the masterbatch product obtained in the experiment as the experimental output label data.

[0135] The material combination identifier feature vector (e.g., 64-dimensional) and the ratio numerical feature vector (2-dimensional) are concatenated to form a 66-dimensional experimental input feature vector. These vectors from all experimental groups are then stacked to form the experimental input feature tensor. Simultaneously, the measured color effect index vector (e.g., 10-dimensional) corresponding to each experimental group is collected as the experimental output label data. This constructs the training dataset for the model.

[0136] Step S224: Construct an initial ratio optimization deep neural network structure, which includes a feature extraction convolutional layer, an interaction modeling graph neural network layer, a ratio optimization prediction fully connected layer, and an output layer.

[0137] Construct an end-to-end neural network model. The feature extraction convolutional layer uses a one-dimensional convolutional neural network, as described in step S145, to extract initial features from the input vector. The interaction modeling graph neural network layer uses a graph attention network, as described in step S146, to explicitly model the interactions between materials. The proportioning optimization prediction fully connected layer is a multilayer perceptron used to map the interaction features to the proportioning parameter space. The output layer is a fully connected layer with the number of neurons equal to the number of target proportioning parameters (i.e., the number of materials in the combination), used to output the final optimized proportioning parameters.

[0138] Step S225: Input the experimental input feature tensor into the feature extraction convolutional layer for local feature extraction processing to obtain a local feature response map. Input the local feature response map into the interaction modeling neural network layer for material interaction relationship modeling processing to obtain the interaction feature representation.

[0139] During training, a batch of experimental input feature tensors are fed into the model. First, a feature extraction convolutional layer outputs a local feature response map. This local feature response map is then fed into an interaction modeling graph neural network layer. In this layer, the model updates the features of each node through a message-passing mechanism based on the learned graph structure (which can be a fully connected graph or a graph constructed based on material similarity), ultimately obtaining an interaction feature representation that integrates information about the node itself and its neighbors.

[0140] Step S226: Input the interaction feature representation into the fully connected layer for nonlinear mapping to obtain the initial ratio optimization prediction value. Input the initial ratio optimization prediction value into the output layer for dimensional transformation to obtain the model prediction output with the same dimension as the experimental output label data.

[0141] The interaction features are represented as input proportions in a fully connected layer for optimal prediction. Through multiple nonlinear transformations, this prediction reaches the output layer. The output layer has two neurons and outputs a 2-dimensional vector, representing the predicted proportions for the experimental group, for example, [22%, 78%]. This predicted value is the model's estimate of the optimal proportions corresponding to the input feature tensors of that experimental group.

[0142] Step S227: Construct a model training loss function based on the difference between the model's predicted output and the experimental output label data. Use the gradient descent optimization algorithm to update the network weight parameters of the ratio-optimized deep neural network structure until the model training loss function value is reduced to below the preset loss threshold, thus obtaining the trained ratio-optimized deep neural network model.

[0143] A loss function is constructed, such as the mean squared error between the predicted proportion vector and the actual proportion vector used in the experiment. However, a more reasonable loss function could be to input the predicted proportion vector into a simulator (or another pre-trained color prediction model), predict its color effect, and then minimize the difference between this predicted color effect and the measured color effect in the experiment. This constitutes a model-based optimization. Gradient descent algorithms, such as Adam, are used to update the weights of the entire network. Iterative training is performed until the loss function value decreases below a preset threshold, for example, the average prediction error is less than 1%. At this point, the model training is complete, and it can predict the precise proportion that produces the optimal color effect for a given combination of materials and a preliminary range.

[0144] Finally, the supplementary steps of the method of the present invention are described in detail, which constitute a complete closed-loop optimization system.

[0145] Step S310: Collect in real time the measured color effect data of the actual color masterbatch product obtained after production according to the final optimized color masterbatch design formula from the color masterbatch production control system.

[0146] After the product is produced in step S1510, the quality inspection system automatically performs online color detection on the finished products. For example, a colorimeter measures the color of the particles after extrusion granulation every 10 seconds and records the measured L... a b The value is transmitted back to the central data server in real time through the data interface.

[0147] Step S320: Compare and analyze the color effect difference between the measured color effect data of the actual color masterbatch product and the corresponding color effect requirements of the target plastic products in the set of color effect requirements for the target plastic products, and calculate the color difference offset between the measured color effect data and the color effect requirements.

[0148] The system will collect the measured L a b The values ​​(e.g., [25.3, -8.1, -15.2]) are compared with the standard Lab values ​​(e.g., [25.0, -8.5, -15.0]) corresponding to the target color effect requirement "H-NAVY_L-1_S-3". The Euclidean distance between the two in the Lab color space is calculated, i.e., the color difference offset DeltaE=sqrt((25.3-25.0)^2+(-8.1+8.5)^2+(-15.2+15.0)^2).

[0149] Step S330: Determine the degree of compliance of the actual color effect of the final color masterbatch optimized design formula based on the color difference offset, and generate an evaluation index for the compliance of the actual effect of the formula.

[0150] The color difference offset DeltaE is mapped to a compliance assessment index. For example, if DeltaE is less than 0.5, the compliance score is 1.0; if DeltaE is between 0.5 and 1.0, the compliance score is (1.5 - DeltaE) / 1.0; and if DeltaE is greater than 1.5, the compliance score is 0. Assuming a calculated DeltaE of 0.45, the compliance assessment index is 1.0.

[0151] Step S340: Link and store the actual effect compliance evaluation index of the formula with the final color masterbatch optimized design formula to form a formula effect history record entry.

[0152] The compliance evaluation index 1.0 is linked with the previously generated final masterbatch optimized design formula (including information such as materials, batches, and proportions) to form a new historical record entry, which is then stored in the formula effect database.

[0153] Step S350: When the actual effect compliance evaluation index of the formula is lower than the preset compliance threshold, extract the material identifier and optimized ratio value in the corresponding final color masterbatch optimized design formula, and combine the material identifier and optimized ratio value with the collected color effect test data to form a new formula optimization training sample.

[0154] Set a preset compliance threshold, such as 0.9. If the compliance evaluation index of a certain formula is 0.8, which is lower than this threshold, the system will automatically trigger data collection. Extract the material identifiers (such as "Titanium Dioxide R-996", "Ultramarine Blue") and optimized proportions (such as 30%, 70%) from the formula, as well as the actual color effect measurement data of the corresponding batch collected from the production line (such as L...). a b (Value). Combine this set of data into a new training sample, with the same format as the experimental data in step S221.

