Intelligent flavor regulation method for sauce-marinated meat products based on multi-source information fusion

Abstract

The application discloses a kind of intelligent regulation and control methods of sauce marinated meat product flavor based on multi-source information fusion, belong to sauce marinated meat product processing intelligent control technical field, for the flavor state in existing sauce marinated meat product production depends on artificial experience judgment, multidimensional sensory information is difficult to be quantified fusion, marinating process cannot be dynamically closed-loop regulated, the method is by setting the perception unit of containing olfactory, gustation and visual sensor and collecting multi-source information, after time synchronization, the fusion feature vector of current comprehensive flavor state is extracted by multi-modal information fusion model;Raw material information and fusion feature vector are input into prediction model to obtain the flavor state prediction value of future period, and the deviation amount is generated by comparing with the preset target;Deviation amount is input into control model to solve the control amount acting on marinating equipment to carry out pre-adjustment;After marinating, the product flavor state is obtained and compared with the standard, and the parameters of the prediction model are updated accordingly.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for the processing of braised meat products. More specifically, this invention relates to a method for intelligent flavor control of braised meat products based on multi-source information fusion. Background Technology

[0002] The flavor formation of braised meat products involves complex biochemical processes such as the thermal degradation, oxidation, and Maillard reaction of proteins and fats in the raw meat under the braising broth environment, accompanied by the dissolution and penetration of flavor compounds from spices in the braising broth. In large-scale production, flavor control faces the following interrelated technical challenges.

[0003] First, the methods for perceiving flavor state are outdated, and multi-source information is difficult to integrate effectively. Current production relies heavily on subjective assessments by process personnel, who visually inspect the color of the braising liquid, smell its aroma, and taste its saltiness and umami. This approach suffers from significant individual differences, makes quantifiable evaluations difficult, and complicates consistency during long-term continuous production. While intelligent sensory instruments such as electronic noses, electronic tongues, and machine vision are used in food testing, using a single sensor can only acquire single-dimensional flavor information (odor, taste, or color alone), failing to simultaneously capture comprehensive information from the olfactory, gustatory, and visual dimensions. More importantly, the physical meanings and mathematical expressions of these three types of sensors are completely different, and their acquisition times are asynchronous. Simple feature splicing or linear combination cannot capture the inherent interactions between modalities (such as the correlation between changes in braising liquid color and the generation of aroma substances), making it difficult to integrate multi-source information into a unified indicator that can characterize the overall flavor state.

[0004] Secondly, the strong nonlinearity and large hysteresis of the braising process make it difficult for traditional control methods to achieve dynamic closed-loop regulation. Braising involves heat transfer, mass diffusion, and complex chemical reactions, and the composition of the broth continuously changes during repeated boiling, constituting a non-steady-state system. Traditional mechanistic models based on heat conduction and mass diffusion equations can describe the basic laws under ideal conditions, but they struggle to accurately describe the nonlinear flavor changes caused by numerous uncontrollable factors in the broth system; while purely data-driven models show a significant decrease in generalization ability when raw material batches and broth states change. Even more challenging is the significant hysteresis of the braising process—a delay of several minutes to over ten minutes exists between the application of a control action (such as adjusting heating power or adding ingredients) and a detectable change in flavor state. Traditional feedback control is prone to oscillations and over-adjustment in such hysteresis systems, leading to aggravated flavor fluctuations rather than convergence.

[0005] Finally, fixed-parameter prediction models cannot adapt to the slow evolution of production conditions over the long term. Differences in the origin, season, and feeding methods of raw meat, as well as the slow evolution of the stock itself, cause the prediction bias of fixed-parameter prediction models to gradually accumulate and increase after several months of production. Current technology lacks a mechanism that can automatically trigger online updates of model parameters based on actual test results of the finished product's flavor, making it difficult for prediction models to maintain stable performance throughout the entire production cycle.

