Decision-making system for predicting flavor formation mechanism and flavor optimization in food processing based on machine learning

Abstract

The invention relates to the technical field of food processing, in particular to a system for predicting flavor formation and optimization decision in food processing based on machine learning, which comprises a sensing unit, a multi-source data acquisition and fusion module, a dynamic modeling module, an optimization decision module and an execution module which are in signal connection with one another, and the optimization decision module is used for receiving the flavor perception probability distribution data and the updated scoring reference data, solving a Pareto optimal solution set through a multi-target particle swarm optimization algorithm in combination with equipment physical constraint conditions, generating a candidate processing parameter scheme, inverting equipment control parameters for the candidate processing parameter scheme through a physical constraint neural network, and obtaining the flavor perception probability distribution data and the updated scoring reference data. And a final machining parameter adjusting instruction is generated and transmitted to the execution module. According to the method, through dynamic threshold modeling, a semantic-chemical attention mechanism and a time-space preference map dynamic correction technology, multi-source data and a multi-target optimization algorithm are fused, so that closed-loop accurate regulation and control of processing parameters are realized, and the flavor quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of food processing, and in particular to a machine learning-based prediction system for flavor formation in food processing and an optimization decision-making system. Background Art

[0002] In the food processing industry, flavor is a key factor influencing consumer acceptance. Flavor formation is complexly influenced by raw material characteristics, processing parameters (such as temperature and time), and environmental conditions. Traditional methods rely on manual experience or single-indicator measurement to adjust process parameters, making it difficult to accurately quantify the dynamic relationship between flavor compounds and sensory preferences. This can lead to unstable product quality, cost waste, and slow market response. Therefore, there is an urgent need to develop data-driven flavor prediction and optimization decision-making systems to achieve scientific control of processing parameters and personalized flavor design.

[0003] In existing food prediction and optimization decision-making systems, patent publication number CN118820876A proposes a rapid fruit variety prediction method based on flavor analysis. This method relates to the field of fruit variety identification and authentication, addressing the existing technical problem of a lack of a rapid and accurate method for distinguishing fruit varieties. The method includes: obtaining the maximum response values of different fruit samples measured by an electronic nose system (Enose) and the composition and content of volatile compounds measured by gas chromatography-mass spectrometry (GCMS); performing feature screening on the volatile compounds to obtain several key volatile compounds, and constructing several compound feature prediction models and electronic nose feature prediction models based on multiple machine learning algorithm models. This invention uses GCMS and Enose to evaluate the odor characteristics of fruit and develops a rapid and accurate method for distinguishing fruit varieties based on machine learning algorithms. This method can be used to address the quality differences of fruit commodities, provide a reference for fruit quality evaluation research, and has high application and promotion value.

[0004] However, existing technologies suffer from the following key flaws: First, objective indicators such as hexanal concentration detected by GC-MS exhibit a significant nonlinear relationship with consumer sensory scores. Traditional linear regression models are unable to capture threshold mutation characteristics, resulting in parameter optimization deviations from actual needs. Second, there is a lack of interpretable quantitative rules between instrument data and sensory descriptors. Existing methods rely on manually set fixed weights, ignoring the impact of concentration gradients on semantic intensity, resulting in distorted optimization targets. Furthermore, consumer preferences vary dynamically with seasons and regions, but existing systems are trained on static sensory datasets, with parameter update cycles lasting several months. This makes it impossible to adapt to market demand in real time, resulting in reduced product competitiveness. These shortcomings severely hinder the development of intelligent food processing, and breakthroughs are urgently needed through multi-source data fusion, dynamic modeling, and real-time optimization technologies. Summary of the Invention

[0005] To address the above problems, the present invention provides a machine learning-based prediction and optimization decision-making system for flavor formation in food processing. By integrating dynamic threshold modeling, semantic-chemical attention mechanism, and spatiotemporal preference map dynamic correction technology, multi-source data and multi-objective optimization algorithms are used to achieve closed-loop precise control of processing parameters and improve flavor quality.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: a machine learning-based prediction and optimization decision-making system for flavor formation in food processing, comprising a sensor unit connected to each other, a multi-source data acquisition and fusion module, a dynamic modeling module, an optimization decision-making module, and an execution module, wherein: The sensor unit is used to acquire and integrate sensor data from multiple different channels and transmit it to the multi-source data acquisition module. The sensor data includes temperature, time, and pressure sensor signals from processing equipment, physical and chemical parameters of food processing scenes, peak area data from chromatography-mass spectrometry, text data of consumer sensory evaluations, and geographic location, season, temperature and humidity data from environmental sensors. The multi-source data acquisition and fusion module is used to receive sensor data and generate corresponding multi-source data. The multi-source data includes time series data of equipment operation status, quantitative concentration data of target volatile compounds, standardized sensory score data, physical and chemical parameters of food matrix, and environmental feature vectors. After feature alignment and normalization of the multi-source data, a multi-dimensional joint feature matrix is obtained and transmitted to the dynamic modeling module. A dynamic modeling module is used to receive a multi-dimensional joint feature matrix and synchronously perform model training through a multi-task deep neural network. The dynamic modeling module includes a nonlinear mapping submodule, a knowledge transfer submodule, and a preference tracking submodule; A nonlinear mapping submodule is used to fit the threshold mutation curve of the sensory score using a piecewise activation function based on the volatile compound concentration data, generate flavor perception probability distribution data, and transmit it to the optimization decision module; The knowledge transfer submodule is used to receive a database of material synergies from an external industrial field, eliminate scenario differences through feature distillation and domain adaptation algorithms, generate transfer-enhanced model parameters, and update the transfer-enhanced model parameters to the activation function of the nonlinear mapping submodule in real time, so that the generated flavor perception probability distribution data adapts to the current processing scenario; The preference tracking submodule is used to construct a spatiotemporal preference map based on the environmental feature vector using a graph convolutional network, dynamically modify the sensory scoring weight coefficient, generate updated scoring benchmark data, and transmit it to the optimization decision module; The optimization decision module is used to receive the flavor perception probability distribution data and the updated scoring benchmark data, and solve the Pareto optimal solution set through the multi-objective particle swarm optimization algorithm in combination with the physical constraints of the equipment to generate candidate processing parameter solutions. The equipment control parameters of the candidate processing parameter solutions are inverted through the physical constraint neural network to generate the final processing parameter adjustment instructions and transmit them to the execution module; The execution module is used to receive processing parameter adjustment instructions. The execution module signal is connected to the equipment controller to achieve closed-loop control of the processing parameters through the equipment controller.