[0155] Step S360: Add the new formula optimization training samples to the training dataset of the ratio optimization deep neural network model, trigger incremental update training of the ratio optimization deep neural network model, and generate the updated ratio optimization deep neural network model.

[0156] This new, substandard training sample from actual production is added to the existing training dataset of the ratio optimization deep neural network model. Then, starting with the current model parameters, incremental updates are performed in small batches. This helps the model learn why the original ratio fails under new production conditions or material batches, and how to adjust it, thus enabling the model to continuously evolve and adapt to changes in production.

[0157] For example, step S410: receiving real-time production process parameters returned by the color masterbatch production control system when producing according to the final color masterbatch optimized design formula, the real-time production process parameters include the actual value of material weighing, the actual value of mixing time, the actual value of extrusion temperature, and the actual value of granulation speed.

[0158] During production, the production control system collects and reports key process parameters in real time. For example, the set weighing value for "titanium dioxide R-996" in the formula is 30 kg, while the actual weighing sensor reading is 30.05 kg. The set mixing time for the mixer is 5 minutes, while the actual mixing time is 5.02 minutes. The set temperature for zone one of the extruder is 180 degrees Celsius, while the actual temperature is 182 degrees Celsius. The set rotation speed for the granulator is 300 rpm, while the actual rotation speed is 298 rpm.

[0159] Step S420: Calculate the parameter deviation between the real-time parameters of the production process and the optimized ratio values ​​in the final masterbatch optimized design formula to obtain the material weighing deviation value, mixing time deviation value, extrusion temperature deviation value, and granulation speed deviation value.

[0160] Calculate the deviation between the actual value and the set value for each parameter. Material weighing deviation = Actual value - Set value = 30.05 - 30 = 0.05 kg. Mixing time deviation = 5.02 - 5 = 0.02 minutes. Extrusion temperature deviation = 182 - 180 = 2 degrees Celsius. Granulation speed deviation = 298 - 300 = -2 revolutions per minute. Use these deviation values ​​as input for subsequent analysis.

[0161] Step S430: Input the material weighing deviation value, mixing time deviation value, extrusion temperature deviation value, and granulation speed deviation value into the production process stability assessment model for stability assessment processing, and generate a comprehensive production process stability score.

[0162] Construct a production process stability assessment model, which can be a rule-based scoring system or a trained regression model. For example, normalize each deviation value and input it into a weighted summation formula. Assume the weights of each deviation are 0.4, 0.1, 0.25, and 0.25, respectively. First, normalize the deviation values: normalized value of weighing deviation = 0.05 / (preset tolerance, e.g., 0.1) = 0.5. Similarly, normalized value of mixing time = 0.02 / 0.1 = 0.2, normalized value of temperature = 2 / 5 = 0.4, normalized value of rotational speed = |-2| / 5 = 0.4. Then calculate the comprehensive score = 1 - (0.4) / 5. 0.5 + 0.1 0.2 + 0.25 0.4 + 0.25 0.4) = 1 - (0.2 + 0.02 + 0.1 + 0.1) = 0.58.

[0163] Step S440: When the overall stability score of the production process is lower than the preset stability threshold, a production process parameter adjustment suggestion is generated based on the material weighing deviation value, mixing time deviation value, extrusion temperature deviation value, and granulation speed deviation value.

[0164] The preset stability threshold is set to 0.7. The overall score of 0.58 is below the threshold, indicating an unstable production process. The system generates adjustment suggestions based on the deviation values: for positive material weighing deviations, it is recommended to calibrate the weighing hopper or check the feeding speed; for positive extrusion temperature deviations, it is recommended to slightly reduce the heater power or check the cooling water flow rate; for negative granulation speed deviations, it is recommended to check the motor drive or belt tension. Specific suggestions can be a standardized operating guide code.

[0165] Step S450: Send the production process parameter adjustment suggestions to the parameter adjustment module of the color masterbatch production control system, triggering the color masterbatch production control system to automatically adjust the material weighing, mixing time, extrusion temperature and granulation speed according to the production process parameter adjustment suggestions.

[0166] The generated adjustment suggestions are sent as commands to the parameter adjustment module of the production control system. For example, a command might be issued to lower the set temperature of zone one of the extruder by 2 degrees Celsius, or to increase the speed setpoint of the granulator by 2 revolutions per minute. Upon receiving the command, the control system automatically modifies the setpoints of the process parameters to bring the production process back to a stable state, thereby ensuring the consistency of the final product.

[0167] For example, step S510: Obtain microstructure characterization data of the trial production masterbatch sample obtained from the masterbatch production control system after the first batch of trial production based on the final masterbatch optimized design formula. The microstructure characterization data includes cross-sectional scanning electron microscope images of the masterbatch sample, X-ray diffraction patterns of material crystallinity, and thermogravimetric analysis curves of material thermal stability.

[0168] Before mass production of the formulation, small-batch pilot production may be conducted first, and the microstructure of the pilot production samples may be analyzed. For example, a small amount of the pilot production masterbatch may be taken and fractured in liquid nitrogen, and the cross-section may be photographed using a scanning electron microscope to obtain a high-resolution SEM image. At the same time, powder samples may be taken for X-ray diffraction analysis to obtain diffraction patterns. Then, thermogravimetric analysis may be performed to obtain a TGA curve showing the change in sample mass with temperature.

[0169] Step S520: Perform image segmentation processing on the cross-sectional scanning electron microscope image to extract the actual spatial distribution morphology characteristics of the basic masterbatch material in the plastic product matrix from the masterbatch sample. The actual spatial distribution morphology characteristics include the material agglomerate size distribution parameters, the material dispersion uniformity index in the matrix, and the porosity of the material-matrix interface.

[0170] The SEM image is input into an image segmentation neural network, such as U-Net. This image segmentation neural network separates the pigment particles, matrix, and interface voids in the image. Then, image analysis is performed on the segmentation results: the equivalent diameter of all pigment particles is calculated to obtain aggregate size distribution parameters (such as D10, D50, D90); the distribution variance of particles in the image is calculated to obtain the dispersion uniformity index; and the percentage of the total area of ​​the interface void region to the total area of ​​the image is calculated to obtain the interface porosity.

