A method for evaluating the quality of formulated products

Abstract

The present invention discloses a method for evaluating the quality of a formula product, which digitally collects the raw material grade, chemical composition index, process data, etc. generated in the formula industry, then uses deep neural networks, fuzzy neural networks, recurrent neural networks and other technologies to establish a formula multi-dimensional model, and then performs comprehensive quality scoring on the formula products in various complex scenarios, and finally uses artificial neural networks and expert scoring values ​​to perform multi-level model combination self-learning optimization on the contribution parameters of each model. Based on the method proposed by the present invention, not only can the quality of the formula product be comprehensively predicted from all angles and multiple perspectives, but also the linkage relationship between the formula models at different levels and different granularities can be extracted based on the multi-level model combination optimization contribution parameter method, which is more suitable for the complex scenarios of the formula model, so that the domain knowledge can be digitally expressed and has transplantability and inheritance, and the development process of the formula product is simulated, saving time and material costs.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial multi-source data processing, and in particular, relates to a formula product quality evaluation method of multi-dimensional model multi-level self-learning fusion. Background Art

[0002] Important factors such as human-machine interconnection, cloud computing and edge computing, deep learning and capital have provided big data environment, computing power, algorithm models and industrial support for the development of artificial intelligence (AI) technology, enabling AI technology to develop rapidly. For formulation companies, accurately predicting and evaluating the quality performance of formulated products based on data and models can effectively simulate the test process of formulated products and provide scientific basis for formulation designers, thereby optimizing formulation processes and resource allocation.

[0003] At the same time, technologies such as computer vision and near-infrared spectroscopy are developing rapidly, providing advanced productivity for formulation industrial production.

[0004] Generally, the quality of formulated products is related to many factors, including the chemical composition of the raw materials, physical properties, formula structure and compatibility, and process data. The process of expressing complex information and internal laws of multiple links such as raw materials, formulas, and processes is complex and cannot be completed by a single algorithm model. This mapping relationship containing multiple complex sub-problems needs to be oriented towards specific decision-making goals and build multiple performance indicator models (such as sensory evaluation performance, raw material evaluation model, process data model, etc.).

[0005] However, current research is limited to the prediction and evaluation of a single model and does not consider the linkage between model parameters. There are problems such as poor prediction stability, one-sided feature information and unsatisfactory prediction accuracy. Summary of the invention

[0006] The present invention proposes a method for evaluating the quality of formulated products, which fully collects near-infrared spectral data, images, process data and expert experience data of the formulated industry, uses multi-dimensional models to predict the performance of formulated products, and achieves the goal of comprehensive evaluation of the quality of formulated products. It meets the needs of all-round comprehensive evaluation of the quality of formulated products in the production process of the formulated industry, simulates the formulation product development process, and reduces the experimental cost and time cost of the formulated products.

[0007] The present invention is implemented by the following technical solutions:

[0008] A method for evaluating formulated products is proposed, including:

[0009] S1, collect recipe data;

[0010] S2, building a multidimensional model of the recipe;

[0011] S3, generating several formulation schemes and scoring them to obtain a score value for each formulation scheme;

[0012] S4, assign contribution parameters and values ​​to each model constructed in S3;

[0013] S5, designing the neural network structure to optimize the contribution parameters and obtain the weight coefficients of each model, including:

[0014] (1) Randomly select from the t constructed models Models, selected The corresponding formula data is input into each selected model for prediction and output results; the mean of the prediction results of each formula and the loss value of the corresponding score value are calculated and fed back to the neural network to optimize the contribution parameters; among them, From 1 to t;

[0015] (2) After traversing all model combinations, the contribution parameter matrix is ​​obtained ,and , find the average ;

[0016] (3) Calculate the weight coefficient of each model as:

[0017] ;

[0018] S6, constructing an evaluation system according to the weight coefficient calculated in S5, and inputting the data of the formulation to be evaluated to obtain a quality evaluation prediction.