[0007] Furthermore, the multi-source data acquisition and fusion module includes a device sensing unit, a material detection unit, a sensory analysis unit, an environmental perception unit, and a data fusion unit, wherein: The equipment sensing unit is used to receive the temperature, time and pressure sensing signals of the processing equipment transmitted by the sensing unit, process the useful information frequency higher than the sensing signal itself through the wavelet transform denoising algorithm, generate time series data of the equipment operation status, and transmit it to the data fusion unit; The substance detection unit is used to receive the detection peak area data of the chromatography-mass spectrometer transmitted by the sensing unit, calibrate it using the internal standard method, and then eliminate the chromatographic drift error using the retention index correction algorithm. After performing double calibration, it generates quantitative concentration data of the target volatile compound and transmits it to the data fusion unit. It is also used to obtain the physicochemical parameters of the food matrix through the dielectric sensor and water activity meter based on the received physicochemical parameters of the food processing scene, and transmit them to the knowledge transfer submodule; The sensory analysis unit is used to receive the consumer sensory evaluation text data transmitted by the sensing unit, extract semantic feature vectors through the natural language processing model, generate standardized sensory score data, and transmit it to the data fusion unit; The environmental perception unit is used to receive the geographic location, season, temperature and humidity data transmitted by the sensing unit, generate an environmental feature vector through a spatiotemporal coding algorithm, and transmit it to the data fusion unit; The data fusion unit is used to receive time series data, quantitative concentration data, sensory score data and environmental feature vectors, match the time series of different data sources through the timestamp alignment algorithm, use feature embedding technology to map heterogeneous data to a unified dimensional space, perform Z-score normalization on the mapped multidimensional features to eliminate dimensional differences, dynamically allocate the weight of each feature through the attention mechanism, generate a multidimensional joint feature matrix including equipment status, substance concentration, sensory score and environmental factors, and transmit it to the dynamic modeling module.

[0008] Furthermore, the nonlinear mapping submodule includes a threshold modeling unit and a semantic fusion unit, wherein: The threshold modeling unit is used to receive the quantitative concentration data in the multidimensional joint feature matrix. The quantitative concentration data includes the time series concentration values of the target compound during the processing process, and the values are calculated in the preset dynamic threshold interval. A differentiable piecewise activation function is used to construct a concentration-sensory score mapping model, where the differentiable piecewise activation function is defined as: Where c is the concentration of the target volatile compound, is the dynamic sensory mutation threshold, and are all linear gain coefficients, and All are nonlinear adjustment factors; Dynamic adjustment through Bayesian optimization framework , the optimization target is the maximum sensory score mutation point in historical data and The matching degree of , its acquisition function is defined as the improved expected improvement function: in, Indicates that the threshold The newly collected sensory scores are represents the best sensory score in historical data, The concentration-sensory association data including the threshold response curve are generated as the standard deviation of the sensory scores and transmitted to the semantic fusion unit.

[0009] Furthermore, the semantic fusion unit is used to receive the consumer evaluation text data transmitted by the sensory analysis unit. The text data contains several different flavor description words and their frequencies of occurrence. The pre-trained language model is used to extract the contextual semantic features of the text and generate a word vector matrix with a dimension of [d×1] ; The semantic features are associated with the concentration data output by the threshold modeling unit through the attention mechanism to calculate the semantic-chemical weight matrix: in, represents the association strength between the i-th chemical substance and the j-th semantic label, which is used to quantify the effect of concentration on sensory description; is the eigenvector of chemical concentration data, is the semantic feature vector extracted from the consumer evaluation text, is the word vector corresponding to the semantic label, and d is the feature dimension; Generate a 3D tensor containing association rules between semantic strength and chemical concentration , where N is the type of compound and M is the number of semantic labels. The three-dimensional tensor is transferred to the optimization decision module.

[0010] Furthermore, the knowledge transfer submodule includes a rule distillation unit and a domain alignment unit, where: The rule distillation unit is used to receive the material synergy database of the external industrial field, which contains the synergistic and inhibitory relationship data of cross-domain materials; it is also used to extract high-order interaction patterns between materials through the graph attention network to generate a cross-domain knowledge graph , where the vertex set V represents the material category, the edge set E represents the synergistic relationship, and the weight W represents the intensity of the effect. The knowledge graph is transferred to the domain alignment unit; Domain alignment unit for receiving physicochemical parameters of food processing scenarios, including matrix dielectric constant and water activity , calculate the difference in physical properties between external industry and food processing: Optimize the transfer loss function through the domain adaptation algorithm: in, and are adaptive weight coefficients, Represents the maximum mean difference between the external industry and food processing knowledge graphs, generates the model parameters after migration enhancement, and updates them to the threshold modeling unit.

[0011] Furthermore, the preference tracking submodule includes a graph construction unit and a weight updating unit, where: The graph construction unit is used to receive the environmental feature vector transmitted by the environmental perception unit. The environmental feature vector contains latitude and longitude coordinates and regional cultural feature labels. The modularity-optimized Louvain algorithm is used to divide the flavor preference community and generate the region-season-preference association matrix. , where r represents the number of regions, S is the number of seasons, and K is the preference dimension, and is transmitted to the weight update unit; The weight update unit is used to receive the temperature and humidity time series data in the environmental feature vector, use the LSTM network to predict the preference drift trend, and dynamically correct the sensory score weight coefficient to generate updated score benchmark data and transmit it to the optimization decision module.