[0171] Step S530: Perform spectrum analysis on the X-ray diffraction pattern of the material's crystallinity, extract the position shift of the crystallization peak, the change in the full width at half maximum (FWHM) of the crystallization peak, and the percentage value of crystallinity of the basic masterbatch material in the masterbatch sample, and calculate the crystal transformation coefficient of the material during the processing based on the position shift of the crystallization peak and the change in the FWHM of the crystallization peak.

[0172] Baseline correction and peak fitting were performed on the XRD pattern. The 2θ values ​​of the main crystalline peaks were extracted and compared with standard cards to obtain the peak position shift. The full width at half maximum (FWHM) of the peaks was extracted and compared with standard values ​​to obtain the FWHM change. The ratio of the crystalline peak area to the amorphous peak area was calculated to obtain the crystallinity percentage. The crystal form transformation coefficient can be calculated by weighting the peak position shift and the FWHM change, for example, coefficient = w1. Offset + w2 Change in half-width at half-height.

[0173] Step S540: Extract the curve feature points from the thermogravimetric analysis curve of the material's thermal stability, extract the initial decomposition temperature, maximum weight loss rate temperature, and residual mass percentage of the masterbatch sample, and construct a comprehensive evaluation index of the material's thermal stability based on the initial decomposition temperature, maximum weight loss rate temperature, and residual mass percentage.

[0174] The first derivative of the TGA curve is used to obtain the DTG curve. The temperature at which the mass begins to decrease by 5% is determined from the TGA curve as the initial decomposition temperature. The temperature corresponding to the peak value on the DTG curve is identified as the temperature of the maximum rate of weight loss. The percentage of residual mass at the end of the experiment (e.g., 800℃) is recorded. These three indicators are normalized, for example, multiplied by different weights, and then summed to obtain a comprehensive thermal stability evaluation index between 0 and 1.

[0175] Step S550: Standardize the aggregate size distribution parameters, dispersion uniformity index, interfacial porosity, crystal transformation degree coefficient, and thermal stability comprehensive evaluation index of the material to obtain a set of standardized microstructure performance parameters. Input these parameters into the microstructure performance correlation analysis model of the formulation for correlation analysis to generate the comprehensive evaluation value of the microstructure performance of the final masterbatch optimized design formulation.

[0176] The extracted parameters (e.g., aggregate D50, uniformity index, porosity, crystal transformation coefficient, and thermal stability index) are constructed into a multidimensional vector. This vector is then Z-score standardized to have a mean of 0 and a variance of 1. The standardized vector is then input into a pre-trained microstructure-performance correlation analysis model (e.g., a support vector regression machine or neural network). This model outputs a scalar value, representing the comprehensive microstructure-performance evaluation value, which reflects the overall performance of the formulation at the microscopic level.

[0177] Step S560: Compare the comprehensive evaluation value of microstructure performance with the preset target value of microstructure performance. When the comprehensive evaluation value of microstructure performance is lower than the preset target value of microstructure performance, extract the area in the cross-sectional scanning electron microscope image where the size of the material agglomerates exceeds the preset agglomerate size threshold as the defect area, and mark the basic masterbatch material identifier corresponding to the defect area.

[0178] The preset microstructure performance target value is set to 0.8. If the comprehensive evaluation value is 0.6, which is lower than the target value, the microstructure is determined to have defects. Returning to the SEM image segmentation results of step S520, all particles with aggregate sizes exceeding a preset threshold (e.g., 5 micrometers) are identified, and the areas containing these particles are marked as defect areas. Then, based on the particle morphology, energy dispersive spectroscopy, and other information in the image, the material corresponding to these defect areas is identified as either "titanium dioxide R-996" or "ultracyan blue," and its material identifier is marked.

[0179] Step S570: Based on the identifier of the basic masterbatch material corresponding to the defect area, retrieve the optimized ratio value of the basic masterbatch material from the final optimized design formula of the masterbatch, compare the optimized ratio value with the theoretical optimal dispersion concentration range of the material in the formula, and calculate the degree of deviation between the optimized ratio value and the theoretical optimal dispersion concentration range.

[0180] Assume the material identifier corresponding to the defect area is "Titanium Dioxide R-996". The optimized proportion is retrieved from the final formulation and found to be 30%. The theoretical optimal dispersion concentration range of "Titanium Dioxide R-996" in this resin matrix is ​​then searched from a material database, for example, [20%, 25%]. The deviation of 30% from this range is calculated. For example, if 30% is greater than the upper limit of 25%, then the deviation = (30-25) / 25 = 0.2, or 20%.

[0181] Step S580: When the optimized ratio value deviates from the theoretical optimal dispersion concentration range by more than a preset deviation threshold, a microstructure defect warning message for the formulation is generated. The microstructure defect warning message includes the basic masterbatch material identifier corresponding to the defect area, the optimized ratio value, the theoretical optimal dispersion concentration range, and the deviation value.

[0182] The preset deviation threshold is set to 10%. The calculated deviation of 20% exceeds 10%, therefore a microstructure defect warning message is generated. This microstructure defect warning message includes: defective material: "Titanium Dioxide R-996"; current ratio: 30%; theoretical range: [20%, 25%]; deviation degree: 20%.

[0183] Step S590: Send the microstructure defect warning information of the formula to the formula optimization design terminal, triggering the formula optimization design terminal to adjust the optimized ratio value of the corresponding basic color masterbatch material in the final color masterbatch optimization design formula according to the defect warning information, and generate the modified color masterbatch formula after microstructure optimization.

[0184] The warning information is sent to the formulation engineer's terminal or the automated optimization system. Upon receiving the warning, the system can adjust the formulation according to preset rules (e.g., adjusting the ratio to the midpoint of the theoretical range) or by invoking a more refined optimization model. For example, the ratio of "Titanium Dioxide R-996" might be reduced from 30% to 22.5%. Simultaneously, to maintain the overall hue, the ratio of "Ultramarine Blue" or other additives may need to be fine-tuned. After these adjustments, a new modified masterbatch formulation designed to improve the microstructure is generated.

[0185] Step S5100: The modified masterbatch formula with optimized microstructure is re-output to the masterbatch production control system, triggering the masterbatch production control system to perform a second trial production operation based on the modified masterbatch formula, and repeatedly acquiring the microstructure characterization data of the second trial production masterbatch samples until the comprehensive evaluation value of microstructure performance reaches the preset microstructure performance target value.