[0019] In some embodiments of the present invention, the data collected in step S1 includes near infrared spectroscopy data, image data, process data and sensory evaluation data of the formula raw materials;

[0020] Performing data enhancement, smoothing, derivation and / or orthogonal signal correction processing on near infrared spectroscopy data;

[0021] Performing affine transformation, horizontal flipping, vertical flipping, grayscale conversion, normalization and / or exponential moving average processing on the image data;

[0022] The collection of sensory evaluation data includes: establishing a sensory index system for formula raw materials and describing it in words, vectorizing the description text to form index system variables, and scoring the formula raw materials to obtain the raw material scoring values.

[0023] In some embodiments of the present invention, step S2 includes:

[0024] Use neural network models or machine learning methods to build raw material classification and chemical composition index prediction models;

[0025] A sensory evaluation model was constructed using fuzzy neural network;

[0026] The process data model is constructed using a recurrent neural network with an added time dimension.

[0027] Compared with the prior art, the advantages and positive effects of the present invention are: in the quality evaluation method of the formula product proposed by the present invention, different grades of raw materials, chemical composition indicators, process data, etc. generated in the formula industry are digitally collected, and then a multi-dimensional model of the formula is established using technologies such as deep neural networks, fuzzy neural networks, and recurrent neural networks, and then the experience of field experts (such as formulators and tasters) is used to conduct comprehensive quality scoring of formula products in various complex scenarios, and finally the contribution parameters of each model are optimized by multi-level model combination self-learning using artificial neural networks and expert scoring values. Based on the quality evaluation method proposed by the present invention, the formula products can be scientifically evaluated from different dimensions, and not only can the quality of the formula products be comprehensively predicted from all-round and multiple perspectives, but also the multi-level model combination optimization contribution parameter method proposed by the present invention can extract the linkage relationship between formula models from different levels and different granularities, which is more suitable for complex scenarios of formula models, so that the field knowledge is digitally expressed and has transplantability and inheritance, simulating the development process of formula products, and saving time and material costs.

[0028] Other features and advantages of the present invention will become more apparent after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0030] Figure 1 The following is a schematic diagram of the steps of the method for evaluating the quality of a formula product proposed in the present invention;

[0031] Figure 2 The process and flow diagram of the quality evaluation method of the formula product proposed in the present invention;

[0032] Figure 3 A schematic diagram of a deep neural network model constructed in an embodiment of the present invention;

[0033] Figure 4 The figure is a schematic diagram of the structure of a fuzzy neural network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] Establishing raw material grade models and chemical composition index models is the key to quality testing of formula raw materials, which directly affect the style and quality of formula products. At present, the quality evaluation of formula products is based on a single dimensional model, but the sub-models with different objectives extract limited information and often only contain partial information about the formula products, and cannot capture the multi-dimensional information in the production process of formula products. In addition, in actual applications, the information provided by each sub-model is different, and a single model cannot systematically and comprehensively describe the quality of formula products.

[0036] Different models provide effective information from multiple dimensions and angles, which can fully tap into deep information, comprehensively consider the performance of formulated products, and improve the prediction credibility of formula quality, thereby improving the stability and confidence of the model. Therefore, the multidimensional model fusion modeling method is a perfection of the single modeling method.

[0037] Traditional multi-model fusion methods aim to jointly learn multiple related tasks so that the knowledge contained in a single task can be used by other tasks to improve the overall predictive ability of the model, including averaging, voting, stacking and boosting methods. However, the multi-dimensional evaluation models of formulated products vary greatly, and the traditional multi-model joint drive common parameter optimization method is not suitable for the comprehensive evaluation of formulated products.

[0038] In addition, model weight distribution is also crucial to avoid the situation where one or more models dominate the network weight. Traditional model weight optimization methods include loss-based, gradient descent, etc. However, these methods all optimize the contribution parameters from a single level and ignore the relationship between multiple model combinations of different granularities.

