Food processing technology intelligent prediction method and system based on target flavor

Abstract

The invention discloses a food processing technology intelligent prediction method and system based on a target flavor. The method comprises the following steps: acquiring a target flavor parameter, and constructing a target flavor vector based on the target flavor parameter; inputting into a process prediction neural network model for processing process prediction to obtain a prediction result; wherein the target flavor vector is converted into a vector representing a specific flavor intention through a flavor target encoder, so as to obtain a flavor intention vector; generating a parameter output vector meeting the process requirement based on the flavor intention vector by the constraint equation network; and based on the parameter output vector, using a sparse attention mask generator to create a mask containing process feasibility constraints, and outputting an optimal parameter combination satisfying all process constraints by a process parameter generator to form a prediction result. By implementing the method provided by the invention, learning and prediction can be carried out in a solution space meeting constraint conditions, and space is left for the development of a new process and a new product while the process stability is ensured.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for intelligent prediction of food processing technology based on target flavor. Background Technology

[0002] The flavor and texture of food are key factors determining its market competitiveness. Soy products, represented by tofu, possess unique flavors such as bean aroma, umami, and even a slightly bitter or fishy taste, as well as textures like firmness, elasticity, and smoothness. These are determined by dozens of parameters across multiple stages, from raw material selection to processing techniques. These parameters exhibit complex non-linear relationships. For example, increasing the concentration of soy milk enhances the bean aroma but may result in a coarser texture; using bittern as a coagulant can impart a unique flavor to tofu, but its dosage and operating temperature are extremely difficult to control, and even slight errors can lead to overcooked tofu or a bitter taste.

[0003] Traditional food processing technology development relies on experienced experts optimizing through repeated trials. This approach has significant drawbacks: First, it is time-consuming and costly, requiring numerous experiments to find the optimal combination of processes. Second, the knowledge of senior process engineers is difficult to quantify and standardize, leading to inconsistent product quality and the risk of a talent shortage. Finally, traditional methods often fail to provide effective guidance when developing entirely new flavor products. In recent years, while machine learning has been introduced to predict food quality, commonly used models neglect the sparsity of the process-flavor space, meaning that many combinations of process parameters are actually invalid or illogical. This makes general-purpose models inefficient to train in invalid data regions and prone to getting trapped in local optima.

[0004] Therefore, it is necessary to design a new method to design constraints as embedded differentiable equations, enabling neural networks to learn and predict within the solution space that satisfies the constraints. This ensures process stability while also leaving room for the development of new processes and products. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for intelligent prediction of food processing technology based on target flavor.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligent prediction of food processing technology based on target flavor, comprising: Obtain the target flavor parameters and construct a target flavor vector based on the target flavor parameters; The target flavor vector is input into the process prediction neural network model to predict the processing technology and obtain the prediction result. The step of inputting the target flavor vector into a process prediction neural network model to predict the processing technology and obtain the prediction result includes: The target flavor vector is input into the process prediction neural network model, and the target flavor vector is converted into a vector representing a specific flavor intent by the flavor target encoder to obtain the flavor intent vector. The constraint equation network generates a parameter output vector that meets the process requirements based on the flavor intention vector; Based on the parameter output vector, a sparse attention mask generator is used to create a mask containing process feasibility constraints, and the process parameter generator outputs the optimal combination of parameters that satisfies all process constraints to form the prediction result.

[0007] The further technical solution is as follows: the step of converting the target flavor vector into a vector representing a specific flavor intent through a flavor target encoder to obtain a flavor intent vector includes: After normalizing the target flavor vector and processing it through a fully connected layer by a flavor target encoder, it is converted into a high-dimensional vector representing a specific flavor intent, thus obtaining a flavor intent vector.

[0008] The further technical solution is as follows: the generation of parameter output vectors that meet the process requirements based on flavor intention vectors by the constraint equation network includes: Based on the flavor intent vector, the constraint equation network is processed using a multi-head output neural network structure, so that each output head outputs one or a set of related parameters that need to be dynamically generated in the corresponding differentiable constraint equation, and the parameters are combined to generate a parameter output vector that meets the process requirements.

[0009] The further technical solution is as follows: Based on the parameter output vector, a sparse attention mask generator is used to create a mask containing process feasibility constraints, and the process parameter generator outputs the optimal parameter combination that satisfies all process constraints to form the prediction result, including: For categorical parameter prediction, based on the parameter output vector, a sparse attention mask generator is used to generate the original score vector, and a mask vector is designed according to the constraint rules. The mask vector is added to the original score vector to filter infeasible options, forming a masked score vector. The process parameter generator processes the masked score vector through the Softmax function to obtain the categorical parameter prediction result. For continuous parameter prediction, unconstrained initial values ​​are predicted based on the parameter output vector. The process parameter generator then processes these unconstrained initial values ​​using the Sigmoid function and linear transformation to obtain the continuous parameter prediction results. The prediction results of the categorical parameters and the prediction results of the continuous parameters are combined to obtain the prediction result.

[0010] The further technical solution is as follows: after creating a mask containing process feasibility constraints using a sparse attention mask generator based on the parameter output vector, and having the process parameter generator output the optimal parameter combination that satisfies all process constraints to form the prediction result, it also includes: After the prediction results are processed by a multilayer perceptron, a risk vector corresponding to a predefined potential risk dimension is output, and the risk value of negative results caused by different process parameters is accurately predicted based on historical failure and boundary experiment data.

[0011] The further technical solution is as follows: the training process of the process prediction neural network model includes: Integrate historical production data, laboratory orthogonal experimental data, and physicochemical model simulation data to form standardized data; Constructing a neural network model and its loss function; Based on the standardized data and loss function, the Adam optimizer is used to update the parameters of the neural network model using gradient descent with an adaptive learning rate, and sparse attention mask pruning is used to search the space.

[0012] The further technical solution is as follows: the neural network model includes the flavor target encoder, the constraint equation network, the sparse attention mask generator, the process parameter generator, and the risk predictor.

[0013] The further technical solution is as follows: the loss function is formed by weighting the mean square error loss of continuous parameters, the cross-entropy loss of discrete parameters, the interval constraint violation loss, the mask sparsity loss, and the risk prediction loss.