[0186] The revised formula is then sent back to the production control system for a new round of trial production. Steps S510 to S560 are then repeated to perform microstructure analysis on the new samples and calculate the comprehensive evaluation value. If the evaluation value still does not meet the standard, the analysis, adjustment, and retesting continue, forming a closed-loop optimization cycle, until the comprehensive evaluation value of the microstructure performance reaches or exceeds the preset target value of 0.8. At this point, the obtained formula has achieved ideal results in both macroscopic color effect and microstructure uniformity.

[0187] For example, after step S120, the method further includes: After obtaining preliminary material effect correlations through the formulation component correlation mapping model, in order to further improve the long-term usability and stability of the formulation, it is necessary to conduct in-depth analysis and correction of the physical properties of the materials and their impact on color stability. This sequence of steps is designed for this purpose.

[0188] Step S610: Obtain the component contribution parameter corresponding to each material effect component association mapping relationship in the set of material effect component association mapping relationships. The component contribution parameter includes the positive contribution component and negative interference component of each candidate associated material to the color effect requirements of the target plastic product.

[0189] First, each mapping entry is parsed from the set of material effect component association mapping relationships obtained in step S128. For the target effect identifier "H-NAVY_L-1_S-3" and its candidate associated material "Pigment Blue-15:3", its component contribution parameter is a multi-dimensional vector, not a single numerical value. In this embodiment, the contribution parameter consists of a triple: {positive contribution component: P_contrib, negative interference component: N_interf, overall confidence level: C_conf}. For example, for "Pigment Blue-15:3", its positive contribution component P_contrib = 0.82, indicating the degree of dominant contribution of the material to achieving the target navy blue phase; the negative interference component N_interf = 0.12, indicating the degree of negative impact such as color shift or poor dispersion that the material may bring; and the overall confidence level C_conf = 0.95, indicating the reliability of the contribution assessment. These components are calculated through the multi-output architecture of the component relationship inference layer in step S125.

[0190] Step S620: Extract the material physical morphology description information from the material property feature vector of each candidate associated material. The material physical morphology description information includes the particle size distribution range of the material particles, the surface morphology feature parameters of the material particles, and the fluidity index of the material in the molten state.

[0191] Next, detailed physical morphological descriptions of each candidate associated material are extracted from the basic masterbatch material library. Taking "Pigment Blue-15:3" as an example, its material property feature vector includes: particle size distribution parameters, such as D10=0.3 μm, D50=0.8 μm, D90=1.5 μm; particle surface morphology parameters, which is a multi-dimensional vector obtained through image analysis, such as sphericity=0.85, roughness index=0.2, specific surface area=12 m² / g; and the material's flowability index in the molten state, a value measured by a capillary rheometer, such as melt index MI=15 g / 10 min (under specific temperature and load conditions). These parameters are extracted completely.

[0192] Step S630: Calculate the spatial distribution uniformity estimate of each candidate associated material in the plastic product matrix based on the particle size distribution range of the material particles, and construct a prediction model of the interfacial bonding strength of the material in the plastic product matrix based on the surface morphology characteristic parameters of the material particles and the fluidity index of the material in the molten state.

[0193] Based on the extracted physical parameters, further calculations and modeling are performed. First, according to the particle size distribution parameters, especially D50 and distribution width (e.g., (D90-D10) / D50), the spatial distribution uniformity prediction value U_pred is calculated using an empirical formula or a trained regression model. This prediction value is a value between 0 and 1, with a higher value indicating that the material is more easily and uniformly dispersed in the matrix. For example, U_pred = f(D50, distribution width), where f is a function fitted based on a large amount of experimental data.

[0194] Simultaneously, a model for predicting interfacial bonding strength is constructed. The model's input consists of material particle surface morphology parameters (e.g., sphericity, roughness) and the melt flow index (MI). The model's output is a predicted interfacial bonding strength value, S_pred. This model can employ a multilayer perceptron structure, comprising an input layer (number of nodes equal to the input feature dimension, e.g., 5), a hidden layer (10 nodes, activation function: linear rectified function), and an output layer (1 node, activation function: sigmoid function, mapping the output value to between 0 and 1). The model's weight parameters are obtained through pre-collected datasets containing measured interfacial bonding strength values ​​for different materials.

[0195] Step S640: The interfacial bonding strength of each candidate associated material in the plastic product matrix is ​​predicted using the interfacial bonding strength prediction model to obtain the predicted value of the interfacial bonding strength between the material and the matrix. The predicted value of spatial distribution uniformity and the predicted value of interfacial bonding strength are standardized respectively. The standardized predicted value of spatial distribution uniformity and the standardized predicted value of interfacial bonding strength are fused together to generate a comprehensive physical compatibility representation vector for each candidate associated material.

[0196] The interface bonding strength prediction model constructed in step S630 is applied. For "Pigment Blue-15:3", its surface morphology characteristic parameters and fluidity index are input into the model. After forward propagation calculation, the predicted interface bonding strength value S_pred is obtained at the output layer, assumed to be 0.75. At the same time, its spatial distribution uniformity prediction value U_pred is assumed to be 0.85. Then, the Z-score standardization is performed on the two values ​​U_pred and S_pred respectively. Standardization requires the mean μ_U, μ_S and standard deviation σ_U, σ_S of these two indicators obtained statistically from a large amount of material datasets. For example, U_norm=(U_pred-μ_U) / σ_U, S_norm=(S_pred-μ_S) / σ_S. Finally, the standardized U_norm and S_norm are concatenated to form a 2-dimensional physical adaptability comprehensive characterization vector V_phy=[U_norm, S_norm]. This vector sum represents the degree of physical compatibility between the material and the matrix.

[0197] Step S650: Input the physical compatibility comprehensive characterization vector of each candidate associated material and the positive contribution component and negative interference component in the corresponding component contribution parameter into the material effect stability analysis module for joint analysis and processing, and calculate the color effect stability retention coefficient of each candidate associated material during long-term use.

[0198] The physical adaptability comprehensive characterization vector V_phy (2D) obtained in step S640 is concatenated with the positive contribution component P_contrib (scalar) and the negative interference component N_interf (scalar) from the component contribution parameters extracted in step S610 to form a 4D comprehensive feature vector V_stability_input=[P_contrib, N_interf, U_norm, S_norm]. This vector is then input into the material effect stability analysis module. This module is a pre-trained regression model, such as a gradient boosting decision tree or a support vector regression machine. The model is trained using a large amount of historical material color retention rate data from accelerated aging experiments as training labels. For "Pigment Blue-15:3", V_stability_input is input into the model, and the model outputs a scalar value K_stab, which is the color effect stability retention coefficient, assumed to be 0.92. This coefficient represents the expected proportion of the material's initial color effect that it can maintain under long-term use or aging conditions.