[0012] Furthermore, the optimization decision module includes a constraint processing unit, a particle swarm optimization unit, and a physical inversion unit, wherein: The constraint processing unit is used to receive time series data of the equipment operation status, including the maximum allowable temperature, pressure threshold and energy consumption per unit time threshold, generate a multi-dimensional feasible solution space through the constraint space modeling algorithm, encode the feasible solution space into a boundary constraint matrix, and transmit it to the particle swarm optimization unit; The particle swarm optimization unit receives the flavor perception probability distribution data and updated rating benchmark data transmitted by the dynamic modeling module, and receives the device feasible solution space boundary matrix generated by the constraint processing unit. It performs parallel search in the feasible solution space through the target particle swarm optimization algorithm, combined with the semantic-chemical association tensor Dynamically adjust the weight distribution of flavor intensity and sensory preference to generate a Pareto candidate solution set that maximizes flavor intensity and optimally matches sensory preference. The selected solution set is then transferred to the physical inversion unit for feasibility verification of processing parameters. The physical inversion unit is used to receive the Pareto candidate solution set transmitted by the particle swarm optimization unit. The solution set contains processing parameter combinations that meet the flavor and preference optimization. The feasibility of the parameters is verified through a neural network model embedded with physical constraints. The parameter combinations that have passed the physical feasibility verification are screened, and the processing parameter adjustment instructions are generated. The adjustment instructions are transmitted to the execution module.

[0013] Furthermore, the execution module includes an instruction parsing unit and a closed-loop control unit, wherein: An instruction parsing unit is used to receive the processing parameter adjustment instructions transmitted by the physical inversion unit. The instructions include temperature setting values, time setting values, and pressure setting values. The instruction is verified for physical feasibility through a finite state machine, and a control instruction sequence executable by the equipment is generated and transmitted to the closed-loop control unit. The closed-loop control unit is used to receive the control instruction sequence, adjust the equipment operating parameters in real time through the proportional-integral-differential algorithm, and output the pulse width modulation signal to the processing equipment to realize the closed-loop dynamic control of the processing parameters.

[0014] Furthermore, in the dynamic modeling module, the multi-task deep neural network synchronous execution processing adopts a hard parameter sharing architecture. The shared layer is a 3-layer bidirectional LSTM network, and the private task is a fully connected network. The dynamic weight averaging strategy is adopted when training the model. The task weight update formula is: in, represents the weight of task k in the tth round of training, N represents the total number of tasks, represents the temperature coefficient, It represents the relative loss reduction rate of the t-1 round under a certain task, and K represents the total number of three subtasks of the nonlinear mapping submodule, knowledge transfer submodule and preference tracking submodule.

[0015] Furthermore, the multi-source data acquisition and fusion module also includes an anomaly detection unit, which is used to detect abnormal data of the sensor unit in real time and uses the isolation forest algorithm to make real-time judgments: Expressed as, if the isolation path length PathLength of a data point x is less than the threshold , then x is determined to be an outlier, where and They represent the mean and standard deviation of the normal data path length respectively. When the abnormality detection unit detects abnormal data from the sensing unit, the redundant sensing data replacement mechanism is activated.

[0016] The above scheme has the following beneficial effects: 1. This solution addresses the problem of linear models in existing technologies being unable to capture the nonlinear threshold mutations between volatile compound concentrations and sensory scores. This system dynamically fits the concentration-score mapping relationship using a differentiable piecewise activation function and a Bayesian optimization framework, solving the problem of adaptively identifying threshold mutation points. By dynamically adjusting the sensory mutation threshold, the system can accurately capture, for example, the positive "freshness" effect of hexanal in the low concentration range and the negative "grassy" effect in the high concentration range. This overcomes the problem of process parameter overshoot or undershoot caused by fixed thresholds in traditional methods, significantly improving the sensitivity and accuracy of flavor prediction.

[0017] 2. In this solution, traditional methods rely on manually setting sensory evaluation weights, resulting in a lack of explainable quantitative associations between chemical detection data and semantic descriptors. This system innovatively introduces an attention mechanism and cross-domain knowledge graph transfer technology. Through a semantic fusion unit, text features such as "rich burnt aroma" described by consumers are mapped to the chemical concentration space, generating a three-dimensional tensor of semantic-chemical association rules. At the same time, a domain adaptation algorithm is used to eliminate the physical property differences (such as dielectric constant and water activity) between the chemical industry and food processing scenarios, achieving efficient cross-domain knowledge transfer. As a result, the system can dynamically quantify the association strength between flavor descriptors and compound concentrations, avoiding the subjective bias of manual weight setting and ensuring that the optimization goals are highly consistent with real consumer preferences.

[0018] 3. This solution addresses the parameter lag caused by the static datasets used in existing systems. By combining a graph convolutional network with an LSTM time series prediction model, this system constructs a spatiotemporal preference map, tracking in real time the impact of seasonal variations (e.g., a preference for rich flavors in winter) and regional cultural differences (e.g., a preference for burnt aromas in Northern Europe and fruity aromas in Southeast Asia) on sensory ratings. Combined with a modularity-optimized community segmentation algorithm, this system dynamically adjusts the sensory weight coefficients for different regions, enabling recommended parameters to adapt to shifting consumer preferences in real time. Compared to the multi-month update cycles of traditional static models, this system achieves millisecond-level dynamic response, completely resolving the quality fluctuations caused by parameter lag.