[0014] The further technical solution is that the discrete parameters include categorical parameters.

[0015] This invention also provides an intelligent prediction system for food processing techniques based on target flavor, comprising: An acquisition unit is used to acquire target flavor parameters and construct a target flavor vector based on the target flavor parameters; The prediction unit is used to input the target flavor vector into the process prediction neural network model to predict the processing technology and obtain the prediction result. The prediction unit includes: The encoding subunit is used to input the target flavor vector into the process prediction neural network model, and convert the target flavor vector into a vector representing a specific flavor intention through the flavor target encoder to obtain the flavor intention vector; The constraint processing subunit is used to generate a parameter output vector that meets the process requirements based on the flavor intention vector from the constraint equation network; The generation sub-unit is used to create a mask containing process feasibility constraints based on the parameter output vector using a sparse attention mask generator, and the process parameter generator outputs the optimal combination of parameters that satisfies all process constraints to form the prediction result.

[0016] The advantages of this invention compared to existing technologies are as follows: This invention transforms target flavor parameters into standardized target flavor vectors and inputs them into a process prediction neural network model. A flavor target encoder then converts these vectors into vectors representing specific flavor intentions. Next, a constraint equation network dynamically generates parameter output vectors that meet process requirements based on this flavor intention vector. These parameter output vectors embed differentiable constraint equations, representing the solution space under specific process conditions. Subsequently, a sparse attention mask generator creates a mask containing process feasibility constraints based on these parameter output vectors, ensuring that the optimal parameter combination output by the process parameter generator naturally satisfies all preset process constraints. Furthermore, the introduction of a soft constraint mechanism allows for deviations in certain parameters. Risk function evaluation ensures process stability while also providing the possibility of exploring new processes and products beyond traditional experience. This enables the neural network to effectively learn and predict within the solution space that satisfies the constraints.

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram illustrating an application scenario of the intelligent prediction method for food processing technology based on target flavor provided in this embodiment of the invention; Figure 2 A flowchart illustrating the intelligent prediction method for food processing technology based on target flavor provided in an embodiment of the present invention; Figure 3 A schematic block diagram of a food processing technology intelligent prediction system based on target flavor provided in an embodiment of the present invention; Figure 4 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0022] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the intelligent prediction method for food processing technology based on target flavor provided in an embodiment of the present invention. Figure 2This is a schematic flowchart illustrating the intelligent prediction method for food processing technology based on target flavor provided in this invention. This method is applied in a server. The server interacts with the terminal by designing complex process constraints as embedded differentiable equations and using a neural network model to learn and predict within the solution space satisfying these constraints. Specifically, the method first obtains target flavor parameters and constructs corresponding vectors, then converts them into high-dimensional vectors representing specific flavor intentions using a flavor target encoder. Next, a constraint equation network generates parameter output vectors that meet process requirements based on the flavor intention vectors, involving the use of a multi-head output structure to ensure dynamic correlation between parameters. Then, a sparse attention mask generator creates a mask containing process feasibility constraints based on the output vectors, ensuring that infeasible options are filtered out. Finally, a process parameter generator outputs the optimal parameter combination. Furthermore, different processing strategies (such as the Sigmoid function and the Softmax function) are used for continuous and categorical parameters, and potential risk values ​​are evaluated after prediction. The entire process not only ensures process stability but also promotes the development potential of new processes and products by optimizing the loss function and pruning the search space.

[0025] Figure 2 This is a flowchart illustrating the intelligent prediction method for food processing technology based on target flavor provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S120.

[0026] S110. Obtain the target flavor parameters and construct the target flavor vector based on the target flavor parameters.

[0027] In this embodiment, the target flavor parameters are a series of quantitative indicators defined by the user based on their desired final product characteristics. These parameters directly reflect the specific requirements of the finished soy products in terms of sensory and physicochemical properties. For example, in the tofu production process, target flavor parameters may include, but are not limited to, firmness (N), elasticity (%), bean aroma (GC-MS peak area), fishy smell (GC-MS peak area), and smoothness (sensory score 1-10). Each parameter has its specific meaning and unit of measurement; for example, firmness can be expressed in Newtons (N), while smoothness may be a value between 1 and 10 obtained through sensory evaluation.

[0028] The target flavor vector is a vector representation formed by standardizing the aforementioned individual target flavor parameters. This vector is used as input data to the process prediction neural network model to guide the model in generating a combination of processing parameters that meet the expected flavor characteristics. Specifically, the construction process of the target flavor vector is as follows: Normalization: Since different flavor parameters have different dimensions and scales, they first need to be normalized to map them to a uniform range (e.g., between 0 and 1 or -1 and 1). This step helps to eliminate the influence of differences in magnitude between different features, ensuring that each parameter contributes information fairly in subsequent calculations.

[0029] Forming a vector: After normalization, all target flavor parameters are organized into an ordered array or vector. Suppose we have five target flavor parameters, and the normalized results are v1, v2, v3, v4, and v5, then the target flavor vector can be represented as V = [v1, v2, v3, v4, v5].

[0030] Encoding Transformation: To enable the target flavor vector to be better understood and utilized by the neural network, it may be transformed into a higher-dimensional "flavor intent" vector through one or more fully connected layers. This high-dimensional vector can capture more complex flavor relationships and facilitates integration with subsequent components such as sparse attention mask generators and process parameter generators.

[0031] In summary, the target flavor vector not only contains users' clear needs for ideal products, but also, through standardization and encoding transformation, enables this information to be effectively expressed and processed mathematically and in computer science, thereby supporting intelligent food processing technology prediction.

[0032] S120. Input the target flavor vector into the process prediction neural network model to predict the processing technology and obtain the prediction result.

[0033] In this embodiment, the prediction results are a set of optimal process parameter combinations generated by processing the target flavor vector through a process prediction neural network model. These parameter combinations aim to maximize the achievement of user-defined target flavor characteristics and ensure that all process constraints are met. Specifically, the prediction results include, but are not limited to, key process parameters in soybean product production such as coagulant type, coagulation temperature, and pressing pressure, which together determine the flavor and texture of the final product.