[0199] Step S660: Perform stability weighting correction processing on each material effect component association mapping relationship in the set of material effect component association mapping relationships according to the color effect stability retention coefficient, and multiply the component contribution parameter corresponding to the candidate associated material in each material effect component association mapping relationship by the corresponding color effect stability retention coefficient to generate the stability weighted component contribution parameter.

[0200] The original component contribution parameters were corrected using the calculated stability retention coefficient. The original overall contribution of "Pigment Blue-15:3" (e.g., its positive contribution component P_contrib = 0.82) was multiplied by the stability retention coefficient K_stab = 0.92 to obtain the stability-weighted overall contribution P_stab_weighted = 0.82. 0.92 = 0.7544. Similarly, if its negative interference components also need to be weighted, they can be treated in the same way. This correction process reflects the consideration that "although a material has a high initial contribution, its long-term effectiveness will be reduced if its stability is poor."

[0201] Step S670: Replace the component contribution parameter in the raw material effect component association mapping relationship with the stability-weighted component contribution parameter to generate the stability-corrected material effect component association mapping relationship set.

[0202] The calculated P_stab_weighted=0.7544 from step S660 is used to replace the comprehensive contribution parameter (e.g., originally 0.85) corresponding to "pigment blue-15:3" in the original mapping relationship in step S610. This replacement operation is performed on each candidate material in all mapping relationships in the material effect component association mapping relationship set, thereby generating a completely new mapping relationship set containing long-term material stability information, i.e., the stability-corrected material effect component association mapping relationship set.

[0203] Step S680: Extract multiple candidate associated materials and their stability-weighted component contribution parameters corresponding to the color effect requirements of each target plastic product from the set of material effect component association mapping relationships after stability correction. Normalize the stability-weighted component contribution parameters to generate a normalized stability contribution weight distribution.

[0204] Taking the target effect "H-BRIGHTBLUE_L-3_S-2" as an example, the candidate materials after stability correction might include "Titanium Dioxide R-996" (stability-weighted contribution 0.90) and "Ultramarine Blue" (stability-weighted contribution 0.70). To more clearly compare the relative importance of the materials, these weighted contributions are normalized. The total is calculated as 0.90 + 0.70 = 1.60. Therefore, after normalization, the weight of "Titanium Dioxide R-996" is 0.90 / 1.60 = 0.5625, and the weight of "Ultramarine Blue" is 0.70 / 1.60 = 0.4375. This yields a normalized stability contribution weight distribution, reflecting the proportion of weight each material accounts for in achieving the desired effect after considering stability.

[0205] Step S690: Based on the normalized stability contribution weight distribution, reorder the importance of multiple candidate associated materials corresponding to the color effect requirements of each target plastic product, and generate a stability-weighted priority ranking list of candidate associated materials.

[0206] The materials are reordered based on their normalized weights. For the effect "H-BRIGHTBLUE_L-3_S-2", the ranking list is updated to: ["Titanium Dioxide R-996" (weight 0.5625), "Ultramarine Blue" (weight 0.4375)]. This new ranking list is the stability-weighted priority list of candidate associated materials, which better reflects the actual application value of the materials than the unweighted ranking (which may only consider the initial contribution).

[0207] Step S6100: Compare and analyze the priority ranking list of candidate related materials after stability weighting with the priority ranking list of candidate related materials before unweighting, identify candidate related materials whose ranking position increases by more than a preset increase threshold as stability-enhancing candidate materials, and identify candidate related materials whose ranking position decreases by more than a preset decrease threshold as stability-weakening candidate materials.

[0208] The stability-weighted ranking list generated in step S690 is compared with the original ranking list generated in step S132. Assume that in the original ranking list, "Ultramarine Blue" is ranked before "Titanium Dioxide R-996", i.e., ["Ultramarine Blue", "Titanium Dioxide R-996"]. After stability weighting, the ranking becomes ["Titanium Dioxide R-996", "Ultramarine Blue"]. "Titanium Dioxide R-996" rises from 2nd to 1st place, an increase of 1 place. If the preset threshold for this increase is set to 1, the material reaches the threshold and is identified as a candidate material with enhanced stability, indicating excellent long-term stability. "Ultramarine Blue" falls from 1st to 2nd place, a decrease of 1 place. If the preset threshold for this decrease is also 1, it is identified as a candidate material with weakened stability, indicating relatively insufficient stability.

[0209] Step S6110: Extract features from the physical morphology description information of the stability-enhancing candidate materials to construct a stability-enhancing material feature template; simultaneously extract features from the physical morphology description information of the stability-weakening candidate materials to construct a stability-weakening material feature template.

[0210] For the identified stability-enhancing material "Titanium Dioxide R-996", its detailed physical morphology description information (including particle size distribution, surface morphology, flowability index, etc.) is extracted to form a feature template, such as a vector T_enhanced containing all relevant parameters. This template represents the typical physical morphological characteristics of a material with "good stability". Similarly, for the stability-weakening material "Ultramarine Blue", its physical morphology information is also extracted to construct a feature template T_weakened, representing the typical characteristics of a material with "poor stability".

[0211] Step S6120: Store the stability-enhancing material feature template and the stability-weakening material feature template in the material stability feature library.

[0212] Finally, the constructed feature templates T_enhanced and T_weakened, along with their corresponding material identifiers and stability classification labels ("enhanced" / "weakened"), are stored in a dedicated material stability feature library. This feature library can be used for the rapid screening and evaluation of new materials in the future: when a new material is added to the library, its physical morphological characteristics can be compared with the T_enhanced and T_weakened templates in the library to quickly predict its potential color effect stability.

[0213] Based on the same inventive concept, please refer to Figure 2The diagram shows a schematic block diagram of a color masterbatch formulation optimization design system 100 combined with deep learning provided in an embodiment of this application. The color masterbatch formulation optimization design system 100 combined with deep learning may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0214] In this embodiment, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 stores machine-executable instructions for executing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the inspection video stream processing method provided in the aforementioned method embodiments.