[0039] Based on the above considerations, the present invention aims at the above problems and proposes a formula product quality evaluation method with multi-dimensional model and multi-level self-learning fusion according to the intrinsic correlation characteristics of raw materials, processes, formulas and product quality. The model is extracted in combination of different dimensions or levels, so as to analyze the influence of formula model combination on the final formula product quality prediction from different granularities. The advantages of expert prior knowledge information and multi-dimensional models are fully utilized to realize the self-learning optimization process of multi-model and multi-level contribution weight parameters, and finally solve the problem of one-sided and low precision of single-dimensional model prediction of formula product quality.

[0040] Specifically, Figure 1and Figure 2 As shown, the following steps are included:

[0041] S1: Collect recipe data.

[0042] The recipe data in the present invention includes but is not limited to near infrared spectrum data, image data, process data and sensory evaluation data of recipe raw materials.

[0043] Preprocess the near-infrared spectral data and raw material image data of the formula raw materials: The preprocessing of the near-infrared spectral data of the formula raw materials includes data enhancement, smoothing, derivation and orthogonal signal correction, etc. The preprocessing of the raw material image data includes affine transformation, horizontal / vertical flipping, grayscale, normalization, exponential moving average (EMA), etc.

[0044] The collection of sensory evaluation data of formula raw materials includes:

[0045] (1) Establish a sensory index system for formula raw materials; taking the cigarette formula industry as an example, the sensory indicators of single-ingredient cigarettes include flavor type, aroma quality, aroma quantity, concentration, impurities, irritation, aftertaste, strength, combustibility, ash content, etc.

[0046] (2) Provide a written description of the formula raw material index system.

[0047] (3) Represent the descriptive text as a vector to form an indicator system variable.

[0048] (4) Ask formulation experts to give a comprehensive score to the formulation raw materials and obtain the scoring value of the raw materials.

[0049] Next, the recipe data is digitized and converted into a form that can be processed by a computer, including:

[0050] (1) Constructing a recipe data set As the input of subsequent tasks; among them, Indicates the number of recipe data types. Represents a kind of recipe data after preprocessing; Indicates the data dimension.

[0051] The collected formula input data can be expressed as a matrix ,Right now .

[0052] (2) Build the recipe data in step (1) Corresponding label or chemical composition index set ; Taking the cigarette industry as an example, the tobacco leaf grade label data include B1F, B1K, B2F, B2K, B2V, etc., the chemical composition indicators include total sugar, total nicotine, reducing sugar, total nitrogen and other content data, and the sensory evaluation data is the expert score.

[0053] but ,in, express Corresponding label data or chemical composition indicators.

[0054] S2: Construct a multidimensional model of the recipe.

[0055] It includes raw material classification and chemical composition index prediction model of formula raw materials, sensory evaluation model and process data model.

[0056] (1) Construct a raw material classification and chemical composition index prediction model.

[0057] The data is in the form of near-infrared spectrum or image data. Since there is no temporal correlation between the data, neural network models or traditional machine learning methods are usually used for training and prediction.

[0058] When the sample size of the formula raw materials is large, that is, in a large sample scenario, a deep neural network model is used to build a classification or prediction model. The process is as follows Figure 3 As shown, represents the parameters to be optimized, Indicates input, represents the output data of the deep neural network, represents the weight of the deep neural network, is the output after weighted summation, and the calculation method is shown in formula (1). Represents a function operation, usually a one-dimensional convolution or a two-dimensional convolution operation. For example, near-infrared spectroscopy uses one-dimensional convolution, and images use two-dimensional convolution. is the predicted value of the raw material grade or chemical composition index by the deep neural network model, and the calculation method is shown in formula (2). Represents an activation function, such as the Sigmoid function, Tanh function, ReLU function, or Softmax function.

[0059]

[0060]

[0061] When there are n samples, the predicted value can be obtained ; Then use the reference value of the label or chemical composition index Calculate the loss and then feed it back to finally achieve the parameter optimization process of the deep neural network. The calculation method of the loss is shown in formulas (3) and (4). Formula (3) is the calculation method of the classification model, and formula (4) is the calculation method of the regression model; where n is the number of samples.

[0062]

[0063]

[0064] So far, we can get the prediction results of a raw material model in a large sample scenario; the same is true for other models, and finally we can get the prediction value .