[0034] In one embodiment, step S120 described above may include steps S121 to S123.

[0035] S121. The target flavor vector is input into the process prediction neural network model, and the target flavor vector is converted into a vector representing a specific flavor intention by the flavor target encoder to obtain the flavor intention vector.

[0036] In this embodiment, the flavor intent vector refers to a high-dimensional vector processed by the flavor target encoder. It captures the core information in the target flavor vector and maps it to a more abstract space for better interaction with subsequent modules (such as the sparse attention mask generator). This vector not only contains the flavor features desired by the user, but also implicitly contains information on how to achieve these goals by adjusting process parameters.

[0037] In this embodiment, the target flavor vector is normalized and processed by a fully connected layer through a flavor target encoder, and then converted into a high-dimensional vector representing a specific flavor intent, so as to obtain a flavor intent vector.

[0038] The flavor target encoder is responsible for transforming the target flavor parameters input by the user into an internal representation that the model can understand.

[0039] Input the user-defined target flavor vector. ,in, This refers to the quantity of flavor indicators. For example, for tofu, It can be ; Each item is a normalized value (e.g., a 0-10 scale).

[0040] Normalization: To eliminate the influence of dimensions, all input parameters need to be normalized. ;in, This is the original rating. and These are the minimum and maximum values ​​of the scoring range.

[0041] Through one or more fully connected layers (Fully, Connected, Layer, FC), This maps to a higher-dimensional "flavor intention" vector. ,in, It is a dimension of the intention space. ;in, and These are learnable weights and biases. It is an activation function (such as ReLU).

[0042] By normalizing the data, the dimensionality of different flavor indicators is eliminated, allowing the input parameters to be compared and processed on a uniform scale, thereby improving the stability of model training and the accuracy of prediction.

[0043] By using fully connected layers to map flavor vectors to a higher-dimensional "flavor intent" space, it is possible to capture users' complex and subtle needs for flavors, enhance the model's ability to understand and express different flavor characteristics, and provide support for achieving highly customized food production.

[0044] This conversion mechanism ensures that users can specify their desired flavor characteristics in an intuitive way (such as a 0-10 rating scale) without having to delve into the production process or formula details, greatly enhancing the user experience and ease of use.

[0045] S122. The constraint equation network generates a parameter output vector that meets the process requirements based on the flavor intention vector.

[0046] In this embodiment, the parameter output vector refers to a set of process parameter values ​​dynamically generated by the Differentiable Constraint Equation Network (DCEN) based on the flavor intent vector. This set of parameters directly reflects which process parameter configurations can ensure the feasibility and efficiency of production under the current target flavor conditions. For example, for tofu production, this may involve specific numerical ranges or fixed values ​​such as coagulant concentration, coagulation temperature, and pressing pressure.

[0047] Specifically, based on the flavor intent vector, the constraint equation network is processed using a multi-head output neural network structure, so that each output head outputs one or a set of related parameters that need to be dynamically generated in the corresponding differentiable constraint equation, and the parameters are combined to generate a parameter output vector that meets the process requirements.

[0048] Unlike rule-based or graph-based discrete logic, constraint equation networks use continuous, differentiable algebraic equations to describe process constraints. It mainly consists of two parts: DCEN and the constraint projection layer.

[0049] Input flavor intent vector .

[0050] The internal structure of DCEN is as follows: It contains a meta-constraint library. This library does not store specific rules, but rather differentiable constraint equation paradigms. All process parameters to be predicted are defined as a long vector. : When describing scenarios where multiple options are available, such as the type of coagulant, then Indicates the first For each type of coagulant, a mutually exclusive constraint equation is used. ;in It is the first The tendency weight coefficients corresponding to each type of coagulant; if the weight of a certain option becomes extremely large, then the corresponding... The value must be 0 to satisfy the inequality.

[0051] When describing parameters related to pressure and concentration, such as the lower limit of pressure increasing with increasing soy milk concentration, dependency constraint equations are used to describe the linear or nonlinear dependence between the two parameters. ,in , These are the equation parameters.

[0052] When describing the dynamic range of parameters such as temperature and time, the optimal empirical range is used. Corresponding deviation constraint equations ;in The deviation tolerance weighting coefficient It is usually set to the midpoint of the optimal experience interval, i.e. It should be noted that Alternatively, the optimal empirical interval probability distribution corresponding to different parameters can be selected. To deviate from the penalty threshold, denoted as the standard deviation of the interval parameter.

[0053] When the model predicts Deviating from the center of experience At this point, a smoothly increasing penalty gradient is generated. This gradient suggests that the model move closer to the empirical center, but does not prevent the model from choosing a value far from the center, as long as the final total loss is better. Through the design of the above soft constraints, the system of this invention not only outputs the optimal process parameters, but also provides risk assessments for exploratory formulations that violate experience but are physically feasible. This ensures the stability of conventional production and provides valuable, data-supported guidance for the discovery of new processes and new flavors, achieving a balance between following scientific principles and breaking through empirical limitations.

[0054] Specifically, DECN employs a multi-head output neural network structure, which receives the flavor intent vector. Each output header is specifically responsible for generating one or a set of related parameters that need to be dynamically generated in the above constraint equations, for example... , , , , And so on, and generate an output vector from these parameters.

[0055] To accurately and flexibly address various flavor requirements and constraints, DECN employs a multi-head output architecture. Each "head" is an independent small neural network, specifically responsible for processing a particular aspect or set of related flavor parameters.

[0056] Each output head focuses on serving a set of related flavor parameters. For example, one head might be dedicated to calculating parameters related to tofu texture (such as firmness, elasticity, etc.), while another head is responsible for flavor intensity (such as saltiness, sweetness, etc.). This design allows the model to be optimized for different types of flavor attributes, improving prediction accuracy.

[0057] Once the system receives the flavor intent vector transformed through the fully connected layer, this information is distributed to the corresponding output heads. The flavor intent vector contains the user's expectations for the flavor characteristics of the final product; it undergoes deep feature extraction and transformation to adapt to subsequent parameter generation tasks.