[0215] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for optimizing the design of color masterbatch formulations using deep learning, characterized in that, The method includes: Obtain a basic masterbatch material library and a set of color effect requirements for target plastic products; Each basic masterbatch material in the basic masterbatch material library and each target plastic product color effect requirement in the target plastic product color effect requirement set are input into the formulation component association mapping model for component association mining processing, so as to obtain the material effect component association mapping relationship set between the basic masterbatch material library and the target plastic product color effect requirement set; The candidate masterbatch material combinations and their initial proportion ranges that meet the color effect requirements of each target plastic product are determined by the set of material effect component correlation mapping relationships, and an initial formula candidate set is generated. For each candidate masterbatch material combination and its initial ratio range in the initial formulation candidate set, the formulation parameters are optimized collaboratively. The deep neural network model for ratio optimization is called to model and analyze the component interaction effect of the candidate masterbatch material combination, and the target optimized ratio parameter set corresponding to each candidate masterbatch material combination is generated. Based on the target optimized ratio parameter set, the final optimized design formula of the masterbatch is generated, and the final optimized design formula of the masterbatch is output to the masterbatch production control system.

2. The method for optimizing masterbatch formulations using deep learning as described in claim 1, characterized in that, The acquisition of the basic masterbatch material library and the set of color effect requirements for the target plastic products includes: Extract basic information records of all registered basic masterbatch materials from the masterbatch material database. The basic information records include the material name identifier, material type classification label, material physical morphology description information, and basic dispersion performance parameters of each basic masterbatch material. The basic information records of each basic masterbatch material are processed by material attribute feature encoding to generate a material attribute feature vector with the material name identifier as the index key value, and the material type classification label, material physical morphology description information and basic dispersion performance parameters as attribute values. The material attribute feature vector is associated with the corresponding material name identifier to form a basic masterbatch material library. The desired color effect description text for each target plastic product is extracted from the design requirements document of the plastic product. The color effect description text includes words related to color hue tendency, words modifying the degree of color saturation, and words limiting the range of color brightness. The text describing the color effect is processed by semantic component extraction to extract color hue feature words that represent color hue tendency, color saturation feature words that represent color saturation degree modifiers, and color brightness feature words that represent color brightness range limit words. The color hue feature words, color saturation feature words, and color brightness feature words are combined and encoded to generate the target plastic product color effect requirements that include color hue feature identifiers, color saturation level identifiers, and color brightness range identifiers. The color effect requirements of the target plastic products are indexed and stored according to their corresponding plastic product design requirement document numbers, forming a set of target plastic product color effect requirements containing multiple target plastic product color effect requirements. Perform a vector dimension consistency check on the material property feature vectors of each basic masterbatch material in the basic masterbatch material library, and uniformly adjust the number of dimensions of the material property feature vectors to the preset standard number of dimensions of the material property feature vectors to obtain the material property feature vectors with unified dimensions. The color effect requirements of each target plastic product in the set of target plastic product color effect requirements are standardized in terms of requirement format. The color hue feature identifier, color saturation level identifier, and color brightness range identifier of each target plastic product color effect requirement are rearranged in a preset requirement format order to generate standardized color effect requirements with a unified format sequence structure.

3. The method for optimizing masterbatch formulations using deep learning as described in claim 1, characterized in that, The process involves inputting each basic masterbatch material in the basic masterbatch material library and each target plastic product color effect requirement in the target plastic product color effect requirement set into a formulation component association mapping model for component association mining, thereby obtaining a set of material effect component association mapping relationships between the basic masterbatch material library and the target plastic product color effect requirement set, including: The material property feature vector of each basic masterbatch material in the basic masterbatch material library after dimensional unification is used as the source input data of the formulation component association mapping model, and the normalized color effect requirement of each target plastic product color effect requirement in the target plastic product color effect requirement set is used as the target input data of the formulation component association mapping model. The material property feature vector after dimension unification is processed by feature embedding encoding through the feature embedding layer of the formula component association mapping model to generate material embedding feature representation. At the same time, the standardized color effect requirements are processed by requirement embedding encoding to generate effect embedding feature representation. The attention interaction module of the formulation component association mapping model is invoked to perform bidirectional attention interaction calculation on the material embedding feature representation and the effect embedding feature representation, and to calculate the material effect attention weight coefficient between the material embedding feature representation of each basic masterbatch material and the effect embedding feature representation of each target plastic product color effect requirement. Based on the material effect attention weight coefficient, a set of candidate related materials that have a correlation degree exceeding a preset correlation degree threshold with the color effect requirements of each target plastic product are selected from the basic masterbatch material library. The set of candidate related materials includes multiple candidate related materials and their corresponding material effect attention weight coefficients. The component relationship reasoning layer of the formulation component association mapping model is used to perform component relationship reasoning processing on the material property feature vector of each candidate associated material in the candidate associated material set after dimensional unification and the effect embedding feature representation of the corresponding target plastic product color effect requirement, so as to generate the component contribution parameter between each candidate associated material and the corresponding target plastic product color effect requirement; Based on the component contribution parameters, a material effect component association mapping relationship is constructed between the color effect requirements of each target plastic product and multiple candidate associated materials. The material effect component association mapping relationship includes the target plastic product color effect requirement identifier, the candidate associated material identifier, and the corresponding component contribution parameters. The constructed multiple material effect component association mapping relationships are grouped and stored according to the color effect requirements of the target plastic product, forming a set of material effect component association mapping relationships with the color effect requirements of the target plastic product as the index and the candidate associated material identifiers and component contribution parameters as the mapping values. Redundant mapping relationships are removed from each material effect component association mapping relationship in the set of material effect component association mapping relationships. The mapping relationships between candidate associated materials and the color effect requirements of the target plastic product with component contribution parameters lower than the preset contribution threshold are deleted. The mapping relationships with component contribution parameters reaching the preset contribution threshold are retained as valid material effect component association mapping relationships. The set of material effect component association mapping relationships between the basic color masterbatch material library and the set of color effect requirements of the target plastic product is obtained.