[0006] In summary, the current production of braised meat products faces a series of interconnected technical challenges, from flavor perception and fusion to dynamic prediction and closed-loop control, and then to the long-term adaptive modeling. There is an urgent need for an intelligent control method that can systematically solve these problems. Summary of the Invention

[0007] To achieve these objectives and other advantages of the present invention, according to one aspect of the present invention, a method for intelligent flavor control of braised meat products based on multi-source information fusion is provided, comprising the following steps: Obtain raw material information characterizing the initial state of the raw meat, including the initial temperature, geometric dimensions, and initial component concentration of the raw meat; A set of sensors integrated into the braising equipment collects multi-source sensory information reflecting the real-time flavor status of the braising broth and meat during the braising process. The set of sensors includes sensors for acquiring olfactory information, sensors for acquiring taste information, and sensors for acquiring visual information. After synchronizing the olfactory information, the gustatory information, and the visual information in time, a fusion feature vector is extracted through a multimodal information fusion model. The fusion feature vector is used to characterize the overall flavor state at the current moment. The raw material information and the fused feature vector are input into a prediction model, and the prediction model outputs the flavor state prediction value after a prediction window of 5-15 minutes. The flavor state prediction value is then compared with a preset target value to generate a deviation. The deviation is input into a control model, which performs rolling optimization at a preset control cycle to obtain the control quantity acting on the actuator of the braising equipment. The control quantity is then output at the control cycle to adjust the braising process. The actuator includes a heating power regulator and a replenishing valve for adding seasoning liquid to the braising broth. The flavor profile of the finished product after braising is obtained, and the flavor profile of the finished product is compared with the standard flavor profile. Based on the comparison results, the parameters of the prediction model are updated.

[0008] Preferably, the olfactory information is collected by an electronic nose device comprising multiple metal oxide semiconductor gas sensors, and the olfactory information is the rate of change of resistance output by each metal oxide semiconductor gas sensor after contact with the headspace gas of the brine during the sampling period, forming an olfactory response vector. The taste information is collected by an electronic tongue device containing multiple ion-selective electrodes. The taste information is a taste response vector composed of the potential difference between each ion-selective electrode and the reference electrode after the electrode comes into contact with the brine during the sampling period. The visual information is collected by an industrial camera. The visual information includes images of the surface of the braising liquid and cross-sectional images of the meat during the braising process, taken within the sampling period. The cross-sectional images are obtained by non-contact imaging of the cross-section of the meat along the thickness direction using an online visual acquisition device installed in the braising equipment.

[0009] Preferably, the step of extracting a fused feature vector through a multimodal information fusion model includes: The time-synchronized olfactory information, gustatory information, and visual information are respectively input into their respective modality coding networks to extract initial olfactory feature vectors, initial gustatory feature vectors, and initial visual feature vectors; The initial olfactory feature vector, the initial gustatory feature vector, and the initial visual feature vector are input into a multimodal interaction module based on a cross-attention mechanism. The cross-attention mechanism is configured to learn the dependency relationship between any two modalities and output the enhanced olfactory feature vector, enhanced gustatory feature vector, and enhanced visual feature vector after interaction. The olfactory enhancement feature vector, the gustatory enhancement feature vector, and the visual enhancement feature vector are concatenated to generate the fused feature vector.

[0010] Preferably, the cross-attention mechanism is configured to learn the dependency relationship between any two modalities, including: For any selected first modality initial feature vector and second modality initial feature vector, the first modality initial feature vector is mapped to a query vector through a linear transformation, and the second modality initial feature vector is mapped to a key vector and a value vector through a linear transformation, respectively. Calculate the dot product of the query vector and the key vector, and perform softmax normalization on the dot product result to obtain the attention weight matrix. The element values ​​in the attention weight matrix represent the degree of correlation between each feature in the second modality initial feature vector and each feature in the first modality initial feature vector. Multiplying the attention weight matrix with the value vector yields cross-modal features aggregated from the second modality by the first modality, which are used to generate the enhanced feature vector corresponding to the first modality.