[0019] 4. In this solution, traditional optimization algorithms ignore equipment physical constraints (such as upper temperature limits and energy consumption thresholds) and chemical kinetics, easily generating infeasible solutions. This system, through the collaborative design of a multi-objective particle swarm optimization algorithm and a physical information neural network, embeds equipment operating constraints and chemical mechanism models such as the Arrhenius equation within the Pareto optimal solution set, ensuring that candidate parameters simultaneously meet the requirements of maximizing flavor intensity, optimizing sensory preference matching, and physical feasibility. Combined with the closed-loop dynamic control of the execution module, the system can accurately convert optimization instructions into equipment control signals, achieving closed-loop management of the entire chain from data perception to process execution.

[0020] 5. This solution addresses issues such as sensor data noise and heterogeneous data timing misalignment. This system utilizes wavelet transform denoising, timestamp alignment, and attention weight allocation techniques to construct a multidimensional joint feature matrix, effectively improving data fusion accuracy. Furthermore, the system integrates an isolation forest anomaly detection algorithm with a redundant sensor switching mechanism to dynamically identify sensor drift or signal distortion and automatically switch to a backup data source, ensuring high system robustness in complex industrial environments.

[0021] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of the system framework of an embodiment of the system for predicting flavor formation and optimizing decision-making in food processing based on machine learning according to the present invention; Figure 2 This is a schematic diagram of a multi-source data acquisition and fusion module framework of an embodiment of a system for predicting flavor formation and optimizing decision-making in food processing based on machine learning of the present invention; Figure 3 Schematic diagram of the dynamic modeling module framework of an embodiment of the present invention's system for predicting flavor formation and optimizing decision-making in food processing based on machine learning. DETAILED DESCRIPTION

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0024] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0026] The following is further described in detail through specific implementation methods: Example 1: As attached Figure 1 、 Figure 2 and Figure 3 As shown: A machine learning-based prediction and optimization decision-making system for flavor formation in food processing includes a sensor unit with interconnected signals, a multi-source data acquisition and fusion module, a dynamic modeling module, an optimization decision module, and an execution module, wherein: The sensing unit is used to acquire and integrate sensor data from several different channels and transmit it to the multi-source data acquisition module. The sensor data includes temperature, time, and pressure sensor signals of processing equipment, physical and chemical parameters of food processing scenarios, detection peak area data of chromatography-mass spectrometry, consumer sensory evaluation text data, and geographic location, season, temperature and humidity data of environmental sensors.

[0027] The multi-source data acquisition and fusion module is used to receive sensor data and generate corresponding multi-source data. The multi-source data includes time series data of equipment operation status, quantitative concentration data of target volatile compounds, standardized sensory score data, physical and chemical parameters of food matrix, and environmental feature vectors. After feature alignment and normalization of the multi-source data, a multi-dimensional joint feature matrix is obtained and transmitted to the dynamic modeling module. The multi-source data acquisition and fusion module includes an equipment sensor unit, a substance detection unit, a sensory analysis unit, an environmental perception unit, and a data fusion unit, wherein: The equipment sensing unit is used to receive the temperature, time and pressure sensing signals of the processing equipment transmitted by the sensing unit, process the useful information frequency higher than the sensing signal itself through the wavelet transform denoising algorithm, generate time series data of the equipment operation status, and transmit it to the data fusion unit; The substance detection unit is used to receive the detection peak area data of the chromatography-mass spectrometer transmitted by the sensing unit, calibrate it using the internal standard method, and then eliminate the chromatographic drift error using the retention index correction algorithm. After performing double calibration, it generates quantitative concentration data of the target volatile compound and transmits it to the data fusion unit. It is also used to obtain the physicochemical parameters of the food matrix through the dielectric sensor and water activity meter based on the received physicochemical parameters of the food processing scene, and transmit them to the knowledge transfer submodule; The sensory analysis unit is used to receive the consumer sensory evaluation text data transmitted by the sensing unit, extract semantic feature vectors through the natural language processing model, generate standardized sensory score data, and transmit it to the data fusion unit; The environmental perception unit is used to receive the geographic location, season, temperature and humidity data transmitted by the sensing unit, generate an environmental feature vector through a spatiotemporal coding algorithm, and transmit it to the data fusion unit; The data fusion unit is used to receive time series data, quantitative concentration data, sensory score data and environmental feature vectors, match the time series of different data sources through the timestamp alignment algorithm, use feature embedding technology to map heterogeneous data to a unified dimensional space, perform Z-score normalization on the mapped multidimensional features to eliminate dimensional differences, dynamically allocate the weight of each feature through the attention mechanism, generate a multidimensional joint feature matrix including equipment status, substance concentration, sensory score and environmental factors, and transmit it to the dynamic modeling module.