[0058] Each output head extracts relevant information from the flavor intent vector based on its specific task and dynamically generates the required parameters by combining them with pre-defined constraint equations. For example, if the saltiness of the product needs to be adjusted, the corresponding output head will calculate an optimal amount of salt to add based on the input flavor intent vector, considering all relevant factors.

[0059] Once all the output heads have completed their respective tasks, the parameters they produce are integrated into a complete output vector. This vector not only includes parameter values ​​directly related to flavor, but may also include factors that indirectly affect flavor, such as processing temperature and time, ensuring that the final product is as close as possible to the user's expectations.

[0060] In this way, DECN can not only effectively solve complex flavor customization problems, but also ensure flexibility and responsiveness in the production process, thereby greatly improving the possibility and technical level of personalized customization in the food industry.

[0061] S123. Based on the parameter output vector, a sparse attention mask generator is used to create a mask containing process feasibility constraints, and the process parameter generator outputs the optimal combination of parameters that satisfies all process constraints to form the prediction result.

[0062] In one embodiment, step S123 described above may include steps S1231 to S1233.

[0063] S1231. For categorical parameter prediction, based on the parameter output vector, a sparse attention mask generator is used to generate the original score vector, and a mask vector is designed according to the constraint rules. The mask vector is added to the original score vector to filter infeasible options, forming a masked score vector. The process parameter generator processes the masked score vector through the Softmax function to obtain the categorical parameter prediction result.

[0064] In this embodiment, the categorical parameter prediction result refers to the probability of each option that meets all process constraints becoming the final selection by processing the masked score vector and applying the Softmax function, and selecting the option with the highest probability as the optimal categorical parameter prediction result.

[0065] In this embodiment, for categorical parameters (e.g., product type, process type, etc.), a sparse attention mask generator is first used to generate an initial score vector based on information derived from the flavor intent vector. This score vector reflects the initial scores of each possible option. A corresponding mask vector is designed based on process feasibility constraints (e.g., equipment limitations, material properties, etc.). This mask vector is used to filter out infeasible options, ensuring that only options that meet all process constraints are retained. Adding the mask vector to the original score vector automatically reduces or eliminates the scores of options that do not meet the constraints, resulting in a new score vector containing only feasible options. Finally, the masked score vector is processed using a Softmax function to convert it into a probability distribution, obtaining the probability that each feasible option will be the final selection. The option with the highest probability is the categorical parameter prediction result.

[0066] S1232. For continuous parameter prediction, unconstrained initial values ​​are predicted based on the parameter output vector. The process parameter generator uses the Sigmoid function and linear transformation to process the unconstrained initial values ​​to obtain the continuous parameter prediction results.

[0067] In this embodiment, the prediction result of the continuous parameter refers to the specific value of the continuous parameter obtained by adjusting the Sigmoid function mapping and linear transformation based on the unconstrained initial value, which satisfies the actual process requirements and optimizes specific objectives (such as efficiency, cost, etc.).

[0068] In this embodiment, for continuous parameters (such as temperature and time), a set of unconstrained initial values ​​are directly predicted based on the parameter output vector. These values ​​represent preliminary estimates without considering any process constraints. To ensure that the predicted values ​​meet actual process requirements (e.g., ensuring the temperature is within a reasonable range), these unconstrained initial values ​​need to be processed. Specifically, the values ​​are first mapped to the (0, 1) interval using the Sigmoid function, and then adjusted to the required range using a linear transformation to obtain the prediction results for the continuous parameters.

[0069] S1233. Combine the categorical parameter prediction results and the continuous parameter prediction results to obtain the prediction result.

[0070] Combining the prediction results of the two types of parameters above forms a complete prediction result. This means that not only is category information such as product type determined, but also the specific values ​​of relevant process parameters (such as temperature, time, etc.) are clarified.

[0071] By using this step-by-step processing method, step S123 can effectively combine process constraints and, starting from the flavor intention vector, generate the optimal parameter combination that meets both user expectations and production realities, greatly improving the feasibility and technical level of personalized customized products.

[0072] Based on the parameter output vector, the sparse attention mask generator creates a mask that defines which process options are feasible and which are not. For example, in some cases, to achieve a high-hardness product, the use of certain types of coagulants (such as lactone) may be prohibited. In this case, the mask will set extremely negative values ​​for the corresponding selections, making these options almost impossible to select.

[0073] The process parameter generator uses the aforementioned mask as guidance to filter out the optimal combination from a large pool of potential process parameter combinations, ensuring that all process constraints are met. These parameters not only guarantee the practical feasibility of production but also optimize product flavor and quality. For example, the final output might include a specific ratio of soybean A and soybean B, precise pulp concentration, and a suitable type and concentration of coagulant.

[0074] In summary, through these three steps, the method of this embodiment can effectively translate users' flavor requirements into specific production process parameters, thereby providing a scientific, efficient, and flexible way to optimize the production process of soy products. This method not only improves production efficiency and product quality but also promotes the innovative exploration of new flavors and processes.

[0075] In this embodiment, the constraint projection layer, based on the output vector of the DECN network, transforms the algebraic constraints generated by DCEN into a sparse attention mask for guided prediction. This is specifically implemented in different types of parameter prediction heads to achieve effective pruning and constraint of the prediction space.

[0076] Category parameters refer to selecting one from multiple discrete options, such as coagulant type (brinest, gypsum, lactone).

[0077] Suppose a categorical parameter has There are several optional values. The output of the corresponding prediction head before applying the Softmax activation function is a raw score vector, called the logits vector. Simultaneously, the sparse attention mask generator generates a mask vector for the head based on the input target flavor and the internal constraint knowledge base. .

[0078] Mask vector Each element The design is as follows: If the first If each option is feasible according to the constraint rules, then .

[0079] If the first If an option is prohibited according to the constraint rules, then... Given a very large negative number (For example, in implementation) ).

[0080] Add the mask vector to the original logits vector element-wise to obtain the masked logits vector. : ; Assume the target flavor is "high firmness," and the rule is "lactones are prohibited for high firmness." Let the coagulant logits be... (corresponding to brine, gypsum, and lactone), then the mask. After fusion .

[0081] The masked logits vector Input the Softmax function to calculate the th Final predicted probabilities for each category : ;in Representing vectors The The element is due to the prohibited item. It is a very large negative number, and its exponent is The value will approach zero infinitely. Therefore, the final predicted probability of this option... It is almost impossible for it to be selected in the prediction.