[0011] Preferably, the step of inputting the raw material information and the fused feature vector into a prediction model, and having the prediction model output a flavor state prediction value after a prediction window, includes: The prediction model is a mechanism-data hybrid driven model, which includes a parallel mechanism prediction module and a data-driven compensation module. The mechanism prediction module uses the temperature of the brine as a boundary condition and calculates the predicted value of the flavor state mechanism after the prediction window by solving the heat conduction equation and mass diffusion equation inside the meat. The data-driven compensation module takes the fusion feature vector at the current moment, the raw material information, and the historical fusion feature vector as input, and calculates the flavor state compensation residual value after the prediction window through a trained temporal neural network. The flavor state compensation residual value represents the part of flavor state change that the mechanism prediction module failed to explain. The flavor state mechanism prediction value is obtained by adding the flavor state compensation residual value.

[0012] Preferably, the raw material information includes the initial temperature, geometric dimensions, and initial component concentration of the raw meat; The mechanism prediction module uses the initial temperature as the initial temperature field of the heat conduction equation, the initial component concentration as the initial concentration field of the mass diffusion equation, the geometric feature size as the spatial solution domain boundary of the heat conduction equation and the mass diffusion equation, and the brine temperature as the temperature boundary condition of the spatial solution domain to calculate the predicted value of the flavor state mechanism.

[0013] Preferably, the step of inputting the deviation into a control model and solving the control model to obtain the control quantity acting on the actuator of the brining equipment includes: The control model is a model predictive controller, which includes a predictive model and an online rolling optimization solver. The prediction model calculates the predicted flavor state at each discrete time point within the prediction window based on the control amount applied by the actuator at a historical time and the deviation amount at the current time. The online rolling optimization solver receives the deviation amount and uses minimizing the cumulative deviation between the predicted flavor state and the preset reference trajectory within the prediction window as the performance index, while satisfying the physical amplitude constraint and rate constraint of the actuator. By solving the online quadratic programming problem, it generates the control sequence in the future control time domain. The control increment at the current moment is extracted from the control sequence, and the control increment is superimposed with the control quantity at the previous moment, and then used as the control quantity to act on the actuator.

[0014] Preferably, the prediction model calculates the current state estimate of the braising process based on the control quantity applied by the actuator at a historical time and the deviation quantity at the current time, and uses the current state estimate as the initial state, takes a set of candidate control sequences in the future control time domain as input, and recursively calculates the predicted flavor state at each discrete time point within the prediction window through the prediction model.

[0015] Preferably, the step of obtaining the flavor state of the finished product after braising, comparing the flavor state of the finished product with the standard flavor state, and triggering an update of the parameters of the prediction model based on the comparison result includes: The flavor state of the finished product is compared with the preset standard flavor state, and the flavor state deviation is calculated. When the absolute value of the flavor state deviation exceeds the preset deviation threshold range, the raw material information, multi-source sensing information, fusion feature vector and finished product flavor state corresponding to the batch are combined into a training sample and added to the incremental training sample set. The prediction model is trained online using the incremental training sample set, and the internal parameters of the prediction model are updated.

[0016] Preferably, the preset deviation threshold range is a preset upper limit threshold, and the parameter update is triggered when the absolute value of the flavor state deviation is greater than the preset upper limit threshold; the preset upper limit threshold is a value between 0.05 and 0.5.

[0017] A flavor intelligent control system for braised meat products based on multi-source information fusion includes: The sensing unit, comprising an electronic nose device, an electronic tongue device, and an industrial camera, is used to collect multi-source sensing information; The multimodal fusion unit is used to extract fused feature vectors after time synchronization of perceived information; The prediction unit is used to output a flavor state prediction value and generate a deviation value based on raw material information and fused feature vectors; The control unit employs a model predictive controller to solve for the control quantity based on the deviation and output it to the actuator. The feedback update unit is used to obtain the flavor status of the finished product and compare it with the standard, triggering the parameter update of the prediction model.