[0028] The dynamic modeling module is used to receive the multi-dimensional joint feature matrix and perform model training synchronously through a multi-task deep neural network. In the dynamic modeling module, the multi-task deep neural network synchronous execution processing adopts a hard parameter sharing architecture. The shared layer is a 3-layer bidirectional LSTM network, and the private task is a fully connected network. The dynamic weight averaging strategy is adopted when training the model. The task weight update formula is: in, represents the weight of task k in the tth round of training, N represents the total number of tasks, represents the temperature coefficient, represents the relative loss reduction rate of the t-1 round under a certain task, K represents the total number of three subtasks of the nonlinear mapping submodule, knowledge transfer submodule and preference tracking submodule; The dynamic modeling module includes nonlinear mapping submodule, knowledge transfer submodule and preference tracking submodule; The nonlinear mapping submodule is used to fit the threshold mutation curve of the sensory score based on the volatile compound concentration data using a piecewise activation function to generate flavor perception probability distribution data and transmit it to the optimization decision module. The nonlinear mapping submodule includes a threshold modeling unit and a semantic fusion unit, wherein: The threshold modeling unit is used to receive the quantitative concentration data in the multidimensional joint feature matrix. The quantitative concentration data includes the time series concentration values of the target compound during the processing process, and the values are calculated in the preset dynamic threshold interval. A differentiable piecewise activation function is used to construct a concentration-sensory score mapping model, where the differentiable piecewise activation function is defined as: Where c is the concentration of the target volatile compound, is the dynamic sensory mutation threshold, and are all linear gain coefficients, and All are nonlinear adjustment factors; Dynamic adjustment through Bayesian optimization framework , the optimization target is the maximum sensory score mutation point in historical data and The matching degree of , its acquisition function is defined as the improved expected improvement function: in, Indicates that the threshold The newly collected sensory scores are represents the best sensory score in historical data, The concentration-sensory association data including the threshold response curve are generated for the standard deviation of the sensory scores and transmitted to the semantic fusion unit; The semantic fusion unit is used to receive the consumer evaluation text data transmitted by the sensory analysis unit. The text data contains several different flavor description words and their frequencies of occurrence. The pre-trained language model is used to extract the contextual semantic features of the text and generate a word vector matrix with a dimension of [d×1] ; The semantic features are associated with the concentration data output by the threshold modeling unit through the attention mechanism to calculate the semantic-chemical weight matrix: in, represents the association strength between the i-th chemical substance and the j-th semantic label, which is used to quantify the effect of concentration on sensory description; is the eigenvector of chemical concentration data, is the semantic feature vector extracted from the consumer evaluation text, is the word vector corresponding to the semantic label, and d is the feature dimension; Generate a 3D tensor containing association rules between semantic strength and chemical concentration , where N is the type of compound and M is the number of semantic labels. The three-dimensional tensor is transferred to the optimization decision module.

[0029] The knowledge transfer submodule is used to receive a database of material synergies from an external industrial field, eliminate scenario differences through feature distillation and domain adaptation algorithms, generate transfer-enhanced model parameters, and update the transfer-enhanced model parameters to the activation function of the nonlinear mapping submodule in real time, so that the generated flavor perception probability distribution data adapts to the current processing scenario; The knowledge transfer submodule includes a rule distillation unit and a domain alignment unit, where: The rule distillation unit is used to receive the material synergy database of the external industrial field, which contains the synergistic and inhibitory relationship data of cross-domain materials; it is also used to extract high-order interaction patterns between materials through the graph attention network to generate a cross-domain knowledge graph , where the vertex set V represents the material category, the edge set E represents the synergistic relationship, and the weight W represents the intensity of the effect. The knowledge graph is transferred to the domain alignment unit; Domain alignment unit for receiving physicochemical parameters of food processing scenarios, including matrix dielectric constant and water activity , calculate the difference in physical properties between external industry and food processing: Optimize the transfer loss function through the domain adaptation algorithm: in, and are adaptive weight coefficients, Represents the maximum mean difference between the external industry and food processing knowledge graphs, generates the model parameters after migration enhancement, and updates them to the threshold modeling unit.

[0030] The preference tracking submodule is used to construct a spatiotemporal preference map based on the environmental feature vector using a graph convolutional network, dynamically modify the sensory scoring weight coefficient, generate updated scoring benchmark data, and transmit it to the optimization decision module. The preference tracking submodule includes a map construction unit and a weight update unit, where: The graph construction unit is used to receive the environmental feature vector transmitted by the environmental perception unit. The environmental feature vector contains latitude and longitude coordinates and regional cultural feature labels. The modularity-optimized Louvain algorithm is used to divide the flavor preference community and generate the region-season-preference association matrix. , where r represents the number of regions, S is the number of seasons, and K is the preference dimension, and is transmitted to the weight update unit; The weight update unit is used to receive the temperature and humidity time series data in the environmental feature vector, use the LSTM network to predict the preference drift trend, and dynamically correct the sensory score weight coefficient to generate updated score benchmark data and transmit it to the optimization decision module.

[0031] The optimization decision module is used to receive the flavor perception probability distribution data and the updated scoring benchmark data, and solve the Pareto optimal solution set through the multi-objective particle swarm optimization algorithm in combination with the physical constraints of the equipment to generate candidate processing parameter solutions. The equipment control parameters of the candidate processing parameter solutions are inverted through the physical constraint neural network to generate the final processing parameter adjustment instructions and transmit them to the execution module. The optimization decision module includes a constraint processing unit, a particle swarm optimization unit, and a physical inversion unit, wherein: The constraint processing unit is used to receive time series data of the equipment operation status, including the maximum allowable temperature, pressure threshold and energy consumption per unit time threshold, generate a multi-dimensional feasible solution space through the constraint space modeling algorithm, encode the feasible solution space into a boundary constraint matrix, and transmit it to the particle swarm optimization unit; The particle swarm optimization unit receives the flavor perception probability distribution data and updated rating benchmark data transmitted by the dynamic modeling module, and receives the device feasible solution space boundary matrix generated by the constraint processing unit. It performs parallel search in the feasible solution space through the target particle swarm optimization algorithm, combined with the semantic-chemical association tensor Dynamically adjust the weight distribution of flavor intensity and sensory preference to generate a Pareto candidate solution set that maximizes flavor intensity and optimally matches sensory preference. The selected solution set is then transferred to the physical inversion unit for feasibility verification of processing parameters. The physical inversion unit is used to receive the Pareto candidate solution set transmitted by the particle swarm optimization unit. The solution set contains processing parameter combinations that meet the flavor and preference optimization. The feasibility of the parameters is verified through a neural network model embedded with physical constraints. The parameter combinations that have passed the physical feasibility verification are screened, and the processing parameter adjustment instructions are generated. The adjustment instructions are transmitted to the execution module.