[0082] The impact of gradients during training: During backpropagation, the gradient calculation of the loss function (such as cross-entropy loss) with respect to the parameters is based on the final probability. This is due to the probability of prohibited terms... The gradient flowing to the network weights corresponding to that option will always be close to zero. This fundamentally prevents the model from learning in the "wrong" direction, forcing it to find the optimal solution within the rules.

[0083] Continuous parameters refer to variables that take values ​​within a range, such as slurry temperature and pressing pressure.

[0084] For these types of parameters, the sparse attention mask generator provides a dynamic feasible range. For example, when the goal is "high fineness," the feasible range of pressing pressure might be constrained to... kPa.

[0085] The parameter prediction head first outputs an unconstrained raw prediction value. The Sigmoid function is used to map the original value to the (0,1) interval, and then a linear transformation is performed to the target interval. This yields the final, constrained predicted value. : ;in, It's the Sigmoid function. This operation guarantees that regardless of the original value... How it changes, and the final output. Always strictly adhere to Within the range.

[0086] The gated scaling mechanism is fully differentiable. The final predicted value... For the original predicted value The gradient is: Because gradients can propagate smoothly backward, the model can "sense" the boundaries of the output during training and learn how to adjust its internal weights to generate values ​​that fall within that dynamic range, rather than performing a simple hard pruning after training.

[0087] Traditional methods typically perform "post-processing" to filter or correct the results after model prediction using external rules. This approach leads to inconsistencies between training and inference, as the model does not learn these constraints during training. This embodiment achieves complete unification of the training and inference processes by internalizing constraints as part of the network's computation graph. This "constraints as computation" paradigm ensures that the model follows domain knowledge at every step of learning, fundamentally improving the efficiency, accuracy, and reliability of predictions. Specifically, this embodiment achieves consistent adherence to constraints during both training and inference phases by directly embedding constraints into the network's computation graph, rather than relying on traditional external post-processing methods to filter or correct model predictions. Specifically, in categorical parameter prediction, the logits vector is dynamically adjusted using a sparse attention mask, eliminating options that do not meet the constraints before Softmax calculation; while in continuous parameter prediction, the Sigmoid function combined with linear transformation ensures that the predicted values ​​always remain within the allowable range. This method allows the model to "sense" these boundaries during the learning process and automatically adjust its internal weights based on domain knowledge to generate outputs that meet the constraints. Therefore, the "constraints as computation" paradigm not only enhances the model's ability to comply with domain-specific rules, but also fundamentally improves prediction efficiency, accuracy, and reliability, because it ensures that the entire process from the start of training to the end of inference closely matches the constraints required in real-world applications.

[0088] In one embodiment, step S120 above further includes step S124.

[0089] S124. After processing the prediction results through a multilayer perceptron, a risk vector corresponding to a predefined potential risk dimension is output, and the risk value of negative results caused by different process parameters is accurately predicted based on historical failure and boundary experiment data.

[0090] This embodiment employs a risk predictor, an independent prediction module running parallel to the main process parameter generator, designed to assess the potential risks posed by the process parameter vectors predicted by the main generator. This risk predictor not only increases the system's robustness but also enhances the predictability and controllability of negative outcomes.

[0091] The input is a vector of process parameters from the master process parameter generator, which is the prediction result. This vector represents a set of potentially "out-of-the-box" process parameters predicted by the model based on the current input.

[0092] The core component of the risk predictor includes a standalone multilayer perceptron (MLP). This type of neural network is well-suited for handling nonlinear mappings and can effectively transform input process parameter vectors into corresponding risk assessment values.

[0093] Output a risk vector R, whose dimensions correspond to a series of predefined potential risks. For example: Rbitterness: Indicates the probability of producing a bitter taste.

[0094] Rstructure_collapse: Indicates the risk of product structure collapse.

[0095] Roff_flavor: Indicates the probability of an unpleasant odor.

[0096] The training data comes from historical "failure experiments" and "boundary experiments," which record the negative results that occurred when specific process parameters were used. In this way, the model can learn which combinations of process parameters may lead to undesirable results. The specific steps are as follows: Collect historical process parameters that have led to negative outcomes (such as bitterness, structural breakdown, or off-flavors) and their corresponding negative outcome labels. For each set P of process parameters leading to a negative outcome, construct a training pair.<P,Rlabel> Rlabel is a label vector that indicates the occurrence of a specific negative outcome. For example, if a certain process parameter causes bitterness, Rbitterness is labeled as 1 in Rlabel, while other items are labeled as 0.

[0097] The model is trained using the training pairs constructed above, enabling the MLP to learn the mapping from process parameters to the probabilities of various negative outcomes. During training, the model continuously adjusts its weights to minimize the difference between the predicted risk and the actual risk.

[0098] This design enables the system not only to predict ideal process parameters but also to provide early warnings of potential problems, thereby helping users optimize production processes and avoid unnecessary losses. By learning directly from historical data, the risk predictor can provide reliable predictions on new, unseen data, greatly enhancing the practicality and reliability of the entire model. Furthermore, this approach promotes a more efficient and safer product development process, ensuring that every step is carried out within controllable limits.

[0099] In one embodiment, the training process of the above-mentioned process prediction neural network model includes: Historical production data, laboratory orthogonal experimental data, and physicochemical model simulation data are integrated to form standardized data; a neural network model and loss function are constructed; based on the standardized data and loss function, the Adam optimizer is used to update the parameters of the neural network model through gradient descent with an adaptive learning rate, and sparse attention mask pruning is used to search the space.

[0100] The neural network model includes the flavor target encoder, constraint equation network, sparse attention mask generator, process parameter generator, and risk predictor.

[0101] The loss function is formed by weighting the mean squared error loss of continuous parameters, the cross-entropy loss of discrete parameters, the interval constraint violation loss, the mask sparsity loss, and the risk prediction loss.

[0102] The discrete parameters include categorical parameters.

[0103] In this embodiment, the training of the model is an end-to-end supervised learning process, the goal of which is to minimize the difference between the process parameters predicted by the model and the actual process parameters.