4. The method for optimizing masterbatch formulations using deep learning as described in claim 1, characterized in that, The process involves determining candidate masterbatch material combinations and their initial proportioning ranges that meet the color effect requirements of each target plastic product through the set of material effect component correlation mapping relationships, generating an initial formula candidate set, including: Analyze each valid material effect component association mapping relationship in the set of material effect component association mapping relationships, and extract multiple candidate associated material identifiers and their component contribution parameters corresponding to the color effect requirement identifier of each target plastic product; Multiple candidate associated material identifiers corresponding to the same target plastic product color effect requirement identifier are sorted from high to low according to their component contribution parameters to generate a priority sorting list of candidate associated materials. A predetermined number of candidate associated material identifiers that are at the top of the priority sorting list of candidate associated materials are selected as the basic candidate material identifier set that meets the color effect requirements of the target plastic product. The basic candidate material identifiers in the basic candidate material identifier set are combined and constructed to generate a candidate color masterbatch material combination containing multiple candidate material combination methods. The candidate color masterbatch material combination contains multiple basic candidate material identifiers and their combination relationship descriptors. Based on the component contribution parameters corresponding to each basic candidate material identifier in the candidate masterbatch material combination, the overall contribution comprehensive evaluation value of the candidate masterbatch material combination is calculated. The overall contribution comprehensive evaluation value is obtained by weighted summation of the component contribution parameters of all basic candidate material identifiers in the combination. From all candidate masterbatch material combinations, the candidate masterbatch material combinations whose overall contribution comprehensive evaluation value reaches the preset overall contribution threshold are selected as the target candidate material combinations corresponding to the color effect requirements of the target plastic product; For each target candidate material combination, extract the historical ratio usage records of each basic candidate material in the combination from the historical formula database. The historical ratio usage records contain the ratio values ​​of each basic candidate material in different historical formulas and their corresponding color effect achievement scores. Based on the distribution range of the ratio values ​​of each basic candidate material in the historical ratio usage records and the corresponding color effect achievement score, the preliminary ratio feasible range of each basic candidate material in the target candidate material combination is determined. The lower limit of the preliminary ratio feasible range is the smallest ratio value in the historical ratio values ​​where the color effect achievement score exceeds the preset score threshold, and the upper limit of the preliminary ratio feasible range is the largest ratio value in the historical ratio values ​​where the color effect achievement score exceeds the preset score threshold. Each target candidate material combination and its corresponding multiple basic candidate materials are associated and stored to form an initial formulation candidate entry containing the target candidate material combination identifier and multiple initial ratio feasible intervals. All initial formulation candidate entries corresponding to the color effect requirements of the target plastic products are aggregated according to the color effect requirement identifier of the target plastic products to generate an initial formulation candidate set.

5. The method for optimizing masterbatch formulations using deep learning as described in claim 1, characterized in that, The process involves co-optimizing the formulation parameters of each candidate masterbatch material combination and its initial ratio range in the initial formulation candidate set. A deep neural network model for ratio optimization is invoked to model and analyze the component interaction effects of the candidate masterbatch material combinations, generating a target optimized ratio parameter set for each candidate masterbatch material combination, including: Extract the basic candidate material identifiers and their corresponding lower and upper limits of the preliminary feasible range for each candidate masterbatch material combination from the initial formulation candidate set; The basic candidate material identifiers are subjected to one-hot encoding to generate a material identifier feature vector corresponding to each basic candidate material identifier. At the same time, the lower limit and upper limit of the preliminary feasible range of proportions are subjected to numerical feature encoding to generate a proportion range feature vector containing the numerical features of the lower limit and the upper limit of the range. The material identification feature vector and the ratio range feature vector are concatenated to generate a combined input feature vector corresponding to each basic candidate material. The combined input feature vectors corresponding to all basic candidate materials in the candidate masterbatch material combination are stacked according to a preset combined feature input order to generate the combined feature input tensor corresponding to the candidate masterbatch material combination. The combined feature input tensor is input into the feature extraction layer of the ratio optimization deep neural network model for component feature extraction processing, generating a component feature map of the candidate color masterbatch material combination. The component feature map includes the local feature response of each basic candidate material in the combination and the interactive feature response between multiple basic candidate materials. The interaction modeling layer of the ratio optimization deep neural network model is used to model the component interaction effect of the component feature map, and the material interaction coefficient between any two basic candidate materials in the candidate masterbatch material combination and the synergistic enhancement coefficient between multiple basic candidate materials are calculated. The material interaction coefficient and synergistic enhancement coefficient are input into the ratio optimization prediction layer of the ratio optimization deep neural network model. The ratio parameter prediction is performed by combining the ratio range feature vector in the combined feature input tensor to generate the initial optimized ratio parameter set corresponding to the candidate masterbatch material combination. For each initial optimized ratio parameter in the initial optimized ratio parameter set, the ratio rationality constraint condition is verified. Initial optimized ratio parameters that do not meet the preset ratio rationality constraint condition are deleted, and initial optimized ratio parameters that meet the preset ratio rationality constraint condition are retained as candidate optimized ratio parameters. Cluster analysis is performed on the candidate optimized ratio parameters. Candidate optimized ratio parameters with similar values ​​are merged into the same ratio parameter cluster. The candidate optimized ratio parameter corresponding to the cluster center point of each ratio parameter cluster is selected as the representative ratio parameter. The representative proportion parameters are associated and stored according to their corresponding candidate masterbatch material combination identifiers to generate a target optimized proportion parameter set containing multiple representative proportion parameters for each candidate masterbatch material combination.