[0018] The present invention provides at least the following beneficial effects: The intelligent flavor control method for braised meat products based on multi-source information fusion, as described in this invention, transforms the flavor control of braised meat products from relying on manual experience judgment into a quantifiable and reproducible automated process by constructing a closed-loop control system encompassing multi-source perception, multi-modal fusion, model prediction, deviation control, and feedback updates. This solves the problems of strong subjectivity in flavor evaluation and difficulty in ensuring batch-to-batch consistency in traditional production methods. By defining the olfactory response vector as the output resistance change rate of the electronic nose device, the taste response vector as the output potential difference of the electronic tongue device, and the visual information as the images of the braising liquid surface and the cross-sectional images of the meat during the braising process obtained by the industrial camera and online visual acquisition device, the specific sources and physical meanings of the multi-source perception information are clarified. Furthermore, the acquisition of the meat cross-sectional images is limited to online non-contact shooting, eliminating the temporal contradiction with real-time flavor status acquisition. By employing a modal coding network to extract initial feature vectors for each modality, and then learning intermodal dependencies through a multimodal interaction module based on a cross-attention mechanism before concatenating and enhancing features, deep interaction and fusion of heterogeneous information are achieved. The dependencies are specifically defined as the mapping of query vectors, key vectors, and value vectors, along with attention weight calculations, ensuring that cross-modal dependencies have a clear mathematical expression and a repeatable computational process. A hybrid driving model, comprising a mechanism prediction module and a data-driven compensation module, is constructed. The heat conduction equation and mass diffusion equation provide prediction trends based on physical laws, while a temporal neural network compensates for residuals not explained by the mechanism. This overcomes the limitation of insufficient prediction capability of a single model in complex halogenated environments. Furthermore, the raw material information is clarified to include initial temperature, geometric feature dimensions, and initial component concentrations, assigning them physical roles as initial temperature fields, initial concentration fields, and spatial solution domain boundaries, respectively, ensuring the mechanism prediction module has complete input conditions. By concretizing the control model into a model predictive controller, an online rolling optimization solver is used to solve a quadratic programming problem under actuator constraints, extracting control increments and applying them to the actuator. By setting an upper limit threshold for the deviation between the finished product's flavor state and the standard flavor state, when the deviation exceeds the threshold, the current batch data is used to form an incremental training sample set and the prediction model is trained online. This gives the prediction model the ability to self-correct based on actual feedback, alleviating the problem of model performance degradation caused by batch differences in raw materials and the slow evolution of the broth. At the same time, the preset upper limit threshold is limited to a value between 0.05 and 0.5, providing a quantitative basis for the selection of threshold parameters.

[0019] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0020] Figure 1This is a flowchart of the intelligent flavor control method for braised meat products based on multi-source information fusion as described in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to specific embodiments, so that those skilled in the art can implement it based on the description.

[0022] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0023] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0024] like Figure 1 As shown, this invention discloses a method for intelligent flavor control of braised meat products based on multi-source information fusion, comprising the following steps: Obtain raw material information characterizing the initial state of the raw meat, including the initial temperature, geometric dimensions, and initial component concentration of the raw meat; A set of sensors integrated into the braising equipment collects multi-source sensory information reflecting the real-time flavor status of the braising broth and meat during the braising process. The set of sensors includes sensors for acquiring olfactory information, sensors for acquiring taste information, and sensors for acquiring visual information. After synchronizing the olfactory information, the gustatory information, and the visual information in time, a fusion feature vector is extracted through a multimodal information fusion model. The fusion feature vector is used to characterize the overall flavor state at the current moment. The raw material information and the fused feature vector are input into a prediction model, and the prediction model outputs the flavor state prediction value after a prediction window of 5-15 minutes. The flavor state prediction value is then compared with a preset target value to generate a deviation. The deviation is input into a control model, which performs rolling optimization at a preset control cycle to obtain the control quantity acting on the actuator of the braising equipment. The control quantity is then output at the control cycle to adjust the braising process. The actuator includes a heating power regulator and a replenishing valve for adding seasoning liquid to the braising broth. The flavor profile of the finished product after braising is obtained, and the flavor profile of the finished product is compared with the standard flavor profile. Based on the comparison results, the parameters of the prediction model are updated.