[0032] The execution module is used to receive the processing parameter adjustment instruction. The execution module signal is connected to the equipment controller, and the closed-loop control of the processing parameters is realized through the equipment controller. The execution module includes an instruction parsing unit and a closed-loop control unit, wherein: An instruction parsing unit is used to receive the processing parameter adjustment instructions transmitted by the physical inversion unit. The instructions include temperature setting values, time setting values, and pressure setting values. The instruction is verified for physical feasibility through a finite state machine, and a control instruction sequence executable by the equipment is generated and transmitted to the closed-loop control unit. The closed-loop control unit is used to receive the control instruction sequence, adjust the equipment operating parameters in real time through the proportional-integral-differential algorithm, and output the pulse width modulation signal to the processing equipment to realize the closed-loop dynamic control of the processing parameters.

[0033] The specific implementation process is as follows: This embodiment takes Arabica coffee beans as an example. Since its roasting process directly affects its flavor quality, the traditional method relies on manual experience to adjust the temperature and time, which has the following defects: 1. When the hexanal concentration exceeds the threshold, the negative evaluation of "grassy taste" suddenly increases, but the linear model cannot capture the mutation point, resulting in repeated adjustment of process parameters; 2. Manually setting fixed weights (such as a "burnt aroma" scoring weight of 0.6) ignores the dynamic changes in 2-acetylpyrazine concentration, and the optimization target deviates from the actual preference; 3. Generally speaking, for foods such as coffee, consumers prefer a rich flavor in winter and a fresh taste in summer, but the static model update cycle is as long as several months, and it cannot adapt to changes in demand in real time.

[0034] In this solution, the baking drum temperature (160–220°C), time (10–50 minutes), and pressure (80–150 kPa) are collected in real time through the multi-source data acquisition and fusion module. The concentrations of key flavor substances such as hexanal, 2-acetylpyrazine, and furfural are detected using GC-MS. Consumers' subjective ratings of descriptions such as "burnt aroma" and "grassy taste" are collected (1–5 points). A total of 500 samples are collected. The weather API is used to obtain the current season, summer or winter, the geographical location in Northern Europe or Southeast Asia, and the temperature and humidity parameters. The obtained multi-source data are subjected to wavelet transform to remove sensor noise. The substance detection unit uses the internal standard method and retention index correction algorithm to double-correct the GC-MS data errors. The temperature curve, chemical concentration, and semantic score are mapped to a unified time axis to generate a 32-dimensional joint feature matrix.

[0035] When the system detected a hexanal concentration of 1.2 μg / g, the "grassy taste" score dropped sharply. Since the traditional linear model had an error of ±0.9 points near the threshold, the error of this system was reduced to ±0.3 points. By dynamically fitting the concentration-score curve through a piecewise function, Bayesian optimization automatically adjusted the threshold. When it was winter or summer, the threshold was lowered by 0.5 μg / g or increased by 1.5 μg / g, respectively, avoiding the risk of over-roasting the coffee beans during the processing decision-making process.

[0036] The semantic features of "burnt aroma" were extracted using the BERT model and associated with 2-acetylpyrazine concentration through an attention mechanism. When the concentration is >0.8μg / g, the semantic weight is automatically increased to 0.82. Because static models require three months of parameter updates, this system implements real-time adjustments, reducing regional score deviations from ±1.2 points to ±0.4 points. Six global flavor preference regions were identified based on graph construction units. The LSTM network predicts weight changes based on real-time temperature and humidity, increasing the weight of burnt aroma in winter from 0.65 to 0.78.

[0037] After a series of data analyses and predictions, the decision-making module constraint modeling was optimized to define a feasible solution space with a temperature ≤ 210°C and an energy consumption ≤ 800kJ. 35% of infeasible parameter combinations were eliminated. Then, the particle swarm optimization unit generated two candidate parameter sets: 185°C / 35min and 190°C / 30min, taking into account both the aroma score and energy efficiency. The physical inversion unit verified the feasibility of the reaction kinetics using the Arrhenius equation and found that the aroma generation rate exceeded the limit, so the 195°C / 28min solution was eliminated. Through dynamic threshold modeling, semantic-chemical association quantification, and spatiotemporal preference tracking, the system systematically addressed the problems of flavor loss, evaluation distortion, and parameter lag in coffee roasting. Practical applications have shown that the system significantly improves the consistency of sensory scores by 66.7% and energy efficiency by 30.9%, providing a reusable innovative solution for intelligent food processing with broad industrial application value.

[0038] Example 2: As attached Figure 1 and Figure 3 As shown, the difference from Example 1 is that the multi-source data acquisition and fusion module further includes an anomaly detection unit, which is used to detect abnormal data of the sensor unit in real time and uses the isolation forest algorithm to make a real-time judgment: Expressed as, if the isolation path length PathLength of a data point x is less than the threshold , then x is determined to be an outlier, where and They represent the mean and standard deviation of the normal data path length respectively. When the abnormality detection unit detects abnormal data from the sensing unit, the redundant sensing data replacement mechanism is activated.

[0039] The specific implementation process is as follows: During the coffee bean roasting process, sensor data anomalies, such as temperature signal drift or GC-MS peak area distortion, will lead to ineffective modeling and decision-making. Traditional methods have the following problems: Static threshold detection uses fixed thresholds (such as temperature ±5°C) to determine anomalies, which cannot adapt to dynamic data drift caused by equipment aging or sudden environmental changes. Furthermore, when a single sensor detects an anomaly, it directly generates an alarm and shuts down the machine, interrupting the production process and causing financial losses.

[0040] Therefore, this solution receives multi-source sensor data in real time and constructs a historical dataset based on the equipment's normal production cycle (e.g., 100 batches). Statistics such as the temperature curve variance, GC-MS peak area slope, and pressure fluctuation frequency are used as features to construct a 128-dimensional feature vector. Using isolation forest modeling, 100 isolation trees are trained, each randomly sampling 256 data points. The mean, standard deviation, and dynamically set threshold of the path length of normal data are calculated. New data points are detected online, such as the path length of a temperature signal x. If the path length PathLength(x) is less than the dynamically set threshold, it is considered an anomaly and the anomaly type is flagged, such as temperature sensor drift or GC-MS injection failure. The sensing unit deploys three redundant temperature sensors (main sensor A, backup sensors B and C) on the baking drum. When primary sensor A fails, the data from sensors B and C are compared with the process model's predictions. The backup sensor with the smallest deviation is selected, and the data fusion weights are updated: the weight of the normal sensor is increased to 0.8, while the weight of the abnormal sensor is reduced to 0.2.