[0104] The training data comes from three sources: a historical production database containing a large amount of process and flavor data for mature products; experimental data systematically generated in the laboratory through orthogonal experimental design; and simulation data generated using validated physicochemical models (such as gel formation kinetic models) to fill in sparse areas of real data. The basic unit constituting the training set is... Data pairs, among which It is a real vector of process parameters. This is the flavor result vector measured under this process. All continuous parameters (such as temperature, pressure, and flavor score) are standardized using Z-score to ensure they follow a normal distribution with a mean of 0 and a standard deviation of 1, thus accelerating model convergence. in, It is the original value. It is the mean of this feature on the training set. It is the standard deviation.

[0105] The model's total loss function It is a multi-task loss, composed of the weighted sum of the losses from each process parameter prediction head. Let the process parameter vector be... It can be divided into including A vector of continuous parameters that cannot tolerate deviations. and contain A vector of discrete (categorical) parameters and including A vector of parameters for a tolerance range. . ;in Continuous parameter loss (MSE) represents the mean square error loss for continuous parameters with unacceptable deviations. , For discrete parameter loss (Cross-Entropy), cross-entropy loss is used for category values ​​such as coagulant type. , It is the first The total number of categories for each discrete parameter. It is a sparse encoding of the real label (the real category position is 1, and the rest are 0). The model predicts that the category is The probability of. The loss for interval constraint violation is expressed as follows: , The expression for mask sparsity loss is as follows: ; Through the above constraints, the model is incentivized to generate a more discriminative mask, thereby guiding the model to learn more accurate and decisive process rules.

[0106] This loss function is used to predict the probability of various predefined negative risks arising from the execution of a certain "out-of-bounds" process. It employs a binary cross-entropy form, and the calculation formula is as follows: ,in For batch size, The number of risk categories defined for the system. Indicates the first The sample at the th True binary label (0 or 1) on risk class. This represents the predicted probability (between 0 and 1) of the risk prediction head for the corresponding risk. The loss function, calculated by averaging the cross-entropy over all risks and all samples, aims to improve the accuracy of the risk prediction head in assessing the probability of various negative consequences that unconventional processes may cause.

[0107] Task preference weight , , , as well as To balance the importance of different tasks, it can be set empirically or optimized as a hyperparameter.

[0108] Gradient descent is performed using the Adam optimizer. Adam combines the advantages of momentum and RMSprop, enabling it to compute adaptive learning rates for different parameters.

[0109] Parameter Update: In each training batch, the model weights are updated according to the following formula: ; ; ; ; ;in, Is the model in the first Step parameters, It is a loss function pair gradient, It's the learning rate. and These are the first-order and second-order moment estimates, respectively. It is the attenuation rate. It is a small constant that prevents the denominator from being zero.

[0110] The role of sparse attention masks is reflected in gradient calculation. During the process, when the mask forces the predicted probability of a certain process option to zero (e.g., by adding a very large negative number to logits), the gradient flowing to the corresponding parameter of that option will also be zero during backpropagation. This fundamentally prevents the model from learning and updating on the "forbidden" paths, thus achieving effective pruning of the search space.

[0111] For example, to optimize and predict tofu production processes, a total of 5000 sets of tofu production data need to be collected or generated. Each set of data contains two main parts: process parameters (P) and flavor parameters (F). Process parameters include the proportion of soybean A (%), the proportion of soybean B (%), the grinding concentration (soybean-to-water ratio), the type of coagulant (bituminous brine / gypsum / glucono delta-lactone), the coagulant concentration (Baumé degree), the coagulation temperature (°C), the pressing pressure (kPa), and the pressing time (min). Flavor parameters cover hardness (N), elasticity (%), soybean aroma (GC-MS peak area), fishy smell (GC-MS peak area), and smoothness (sensory score 1-10). All data are normalized to ensure the accuracy of model training. 80% of the data will be used for model training, and the remaining 20% ​​will be used as the test set.

[0112] The DECN-SAN model architecture is adopted, and a pre-set constraint knowledge base contains approximately 20 rules, such as "when the hardness is greater than 8.0, the coagulant type should not be lactone" and "if the fineness exceeds 8.0, the pressing pressure should be less than 10 kPa". These rules are designed to guide the model's prediction direction under specific conditions and avoid unreasonable results.

[0113] One example is the optimal process prediction for making rich-flavored Northern tofu.

[0114] Target flavor inputs: Hardness: 9.0 / 10; Elasticity: 7.0 / 10; Bean aroma: 9.5 / 10; Fishy smell: 1.0 / 10; Fineness: 4.0 / 10.

[0115] Model prediction process: The flavor target encoder identifies that the user wants a product with high firmness, strong bean aroma and low fishy smell.

[0116] The sparse attention mask generator is activated and creates a mask that cancels the possibility of selecting lactone as a coagulant and limits the pressing pressure to a higher range (e.g., >20 kPa).

[0117] The process parameter generator makes predictions based on the mask mentioned above.

[0118] The optimal process parameters output by the model are: soybean A ratio: 100%; soybean B ratio: 0%; pulping concentration: 1:6; coagulant type: brine; coagulant concentration: 20 Baume degrees; slurry temperature: 88°C; pressing pressure: 30 kPa; pressing time: 25 min.

[0119] Another example is the process prediction for innovative chewy tofu, with the following target flavor inputs: hardness: 7.0 / 10; elasticity: 9.5 / 10 (core appeal); bean aroma: 7.0 / 10; fishy smell: 2.0 / 10; and smoothness: 6.0 / 10.

[0120] Model Prediction Process: The model first determines that improving elasticity is the primary task. According to the knowledge base, "high elasticity" is usually associated with the use of composite coagulants and secondary curing processes. Therefore, the mask generator increases the probability of using gypsum and brine in combination.

[0121] The optimal process parameters output by the model are: soybean A ratio: 70%; soybean B ratio: 30%; pulping concentration: 1:7; coagulant type: gypsum + brine composite; coagulation temperature: 85°C; pressing pressure: 25 kPa; pressing time: 20 min.