[0025] In this technical solution, regarding the technical features of multi-source sensing information acquisition and multimodal fusion, raw material information can be obtained by interfacing with the factory production management system. This raw material information may include parameters such as the variety, part, initial temperature, geometric dimensions, and initial fat content of the raw meat. For acquiring olfactory information, an electronic nose device containing multiple metal oxide semiconductor gas-sensitive elements can be selected. This electronic nose device can extend its probe into the headspace above the braising equipment, contacting the gas components volatilized from the braising liquid during the sampling period, and outputting the rate of change of the ratio of the resistance value of each gas-sensitive element to the baseline resistance value, thus forming an olfactory response vector. For acquiring taste information, an electronic tongue device containing multiple ion-selective electrodes can be selected. This electronic tongue device can be connected to the braising liquid through a bypass sampling pipeline, allowing the braising liquid to flow through the electrode surface during the sampling period, and outputting the potential difference value of each electrode relative to the reference electrode, thus forming a taste response vector. Sensors used to acquire visual information can include industrial cameras and online visual acquisition devices. Industrial cameras can be installed above the braising equipment to capture images of the braising liquid surface. Online visual acquisition devices can be installed on one side of the meat conveying path within the braising equipment to non-contactly capture cross-sectional images of the meat along its thickness direction, obtaining cross-sectional images during the braising process. The aforementioned olfactory response vector, taste response vector, liquid surface image, and cross-sectional image, after being tagged with the same timestamp by a time synchronization module, are then fed into a multimodal information fusion model. This multimodal information fusion model can be deployed in an industrial control computer or edge computing device. Its operation is as follows: first, the olfactory response vector, taste response vector, and visual image are input into their respective modal coding networks to extract initial olfactory feature vectors, initial taste feature vectors, and initial visual feature vectors; then, these initial feature vectors are input into a multimodal interaction module based on a cross-attention mechanism to learn the correlation between any two modal features; finally, the enhanced feature vectors after interaction are concatenated to generate a fusion feature vector representing the overall flavor state at the current moment.

[0026] For the technical feature of flavor state prediction and deviation generation, the prediction model can adopt a mechanism-data hybrid driven structure and be deployed on the same industrial control computer or host server. The mechanism prediction module uses the brine temperature as the boundary condition, the initial temperature in the raw material information as the initial temperature field of the heat conduction equation, the initial component concentration as the initial concentration field of the mass diffusion equation, and the geometric feature size as the solution domain boundary. It solves the heat conduction equation and the mass diffusion equation through the finite difference method or the finite element method to obtain the flavor state mechanism prediction value after the prediction window. The data-driven compensation module can use a trained long short-term memory network or a gated recurrent unit network. It takes the fused feature vector at the current moment, the raw material information, and the historical fused feature vectors of the previous few sampling periods as inputs and outputs the flavor state compensation residual value. After adding the mechanism prediction value and the compensation residual value to obtain the flavor state prediction value, it is compared with the preset target value to generate the deviation. The preset target value can be derived from a standardized formula database, which stores the salinity target value and color target value corresponding to the target product.

[0027] In this invention, the "control period" refers to the time interval between two consecutive outputs of control quantities by the control model, which can be set to 30s to 120s; the "prediction window" refers to the time length for the prediction model to predict the flavor state, which is set to 5min to 15min, typically 10 to 30 times the control period. Both satisfy the following condition: prediction window = prediction time domain length × control period.

[0028] For the technical features of deviation control and model parameter updating, the control model can employ a model predictive controller, deployed within the programmable logic controller (PLC) or dedicated controller of the braising equipment. The model predictive controller internally includes a predictive model and an online rolling optimization solver. The predictive model, as described above, uses the current state estimate as the initial state and recursively calculates the predicted flavor state at each discrete time point within the prediction window using future candidate control sequences as input. The online rolling optimization solver uses minimizing the cumulative deviation between the predicted flavor state within the prediction window and the preset reference trajectory as its performance index. It simultaneously satisfies upper and lower limits of heating power, power change rate constraints, and opening limits of the feeding valve. It generates the control sequence in the future control time domain by solving a quadratic programming problem and extracts the control increment for the current moment from the control sequence, superimposing it with the control quantity from the previous moment before applying it to the actuator. The actuator may include a power regulator for the electric heating element and a solenoid valve on the feeding pipeline. After braising, an electronic nose device, an electronic tongue device, and an industrial camera re-collect the flavor state of the finished product, comparing it with the preset standard flavor state to calculate the flavor state deviation. When the absolute value of the flavor state deviation exceeds the preset upper limit threshold, for example, the threshold can be set to 0.15, the raw material information, multi-source sensing information, fused feature vector and finished product flavor state corresponding to the batch are combined into a training sample and added to the incremental training sample set. The prediction model is then trained online using this sample set to update the internal parameters of the prediction model.