[0041] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A machine learning-based prediction and optimization decision-making system for flavor formation in food processing, characterized by: It includes sensor units with interconnected signals, multi-source data acquisition and fusion modules, dynamic modeling modules, optimization decision modules and execution modules, among which: The sensor unit is used to acquire and integrate sensor data from multiple different channels and transmit it to the multi-source data acquisition module. The sensor data includes temperature, time, and pressure sensor signals from processing equipment, physical and chemical parameters of food processing scenes, peak area data from chromatography-mass spectrometry, text data of consumer sensory evaluations, and geographic location, season, temperature and humidity data from environmental sensors. The multi-source data acquisition and fusion module is used to receive sensor data and generate corresponding multi-source data. The multi-source data includes time series data of equipment operation status, quantitative concentration data of target volatile compounds, standardized sensory score data, physical and chemical parameters of food matrix, and environmental feature vectors. After feature alignment and normalization of the multi-source data, a multi-dimensional joint feature matrix is obtained and transmitted to the dynamic modeling module. A dynamic modeling module is used to receive a multi-dimensional joint feature matrix and synchronously perform model training through a multi-task deep neural network. The dynamic modeling module includes a nonlinear mapping submodule, a knowledge transfer submodule, and a preference tracking submodule; A nonlinear mapping submodule is used to fit the threshold mutation curve of the sensory score using a piecewise activation function based on the volatile compound concentration data, generate flavor perception probability distribution data, and transmit it to the optimization decision module; The knowledge transfer submodule is used to receive a database of material synergies from an external industrial field, eliminate scenario differences through feature distillation and domain adaptation algorithms, generate transfer-enhanced model parameters, and update the transfer-enhanced model parameters to the activation function of the nonlinear mapping submodule in real time, so that the generated flavor perception probability distribution data adapts to the current processing scenario; The preference tracking submodule is used to construct a spatiotemporal preference map based on the environmental feature vector using a graph convolutional network, dynamically modify the sensory scoring weight coefficient, generate updated scoring benchmark data, and transmit it to the optimization decision module; The optimization decision module is used to receive the flavor perception probability distribution data and the updated scoring benchmark data, and solve the Pareto optimal solution set through the multi-objective particle swarm optimization algorithm in combination with the physical constraints of the equipment to generate candidate processing parameter solutions. The equipment control parameters of the candidate processing parameter solutions are inverted through the physical constraint neural network to generate the final processing parameter adjustment instructions and transmit them to the execution module; The execution module is used to receive processing parameter adjustment instructions. The execution module signal is connected to the equipment controller to achieve closed-loop control of the processing parameters through the equipment controller.

2. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 1, characterized in that: The multi-source data acquisition and fusion module includes a device sensing unit, a material detection unit, a sensory analysis unit, an environmental perception unit, and a data fusion unit, among which: The equipment sensing unit is used to receive the temperature, time and pressure sensing signals of the processing equipment transmitted by the sensing unit, process the useful information frequency higher than the sensing signal itself through the wavelet transform denoising algorithm, generate time series data of the equipment operation status, and transmit it to the data fusion unit; The substance detection unit is used to receive the detection peak area data of the chromatography-mass spectrometer transmitted by the sensing unit, calibrate it using the internal standard method, and then eliminate the chromatographic drift error using the retention index correction algorithm. After performing double calibration, it generates quantitative concentration data of the target volatile compound and transmits it to the data fusion unit. It is also used to obtain the physicochemical parameters of the food matrix through the dielectric sensor and water activity meter based on the received physicochemical parameters of the food processing scene, and transmit them to the knowledge transfer submodule; The sensory analysis unit is used to receive the consumer sensory evaluation text data transmitted by the sensing unit, extract semantic feature vectors through the natural language processing model, generate standardized sensory score data, and transmit it to the data fusion unit; The environmental perception unit is used to receive the geographic location, season, temperature and humidity data transmitted by the sensing unit, generate an environmental feature vector through a spatiotemporal coding algorithm, and transmit it to the data fusion unit; The data fusion unit is used to receive time series data, quantitative concentration data, sensory score data and environmental feature vectors, match the time series of different data sources through the timestamp alignment algorithm, use feature embedding technology to map heterogeneous data to a unified dimensional space, perform Z-score normalization on the mapped multidimensional features to eliminate dimensional differences, dynamically allocate the weight of each feature through the attention mechanism, generate a multidimensional joint feature matrix including equipment status, substance concentration, sensory score and environmental factors, and transmit it to the dynamic modeling module.

3. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 2, characterized in that: The nonlinear mapping submodule includes a threshold modeling unit and a semantic fusion unit, where: The threshold modeling unit is used to receive the quantitative concentration data in the multidimensional joint feature matrix. The quantitative concentration data includes the time series concentration values of the target compound during the processing process, and the values are calculated in the preset dynamic threshold interval. A differentiable piecewise activation function is used to construct a concentration-sensory score mapping model, where the differentiable piecewise activation function is defined as: Where c is the concentration of the target volatile compound, is the dynamic sensory mutation threshold, and are all linear gain coefficients, and All are nonlinear adjustment factors; Dynamic adjustment through Bayesian optimization framework , the optimization target is the maximum sensory score mutation point in historical data and The matching degree of , its acquisition function is defined as the improved expected improvement function: in, Indicates that the threshold The newly collected sensory scores are represents the best sensory score in historical data, The concentration-sensory association data including the threshold response curve are generated as the standard deviation of the sensory scores and transmitted to the semantic fusion unit.

4. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 3, characterized in that: The semantic fusion unit is used to receive the consumer evaluation text data transmitted by the sensory analysis unit. The text data contains several different flavor description words and their frequencies of occurrence. The pre-trained language model is used to extract the contextual semantic features of the text and generate a word vector matrix with a dimension of [d×1] ; The semantic features are associated with the concentration data output by the threshold modeling unit through the attention mechanism to calculate the semantic-chemical weight matrix: in, represents the association strength between the i-th chemical substance and the j-th semantic label, which is used to quantify the effect of concentration on sensory description; is the eigenvector of chemical concentration data, is the semantic feature vector extracted from the consumer evaluation text, is the word vector corresponding to the semantic label, and d is the feature dimension; Generate a 3D tensor containing association rules between semantic strength and chemical concentration , where N is the type of compound and M is the number of semantic labels. The three-dimensional tensor is transferred to the optimization decision module.

5. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 4, characterized in that: The knowledge transfer submodule includes a rule distillation unit and a domain alignment unit, where: The rule distillation unit is used to receive the material synergy database of the external industrial field, which contains the synergistic and inhibitory relationship data of cross-domain materials; it is also used to extract high-order interaction patterns between materials through the graph attention network to generate a cross-domain knowledge graph , where the vertex set V represents the material category, the edge set E represents the synergistic relationship, and the weight W represents the intensity of the effect. The knowledge graph is transferred to the domain alignment unit; Domain alignment unit for receiving physicochemical parameters of food processing scenarios, including matrix dielectric constant and water activity , calculate the difference in physical properties between external industry and food processing: Optimize the transfer loss function through the domain adaptation algorithm: in, and are adaptive weight coefficients, Represents the maximum mean difference between the external industry and food processing knowledge graphs, generates the model parameters after migration enhancement, and updates them to the threshold modeling unit.

6. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 5, characterized in that: The preference tracking submodule includes a graph construction unit and a weight update unit, where: The graph construction unit is used to receive the environmental feature vector transmitted by the environmental perception unit. The environmental feature vector contains latitude and longitude coordinates and regional cultural feature labels. The modularity-optimized Louvain algorithm is used to divide the flavor preference community and generate the region-season-preference association matrix. , where r represents the number of regions, S is the number of seasons, and K is the preference dimension, and is transmitted to the weight update unit; The weight update unit is used to receive the temperature and humidity time series data in the environmental feature vector, use the LSTM network to predict the preference drift trend, and dynamically correct the sensory score weight coefficient to generate updated score benchmark data and transmit it to the optimization decision module.

7. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 6, characterized in that: The optimization decision module includes a constraint processing unit, a particle swarm optimization unit, and a physical inversion unit, among which: The constraint processing unit is used to receive time series data of the equipment operation status, including the maximum allowable temperature, pressure threshold and energy consumption per unit time threshold, generate a multi-dimensional feasible solution space through the constraint space modeling algorithm, encode the feasible solution space into a boundary constraint matrix, and transmit it to the particle swarm optimization unit; The particle swarm optimization unit receives the flavor perception probability distribution data and updated rating benchmark data transmitted by the dynamic modeling module, and receives the device feasible solution space boundary matrix generated by the constraint processing unit. It performs parallel search in the feasible solution space through the target particle swarm optimization algorithm, combined with the semantic-chemical association tensor Dynamically adjust the weight distribution of flavor intensity and sensory preference to generate a Pareto candidate solution set that maximizes flavor intensity and optimally matches sensory preference. The selected solution set is then transferred to the physical inversion unit for feasibility verification of processing parameters. The physical inversion unit is used to receive the Pareto candidate solution set transmitted by the particle swarm optimization unit. The solution set contains processing parameter combinations that meet the flavor and preference optimization. The feasibility of the parameters is verified through a neural network model embedded with physical constraints. The parameter combinations that have passed the physical feasibility verification are screened, and the processing parameter adjustment instructions are generated. The adjustment instructions are transmitted to the execution module.

8. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 7, characterized in that: The execution module includes an instruction parsing unit and a closed-loop control unit, wherein: An instruction parsing unit is used to receive the processing parameter adjustment instructions transmitted by the physical inversion unit. The instructions include temperature setting values, time setting values, and pressure setting values. The instruction is verified for physical feasibility through a finite state machine, and a control instruction sequence executable by the equipment is generated and transmitted to the closed-loop control unit. The closed-loop control unit is used to receive the control instruction sequence, adjust the equipment operating parameters in real time through the proportional-integral-differential algorithm, and output the pulse width modulation signal to the processing equipment to realize the closed-loop dynamic control of the processing parameters.

9. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 8, characterized in that: In the dynamic modeling module, the multi-task deep neural network synchronous execution processing adopts a hard parameter sharing architecture. The shared layer is a 3-layer bidirectional LSTM network, and the private task is a fully connected network. The dynamic weight averaging strategy is adopted when training the model. The task weight update formula is: in, represents the weight of task k in the tth round of training, N represents the total number of tasks, represents the temperature coefficient, It represents the relative loss reduction rate of the t-1 round under a certain task, and K represents the total number of three subtasks of the nonlinear mapping submodule, knowledge transfer submodule and preference tracking submodule.

10. The machine learning-based prediction and optimization decision-making system for flavor formation in food processing according to claim 9, characterized in that: The multi-source data acquisition and fusion module also includes an anomaly detection unit, which is used to detect abnormal data of the sensor unit in real time and uses the isolation forest algorithm to make real-time judgments: Expressed as, if the isolation path length PathLength of a data point x is less than the threshold , then x is determined to be an outlier, where and They represent the mean and standard deviation of the normal data path length respectively. When the abnormality detection unit detects abnormal data from the sensing unit, the redundant sensing data replacement mechanism is activated.

Images