[0065] When the number of labeled samples is small, a semi-supervised learning algorithm is used to build the model and output the prediction results at the same time; semi-supervised learning methods include S 3 VM, GSSL, etc.

[0066] When there are fewer training samples, data enhancement methods are used to enhance the training data first, usually including GAN / WGAN, etc., and then traditional machine learning methods are used to model, such as PLSR / SVM / XGBoost, etc.

[0067] (2) Construct a sensory evaluation model.

[0068] Will As input layer parameters, it is input into the fuzzy neural network and uses the recipe data obtained in step S1 Corresponding label or chemical composition index set As prior knowledge to train the neural network, the structure of the fuzzy neural network is as follows Figure 4 As shown, after multiple iterations, the Loss converges, and the Loss formula refers to formula (4).

[0069] (3) Construct a process data model.

[0070] Formula process data include equipment pressure, temperature, humidity and moisture content of formula raw materials after being processed by industrial equipment. Since this type of data is a continuously changing sequence and has a strong time correlation, traditional neural networks are not applicable because they cannot make predictions based on historical data. In addition to the input of current information, each neuron in the recurrent neural network also has previously generated memory information, which retains the dependency of the sequence. Therefore, using a recurrent neural network to establish a formula process data model can integrate historical data and correlation to enable the network to generate memory, and can well mine the characteristics and rules of formula process data.

[0071] The present invention uses a recurrent neural network to establish a recipe process data model, adding a time dimension, assuming represents the time dimension, i.e. The output at time is expressed as , access to The hidden layer at time The output at a given moment depends not only on the input , and also referred to The hidden layer output at time , similarly, The output of the time will also refer to The hidden layer output at time In this way, the process of model building for process data (pressure, temperature, etc.) will have time memory and the model will be more stable.

[0072] The process of building the process data model is as follows:

[0073] 1) Assumptions is the time state weight between the upper layer and the lower layer of the network, then Input at time The calculation method is shown in formula (5), where: is the input process data, is the network weight parameter, for The output value at time, is the bias value:

[0074]

[0075] Then, Output at any time The calculation method is shown in formula (6):

[0076]

[0077] in, Indicates all The matrix composed of represents the activation function, such as .

[0078] 2) Utilize Back propagation is performed, and the Loss is calculated using formula (4).

[0079] 3) Set the convergence conditions and iterate several times until the model converges.

[0080] So far, the formula raw material classification, chemical composition index, sensory and process data models have been established and defined as .

[0081] S3: Generate several formulation schemes and score them to obtain a score value for each formulation scheme.

[0082] The formulator designs the formula scheme, including raw material grade information, chemical composition index information, sensory evaluation information and process data information, converts it into an inputtable data form, and then scores these formula schemes to obtain the score value. ,in Indicates the total number of recipes.

[0083] S4: Assign contribution parameters and values ​​to each constructed model.

[0084] Assign contribution parameters to each established model:

[0085] ,and ; Each value determines the degree of influence of each model on the final formula product quality evaluation output results.

[0086] S5: Design a neural network structure to optimize the contribution parameters and obtain the weight coefficients of each model.

[0087] In the present invention, a neural network structure is designed, such as a feedforward neural network, a multilayer perceptron or a convolutional neural network, and each contribution parameter is optimized by a Loss feedback method, including:

[0088] (1) Randomly select from the t constructed models Models, selected The corresponding formula data is input into each selected model for prediction and output; the mean of the prediction results of each formula and the loss value of the corresponding score value are calculated and sent back to the neural network to optimize the contribution parameters. From 1 to t.

[0089] Specifically include:

[0090] (11) Randomly select a model from the t models constructed, and select The corresponding formula data is input into each selected model for prediction and output. The mean of the prediction results of each formula and the loss value of the corresponding score value are calculated and fed back to the neural network to optimize the contribution parameters.

[0091] exist A model is randomly selected from the models, and the contribution parameter of each model is , then there are a total of A selection method is used to input the corresponding formula data (raw material classification data, chemical composition data, sensory data or process data) into the selected model, and the model is used to make predictions and output the results. ,and .