[0029] This implementation method collects flavor perception information from three dimensions—olfactory, gustatory, and visual—online using multiple sensors. After time synchronization and multimodal fusion based on a cross-attention mechanism, a fusion feature vector characterizing the current overall flavor state is extracted. Raw material information and this fusion feature vector are input into a mechanistic data-driven prediction model to obtain a predicted flavor state for future periods. This predicted value is then compared with a preset target to generate a deviation. Based on this, a model predictive controller is used to solve for control quantities that satisfy the actuator constraints through rolling optimization, pre-adjusting the braising process to effectively address its large hysteresis characteristics. After braising, the online incremental update of the model is triggered by comparing the finished product's flavor state with a standard, enabling the prediction model to adapt to changes caused by batch differences in raw materials and the slow evolution of the broth.

[0030] The braising process of braised meat products exhibits typical characteristics of a large time lag system, specifically reflected in: During the braising process, heat is transferred from the heating source (electric heating element or steam jacket) to the braising liquid, and then from the braising liquid to the interior of the meat. This requires overcoming the thermal resistance of the meat itself and the resistance of the heat transfer boundary layer. The geometric dimensions of the meat (typically 20mm to 80mm) and its thermal diffusivity (approximately 1.4 × 10⁻⁶ for beef) are crucial factors. -7 m 2The time constant of heat transfer is determined by the heat transfer coefficients ∂T / ∂t and ∂T / ∂t. According to Fourier's heat conduction equation, ∂T / ∂t = α∇ 2 When the meat block thickness is 45 mm, the characteristic time required for heat to transfer from the surface to the center is approximately (0.045). 2 / α ≈ 14.5 seconds, but this is only an estimate under an ideal one-dimensional heat transfer model. In the actual braising process, due to the dynamic changes in the thermal boundary layer, the convective heat transfer between the meat surface and the braising liquid, and the influence of the heterogeneous structure inside the meat, it takes much longer for the temperature field to establish and reach a steady state. There is a delay of several minutes to more than ten minutes from the application of heating power adjustment to the detection of a detectable change in the temperature at the center of the meat.

[0031] During the braising process, the diffusion of salt, spices, and flavor compounds from the braising liquid into the meat is governed by Fick's second law: ∂C / ∂t=D∇ 2 According to description C, the diffusion coefficient D of salt is approximately 1.2 × 10⁻⁶. -9 m 2 / s. Assuming a meat block thickness of 45mm, the characteristic diffusion time required for salt to penetrate to the center of the meat is approximately (0.045) seconds. 2 / D≈1.7×10 6 The timescale of seconds is much larger than that of the actual braising process. This calculation shows that the braising process (usually 60-120 minutes) cannot allow salt to completely penetrate to the center of the meat. The penetration of flavor substances is gradual and is influenced by multiple factors, including the internal microstructure of the meat, fat content, and moisture distribution, exhibiting a strongly nonlinear spatiotemporal distribution characteristic. From the opening of the replenishment valve (adding seasoning liquid to the braising broth) to a detectable change in the salinity inside the meat, a significant time lag exists due to the need to overcome penetration barriers and diffusion resistance.

[0032] Flavor formation involves complex biochemical processes such as protein thermal degradation, lipid oxidation, and Maillard reactions, which exhibit specific temperature dependence and time-cumulative effects. For example, the Maillard reaction initiates rapidly at high temperatures, but the formation and accumulation of flavor compounds is a cumulative process, with the reaction rate changing exponentially with temperature (following the Arrhenius equation) rather than linearly. Even if the brine temperature increases rapidly, changes in flavor compound concentration require a certain reaction time to manifest. From the application of regulatory actions (adjusting heating power or adding ingredients) to the detection of flavor states (salinity, color, and aroma compound concentrations), the entire response process exhibits significant hysteresis.