[0122] The last example is the handling of goals that do not conform to conventional logic. The target flavor input (illogical): Hardness: 9.0 / 10 (extremely hard); Fineness: 9.5 / 10 (extremely smooth).

[0123] Model prediction process: DECN performs analysis. Hard constraints work normally, while soft constraints identify conflicts. The empirical requirement for "high hardness" is "coagulant = brine / gypsum", while the empirical requirement for "high fineness" is "coagulant = lactone". DECN does not directly prohibit either one, but instead generates two Gaussian penalty functions simultaneously, one centered on "lactone" and the other on "brine / gypsum".

[0124] Process parameter generator exploration: During optimization, the generator discovered that strictly adhering to the "high hardness" rule (using brine) resulted in extremely high "fineness" values ​​in the flavor prediction loss. Conversely, the opposite was also true. To minimize the total loss... The model might find an "out-of-the-box" solution: for example, using lactone as the primary coagulant, but combined with an unconventional, extremely high curing temperature and a slight secondary pressurization process. This combination does not exist in conventional experience. Although this solution will incur some interval constraint violation losses. However, because it theoretically achieves a better balance between hardness and smoothness, it significantly reduces the loss of other parameters, thus minimizing the overall loss. To reach the minimum.

[0125] The module received this unconventional formula of "lactone + high temperature + micro pressure". Based on the knowledge learned from failed experiments, it predicts the risk vector. "The structure may be partially flocculent, with a moderate (45%) risk of producing a slightly acidic taste."

[0126] Final output to the user:

[0127] Recommended Process 1 (Standard): An experience-based compromise formula that balances hardness and smoothness (e.g., plaster + low pressure), labeled "Expected Result: Hardness 7.0, Smoothness 7.5".

[0128] Exploratory Process 2 (Innovation): A formulation combining "lactone + high temperature + micro-pressure," with an attached risk report: This formulation may achieve a unique combination of hardness 8.0 and fineness 8.5, but carries a moderate risk of acidity and structural flocculation. Small-batch trials are recommended, with close monitoring of pH changes.

[0129] The constraint equation network of the method in this embodiment can dynamically and in real time transform pre-defined physicochemical laws, expert experience, and other domain knowledge into a set of differentiable constraint equations. These equations directly define the process-feasible solution space under the current target flavor. Subsequently, a constraint projection layer generates a sparse attention mask based on this set of equations. This mask guides the subsequent process parameter generator, ensuring that any combination of parameters it outputs naturally and algebraically satisfies all process constraints. By designing these constraints as differentiable equations embedded in the model, the neural network can directly learn and predict within the solution space that satisfies the equations. At the same time, for some parameters that tolerate deviations from empirical values, the existence of deviations is allowed, and these deviations are evaluated through a risk function. This ensures the stability and novelty of the process while also preserving the possibility of new processes and products emerging beyond traditional experience.

[0130] The aforementioned intelligent prediction method for food processing technology based on target flavor transforms target flavor parameters into standardized target flavor vectors, which are then input into a process prediction neural network model. A flavor target encoder converts these vectors into vectors representing specific flavor intentions. Next, a constraint equation network dynamically generates parameter output vectors that meet process requirements based on this flavor intention vector. These parameter output vectors embed differentiable constraint equations, representing the solution space under specific process conditions. Subsequently, a sparse attention mask generator creates a mask containing process feasibility constraints based on these parameter output vectors, ensuring that the optimal parameter combination output by the process parameter generator naturally satisfies all preset process constraints. Furthermore, the introduction of a soft constraint mechanism allows for deviations in certain parameters. Risk function evaluation ensures process stability while also providing the possibility of exploring new processes and products beyond traditional experience. This enables the neural network to effectively learn and predict within the solution space that satisfies the constraints.

[0131] Figure 3 This is a schematic block diagram of a food processing technology intelligent prediction system 300 based on target flavor provided in an embodiment of the present invention. Figure 3 As shown, corresponding to the above-described intelligent prediction method for food processing technology based on target flavor, the present invention also provides an intelligent prediction system 300 for food processing technology based on target flavor. This intelligent prediction system 300 includes a unit for executing the above-described intelligent prediction method for food processing technology based on target flavor, and the system can be configured in a server. Specifically, please refer to... Figure 3The intelligent prediction system 300 for food processing technology based on target flavor includes an acquisition unit 301 and a prediction unit 302.

[0132] The acquisition unit 301 is used to acquire target flavor parameters and construct a target flavor vector based on the target flavor parameters; the prediction unit 302 is used to input the target flavor vector into the process prediction neural network model to predict the processing technology and obtain the prediction result. The prediction unit 302 includes: The encoding subunit is used to input the target flavor vector into the process prediction neural network model, and convert the target flavor vector into a vector representing a specific flavor intent through the flavor target encoder to obtain the flavor intent vector; the constraint processing subunit is used to generate a parameter output vector that meets the process requirements based on the flavor intent vector by the constraint equation network; the generation subunit is used to create a mask containing process feasibility constraints based on the parameter output vector using a sparse attention mask generator, and the process parameter generator outputs the optimal parameter combination that satisfies all process constraints to form the prediction result.

[0133] In one embodiment, the encoding subunit is used to normalize the target flavor vector and process it through a fully connected layer by a flavor target encoder, and then convert it into a high-dimensional vector representing a specific flavor intent, so as to obtain a flavor intent vector.

[0134] In one embodiment, the constraint processing subunit is used to process the flavor intention vector by the constraint equation network using a multi-head output neural network structure, so that each output head outputs one or a set of related parameters that need to be dynamically generated in the corresponding differentiable constraint equation, and combines the parameters to generate a parameter output vector that meets the process requirements.

[0135] In one embodiment, the generating subunit includes: The first prediction module is used for categorical parameter prediction. Based on the parameter output vector, it generates an original score vector using a sparse attention mask generator, designs a mask vector according to constraint rules, and adds it to the original score vector to filter infeasible options, forming a masked score vector. The process parameter generator processes the masked score vector using a Softmax function to obtain the categorical parameter prediction result. The second prediction module is used for continuous parameter prediction. Based on the parameter output vector, it predicts unconstrained initial values. The process parameter generator processes the unconstrained initial values ​​using a Sigmoid function and a linear transformation to obtain the continuous parameter prediction result. The combination module is used to combine the categorical parameter prediction result and the continuous parameter prediction result to obtain the prediction result.