[0092] beg The average , calculated using formula (4) and The Loss value is then passed back to the neural network to optimize .

[0093] (12) Randomly select two models from the t models constructed, and select The corresponding formula data is input into each selected model for prediction and output. The mean of the prediction results of each formula and the loss value of the corresponding score value are calculated and fed back to the neural network to optimize the contribution parameters.

[0094] exist Two models are randomly selected from the models, and the contribution parameter of each model is , then there are a total of A selection method, input the corresponding formula data into the selected model, use the model to predict and output the results ,and .

[0095] beg The average , calculated using formula (4) and The Loss value is then passed back to the neural network to optimize .

[0096] (13) Randomly select three models from the constructed t models, a total of The corresponding formula data is input into each selected model for prediction and output. The mean of the prediction results of each formula and the loss value of the corresponding score value are calculated and fed back to the neural network to optimize the contribution parameters.

[0097] exist Three models are randomly selected from the models, and the contribution parameter of each model is , then there are a total of A selection method, input the corresponding formula data into the selected model, use the model to predict and output the results ,and .

[0098] beg The average , calculated using formula (4) and The Loss value is then passed back to the neural network to optimize .

[0099] (14) The number of randomly selected models is increased by one and the steps of predicting and optimizing the contributing parameters are repeated until t models are selected from the t constructed models.

[0100] (2) Set the convergence conditions and traverse all model combinations to obtain the contribution parameter matrix ,and , find the average .

[0101] The convergence conditions here are, for example: the number of iterations is set to be greater than or equal to 100, and the Loss value is less than 0.5.

[0102] (3) Calculate the weight coefficient of each model as:

[0103] .

[0104] S6: Construct an evaluation system based on the weight coefficients calculated in S5, and input the data of the formula to be evaluated to obtain a quality evaluation prediction.

[0105] The entire formula raw material model parameters are determined in step S2, and the contribution parameters of each model are determined in step S5, thereby the entire rating system can be determined.

[0106] Enter new recipe data into the evaluation system The model obtains the predicted value of the quality evaluation of the formulated product , formulation experts can give evaluation values , by calculating the test error To evaluate the model performance.

[0107] The above-mentioned multi-dimensional model multi-level self-learning fusion formula product quality evaluation method proposed by the present invention, compared with the traditional single model evaluation method, makes full use of the evaluation model of multiple dimensions of the formula and integrates multiple evaluation modes, which can realize the integrated optimization of multi-task collaboration. The evaluation model is more comprehensive, objective, and closer to human evaluation methods, so the accuracy is higher. At the same time, compared with other fixed weight parameter value optimizations, the optimization method for self-learning of multi-model multi-level contribution parameters proposed by the present invention fully considers the potential relationship of different levels and granularity model combinations. Therefore, the method proposed by the present invention can dynamically simulate the formulation process from different levels using multi-factor contribution optimization, and optimize the adjustment parameters through formula simulation credibility evaluation through expert experience, solve the coupling effects of multiple parameters such as raw materials, auxiliary materials, and processes, and the difficulty of coordinated regulation with the formula, and the low integration of knowledge associations of different models, forming a multi-task collaborative design technical method, which is more suitable for complex scenarios of formula industrial product quality evaluation.

[0108] The following is a comparison between the formula product quality evaluation method of the multi-dimensional model multi-level self-learning fusion proposed in the present invention and the existing single model and fixed model contribution parameter methods.

[0109] Table 1 shows the comparison results of the formula product quality evaluation method of multiple single models and the multi-dimensional model multi-level self-learning fusion proposed in the present invention when t=4, where: Table 2 shows the comparison results of the formula product quality evaluation method of multiple single models and the multi-dimensional model multi-level self-learning fusion proposed in the present invention when t=6, among which, .

[0110] The present invention selects the error ERR between the model prediction score and the expert evaluation score as the measurement indicator, where It represents a single model, and the smaller the prediction error, the better the model prediction performance. It can be seen that the prediction error of a single model is high, and the prediction error of the method proposed by the present invention is small.