6. The method for optimizing masterbatch formulations using deep learning as described in claim 1, characterized in that, The process of generating the final optimized design formula for the masterbatch based on the target optimized ratio parameter set, and outputting the final optimized design formula for the masterbatch to the masterbatch production control system, includes: Analyze each target optimization ratio parameter entry in the target optimization ratio parameter set, and extract the candidate masterbatch material combination identifier and the corresponding representative ratio parameter contained in each target optimization ratio parameter entry; Based on the candidate masterbatch material combination identifier, the corresponding basic candidate material identifier list is retrieved from the initial formulation candidate set. Each basic candidate material identifier in the basic candidate identifier list is associated with the representative proportion parameter to generate a preliminary formulation component correspondence. Perform material property verification processing on each basic candidate material identifier in the preliminary formulation component correspondence to confirm the available inventory status and material batch information of the material corresponding to each basic candidate material identifier in the current masterbatch production control system. Based on the material batch information, extract the actual performance parameters of the material corresponding to each basic candidate material identifier from the material batch database, calculate the matching degree between the actual performance parameters and the representative proportion parameters, and obtain the matching degree coefficient between the material batch and the proportion parameters. The basic candidate material identifiers and their corresponding representative ratio parameters that reach the preset matching degree threshold are used as the formula component items to be confirmed, and the basic candidate material identifiers and their corresponding representative ratio parameters that have a matching degree coefficient lower than the preset matching degree threshold are marked as the formula component items to be adjusted. For the basic candidate material identifier marked as the component item to be adjusted, retrieve its list of available alternative materials, select the alternative material identifier with the highest matching degree with the representative ratio parameter from the list of available alternative materials, replace the original basic candidate material identifier with the alternative material identifier, and adjust the representative ratio parameter corresponding to the alternative material identifier to the alternative material suitable ratio parameter. All unconfirmed formulation component entries and adjusted formulation component entries are merged to generate a final formulation component entry list containing multiple formulation component entries. Each formulation component entry includes a material identifier, material batch information, and the corresponding optimized ratio value. The optimized proportion values ​​in the final formula component list are summed and normalized. The sum of the optimized proportion values ​​is adjusted to the preset total weight benchmark value of the formula, and the normalized optimized proportion values ​​are generated. The material identifier, material batch information, and normalized optimized ratio values ​​are packaged according to a preset formula data format to generate the final optimized design formula for the color masterbatch. The final optimized design formula of the masterbatch is transmitted to the formula receiving module of the masterbatch production control system through a data interface, triggering the masterbatch production control system to perform formula configuration operations such as material weighing, mixing and extrusion granulation according to the optimized ratio values.

7. The method for optimizing masterbatch formulations using deep learning as described in claim 1, characterized in that, The process of constructing the formulation component association mapping model includes: Collect historical masterbatch formulation design case data, which includes the basic masterbatch material identification used in each historical case, the corresponding target plastic product color effect requirements, and the color effect achievement evaluation score of the actual produced masterbatch products; The material property feature vector corresponding to the basic masterbatch material identifier is extracted from the data of each historical masterbatch formulation design case as the input sample of the model training source end. The standardized color effect requirements corresponding to the color effect requirements of the target plastic product are extracted as the input sample of the model training target end. The color effect achievement evaluation score is extracted as the supervision label data for model training. An initial formulation component association mapping neural network structure is constructed, which includes a source-end feature embedding layer, a target-end feature embedding layer, a bidirectional attention interaction layer, and a component relationship inference output layer. The source input samples of the model training are input into the source feature embedding layer for source feature embedding learning to obtain the source embedding feature representation. The target input samples of the model training are input into the target feature embedding layer for target feature embedding learning to obtain the target embedding feature representation. The source-end embedded feature representation and the target-end embedded feature representation are input into the bidirectional attention interaction layer to calculate attention weights, generating the source-end to target-end attention weight matrix and the target-end to source-end attention weight matrix. The attention weight matrix from the source end to the target end and the attention weight matrix from the target end to the source end are input into the component relationship reasoning output layer to perform component relationship reasoning calculation, and the predicted correlation score between each source end input sample and each target end input sample is output. The model loss function is constructed based on the difference between the predicted correlation score and the color effect achievement evaluation score. The network weight parameters of the formula component correlation mapping neural network structure are updated using the backpropagation algorithm until the model loss function converges, thus obtaining the trained formula component correlation mapping model.

8. The method for optimizing the formulation of color masterbatch by combining deep learning according to claim 1, characterized in that, The construction process of the ratio optimization deep neural network model includes: Multiple sets of color masterbatch formulation experimental data were collected. The color masterbatch formulation experimental data included the basic color masterbatch material combination identifier used in each set of experiments, the actual ratio value of each basic color masterbatch material, and the measured index of the color effect of the color masterbatch products obtained in the experiment. Material identification encoding is performed on the basic masterbatch material combination identifiers in each group of masterbatch formulation experimental data to generate a material combination identifier feature vector. The actual ratio values ​​of each basic masterbatch material are normalized to generate a ratio value feature vector. The material combination identification feature vector and the ratio numerical feature vector are fused to generate the experimental input feature tensor corresponding to each set of experimental data. The measured color effect index of the masterbatch product obtained in the experiment is used as the experimental output label data. An initial matching optimization deep neural network structure is constructed, which includes a feature extraction convolutional layer, an interaction modeling graph neural network layer, a matching optimization prediction fully connected layer, and an output layer. The experimental input feature tensor is input into the feature extraction convolutional layer for local feature extraction processing to obtain a local feature response map. The local feature response map is then input into the interaction modeling neural network layer to model the interaction relationship between materials, resulting in an interaction feature representation. The interaction feature representation is input into the fully connected layer for ratio optimization prediction and then subjected to nonlinear mapping to obtain the initial ratio optimization prediction value. The initial ratio optimization prediction value is then input into the output layer for dimensional transformation to obtain the model prediction output with the same dimension as the experimental output label data. The model training loss function is constructed based on the difference between the model's predicted output and the experimental output label data. The gradient descent optimization algorithm is used to update the network weight parameters of the ratio-optimized deep neural network structure until the model training loss function value is reduced to below the preset loss threshold, thus obtaining the trained ratio-optimized deep neural network model.

9. The method for optimizing masterbatch formulations using deep learning as described in claim 1, characterized in that, The method further includes: The color effect measurement data of the actual color masterbatch products obtained after production according to the final optimized color masterbatch design formula are collected in real time from the color masterbatch production control system. The color effect measured data of the actual color masterbatch product is compared and analyzed with the color effect requirements of the target plastic products in the set of color effect requirements of the target plastic products. The color difference offset between the measured color effect data and the color effect requirements is calculated. Based on the color difference offset, determine the degree to which the actual color effect of the final masterbatch optimized design formula meets the standard, and generate an evaluation index for the actual effect of the formula. The evaluation index of the actual effect of the formula is associated with the final optimized design formula of the color masterbatch and stored to form a historical record of the formula effect. When the actual effect of the formula is lower than the preset achievement threshold, the material identifier and optimized ratio value in the corresponding final color masterbatch optimized design formula are extracted, and the material identifier and optimized ratio value are combined with the collected color effect test data to form a new formula optimization training sample. The new formula optimization training samples are added to the training dataset of the ratio optimization deep neural network model, triggering incremental update training of the ratio optimization deep neural network model and generating the updated ratio optimization deep neural network model.

10. A color masterbatch formulation optimization design system combining deep learning, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the deep learning-integrated masterbatch formulation optimization design method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images