[0136] In one embodiment, the prediction unit 302 further includes: The risk prediction subunit is used to process the prediction results through a multilayer perceptron and output a risk vector corresponding to a predefined potential risk dimension, and accurately predict the risk value of negative results caused by different process parameters based on historical failure and boundary experiment data.

[0137] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned intelligent prediction system 300 for food processing technology based on target flavor and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0138] The aforementioned intelligent prediction system 300 for food processing technology based on target flavor can be implemented as a computer program, which can be used in, for example... Figure 4 It runs on the computer device shown.

[0139] Please see Figure 4 , Figure 4 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0140] See Figure 4 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0141] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform an intelligent prediction method for food processing technology based on a target flavor.

[0142] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0143] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a smart prediction method for food processing technology based on target flavor.

[0144] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0145] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the intelligent prediction method for food processing technology based on target flavor.

[0146] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0147] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0148] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all steps of the intelligent prediction method for food processing technology based on target flavor.

[0149] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0150] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0151] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0152] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0153] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0154] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent prediction of food processing technology based on target flavor, characterized in that, include: Obtain the target flavor parameters and construct a target flavor vector based on the target flavor parameters; The target flavor vector is input into the process prediction neural network model to predict the processing technology and obtain the prediction result. The step of inputting the target flavor vector into a process prediction neural network model to predict the processing technology and obtain the prediction result includes: The target flavor vector is input into the process prediction neural network model, and the target flavor vector is converted into a vector representing a specific flavor intent by the flavor target encoder to obtain the flavor intent vector. The constraint equation network generates a parameter output vector that meets the process requirements based on the flavor intention vector; Based on the parameter output vector, a sparse attention mask generator is used to create a mask containing process feasibility constraints, and the process parameter generator outputs the optimal combination of parameters that satisfies all process constraints to form the prediction result.

2. The intelligent prediction method for food processing technology based on target flavor according to claim 1, characterized in that, The process of converting the target flavor vector into a vector representing a specific flavor intent using a flavor target encoder to obtain a flavor intent vector includes: After normalizing the target flavor vector and processing it through a fully connected layer by a flavor target encoder, it is converted into a high-dimensional vector representing a specific flavor intent, thus obtaining a flavor intent vector.

3. The intelligent prediction method for food processing technology based on target flavor according to claim 1, characterized in that, The generation of parameter output vectors that meet process requirements by the constraint equation network based on the flavor intent vector includes: Based on the flavor intent vector, the constraint equation network is processed using a multi-head output neural network structure, so that each output head outputs one or a set of related parameters that need to be dynamically generated in the corresponding differentiable constraint equation, and the parameters are combined to generate a parameter output vector that meets the process requirements.

4. The intelligent prediction method for food processing technology based on target flavor according to claim 1, characterized in that, The process involves using a sparse attention mask generator to create a mask containing process feasibility constraints based on the parameter output vector, and then having a process parameter generator output the optimal combination of parameters that satisfies all process constraints to form the prediction result, including: For categorical parameter prediction, based on the parameter output vector, a sparse attention mask generator is used to generate the original score vector, and a mask vector is designed according to the constraint rules. The mask vector is added to the original score vector to filter infeasible options, forming a masked score vector. The process parameter generator processes the masked score vector through the Softmax function to obtain the categorical parameter prediction result. For continuous parameter prediction, unconstrained initial values ​​are predicted based on the parameter output vector. The process parameter generator then processes these unconstrained initial values ​​using the Sigmoid function and linear transformation to obtain the continuous parameter prediction results. The prediction results of the categorical parameters and the prediction results of the continuous parameters are combined to obtain the prediction result.

5. The intelligent prediction method for food processing technology based on target flavor according to claim 1, characterized in that, The process of creating a mask containing process feasibility constraints using a sparse attention mask generator based on the parameter output vector, and then having the process parameter generator output the optimal parameter combination that satisfies all process constraints to form the prediction result, further includes: After the prediction results are processed by a multilayer perceptron, a risk vector corresponding to a predefined potential risk dimension is output, and the risk value of negative results caused by different process parameters is accurately predicted based on historical failure and boundary experiment data.

6. The intelligent prediction method for food processing technology based on target flavor according to claim 1, characterized in that, The training process of the process prediction neural network model includes: Integrate historical production data, laboratory orthogonal experimental data, and physicochemical model simulation data to form standardized data; Constructing a neural network model and its loss function; Based on the standardized data and loss function, the Adam optimizer is used to update the parameters of the neural network model using gradient descent with an adaptive learning rate, and sparse attention mask pruning is used to search the space.

7. The intelligent prediction method for food processing technology based on target flavor according to claim 6, characterized in that, The neural network model includes the flavor target encoder, constraint equation network, sparse attention mask generator, process parameter generator, and risk predictor.

8. The intelligent prediction method for food processing technology based on target flavor according to claim 7, characterized in that, The loss function is formed by weighting the mean squared error loss of continuous parameters, the cross-entropy loss of discrete parameters, the interval constraint violation loss, the mask sparsity loss, and the risk prediction loss.

9. The intelligent prediction method for food processing technology based on target flavor according to claim 8, characterized in that, The discrete parameters include categorical parameters.

10. A smart predictive system for food processing technology based on target flavor, characterized in that, include: An acquisition unit is used to acquire target flavor parameters and construct a target flavor vector based on the target flavor parameters; The prediction unit is used to input the target flavor vector into the process prediction neural network model to predict the processing technology and obtain the prediction result. The prediction unit includes: The encoding subunit is used to input the target flavor vector into the process prediction neural network model, and convert the target flavor vector into a vector representing a specific flavor intention through the flavor target encoder to obtain the flavor intention vector; The constraint processing subunit is used to generate a parameter output vector that meets the process requirements based on the flavor intention vector from the constraint equation network; The generation sub-unit is used to create a mask containing process feasibility constraints based on the parameter output vector using a sparse attention mask generator, and the process parameter generator outputs the optimal combination of parameters that satisfies all process constraints to form the prediction result.

Images