[0111] Table 1: Comparison of multiple single models and the method proposed in this invention when t=4

[0112]

[0113] *This is the formula product quality evaluation method of Multi-level Self-learning Fusion of Multi-dimensional Models, MSFMMs proposed in the present invention.

[0114] Table 2: Comparison of multiple single models and the method of the present invention when t=6

[0115]

[0116] *This is the formula product quality evaluation method of Multi-level Self-learning Fusion of Multi-dimensional Models, MSFMMs proposed in the present invention.

[0117] Tables 3 and 4 show the comparison results of the fixed model contribution parameter and the multi-level self-learning parameter optimization model weight assignment method. It can be seen that the self-learning contribution parameter assignment method proposed in the present invention has a smaller prediction error for the final formula product quality and better performance.

[0118] Table 3: Comparison of fixed contribution parameters and optimized contribution parameters in the present invention when t=4

[0119]

[0120] *This is the formula product quality evaluation method of Multi-level Self-learning Fusion of Multi-dimensional Models, MSFMMs proposed in the present invention.

[0121] Table 4: Comparison of fixed contribution parameters and optimized contribution parameters in the present invention when t=6

[0122]

[0123] *This is the formula product quality evaluation method of Multi-level Self-learning Fusion of Multi-dimensional Models, MSFMMs proposed in the present invention.

[0124] It should be noted that, in the specific implementation process, the above-mentioned control part can be implemented by a hardware processor executing computer execution instructions in software form stored in the memory, which will not be elaborated here, and the programs corresponding to the actions performed by the above-mentioned control circuit can be stored in the computer-readable storage medium of the system in software form, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0125] The computer-readable storage medium mentioned above may include volatile memory, such as random access memory; may also include non-volatile memory, such as read-only memory, flash memory, hard disk or solid-state drive; may also include a combination of the above types of memory.

[0126] The processor mentioned above may also be a general term for multiple processing elements. For example, the processor may be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, field programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc., or may be a dedicated processor.

[0127] It should be pointed out that the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for evaluating the quality of a formulated product, characterized in that: include: S1, collect recipe data; S2, building a multidimensional model of the recipe; S3, generating several formulation schemes and scoring them to obtain a score value for each formulation scheme; S4, assign contribution parameters and values ​​to each model constructed in S3; S5, designing the neural network structure to optimize the contribution parameters and obtain the weight coefficients of each model, including: (1) Randomly select from the t constructed models Models, selected The corresponding formula data is input into each selected model for prediction and output results; the mean of the prediction results of each formula and the loss value of the corresponding score value are calculated and fed back to the neural network to optimize the contribution parameters; among them, From 1 to t; wherein the model is the raw material classification and chemical composition index prediction model, sensory evaluation model and process data model constructed in step S2, and t is the number of constructed models; (2) After traversing all model combinations, the contribution parameter matrix is ​​obtained ,and , find the average ; (3) Calculate the weight coefficient of each model as: ； S6, constructing an evaluation system according to the weight coefficient calculated in S5, inputting the data of the formulation to be evaluated to obtain a quality evaluation prediction; Step S2 includes: Use neural network models or machine learning methods to build raw material classification and chemical composition index prediction models; A sensory evaluation model was constructed using fuzzy neural network; A recurrent neural network with an added time dimension is used to build a process data model; Step S3 includes: Design formulations and convert them into inputtable data forms; Scoring the formula scheme to obtain a scoring value; The data collected in step S1 include near infrared spectrum data, image data, process data and sensory evaluation data of the formula raw materials; Performing data enhancement, smoothing, derivation and / or orthogonal signal correction processing on near infrared spectroscopy data; Performing affine transformation, horizontal flipping, vertical flipping, grayscale conversion, normalization and / or exponential moving average processing on the image data; The collection of sensory evaluation data includes: establishing a sensory index system for formula raw materials and describing it in words, vectorizing the description text to form index system variables, and scoring the formula raw materials to obtain the raw material scoring values.

Images