[0033] The superposition of the aforementioned multi-source hysteresis effects allows the transfer function of the halogenation process to be modeled using a first-order inertial plus pure delay model. An approximate description, where the pure delay time t d It can last from several minutes to more than ten minutes, where τ is a time constant, and τ is related to t. dThe ratio is significantly smaller than the typical acceptable range in food processing. Studies show that the larger the ratio of pure delay time to the dominant time constant, the more difficult it is for the control system to achieve satisfactory performance. Traditional PID control strategies struggle to achieve good control results for such systems with large time delays.

[0034] This invention uses a model predictive controller (MPC) instead of a traditional PID controller, which has inherent advantages in handling systems with large time delays from the perspective of control theory.

[0035] The fundamental difference between MPC and PID controllers lies in their output. A PID controller's output is a weighted combination of proportional, integral, and derivative errors. Its control decisions are entirely based on current and past deviations, making it a purely reactive control strategy. It doesn't predict future system behavior and doesn't explicitly use a process model. When there's a large time delay in the system, the PID controller accumulates deviations for several minutes to over ten minutes before the control action takes effect, easily leading to over-tuning and oscillations. In contrast, the core control law of MPC originates from a defined internal dynamic process model, which explicitly participates in the generation of control decisions. Specifically, MPC solves an online optimization problem of the following form in each control cycle: ; ; MPC utilizes predictive models to anticipate the system state at multiple future time steps in the control law, enabling it to predict future deviations and take corrective actions in advance. When the system has a large time delay, MPC can adjust in advance before the control action takes effect, thereby offsetting the negative impact of lag.

[0036] Large hysteresis systems require sufficiently long prediction windows. For such systems, the controller's prediction window must at least cover the system's hysteresis time to achieve effective forward-looking regulation. This invention sets the prediction window to 5-15 minutes, a range designed to match the large hysteresis characteristics of the braising process—sufficiently covering the time window from the application of the control action to detectable changes in flavor state. This allows the MPC to plan the control sequence in advance within this window, predicting and applying compensatory actions before deviations actually occur, effectively eliminating the impact of large hysteresis on system stability.

[0037] The actuators (heating power regulator and feed valve) involved in this invention have a clear physical limit constraint (heating power P). min ≤P≤P max Feed valve opening degree O min ≤O≤O max ) and rate constraint (power change rate |ΔP|≤ΔP) maxMPC inherently possesses the ability to explicitly handle input / output constraints. Its online rolling optimization solver incorporates these constraints into the optimization framework when solving quadratic programming problems, ensuring that the generated control sequence always satisfies the physical constraints of the actuator. Traditional PID controllers, on the other hand, lack constraint handling capabilities and rely solely on regulator parameter tuning (such as the PID's proportional gain K). p Integration time T i Differential time T d Indirect control of the output cannot explicitly constrain the magnitude and rate of change of the control quantity. Under conditions of large lag, it is more likely to cause overshoot or even system instability due to actuator saturation.

[0038] There is a coupling between the heating power and the opening of the feeding valve during the braising process—increasing the heating power accelerates water evaporation, changes the braising liquid concentration, and thus affects the feeding strategy; feeding introduces low-temperature seasoning liquid, lowering the braising liquid temperature, which in turn affects the heating power requirement. Model predictive controllers (MPCs) inherently support the control of multiple-input multiple-output (MIMO) systems. Their control law is a decoupling strategy obtained by co-optimizing the coupling relationships between variables within a unified quadratic programming framework, rather than indirectly addressing coupling by adjusting the parameters of a single controller. Research indicates that MPCs are particularly suitable for systems with significant interactions between process variables. The model predictive controller of this invention simultaneously manages both the heating power and the feeding valve actuators, co-optimizing the control sequences of the two channels within the prediction window, thus avoiding conflicts between heating power regulation and feeding valve operation.

[0039] This invention concretizes the theoretical advantages of MPC into an executable control scheme through the following technical means: The effectiveness of MPC is highly dependent on the accuracy of the internal prediction model. This invention employs a mechanism-data hybrid-driven prediction model, using the heat conduction equation and mass diffusion equation as the core mechanisms to provide physically reasonable prediction trends, and using an LSTM temporal neural network to compensate for nonlinear dynamics and unmodeled perturbations. This enables MPC to obtain sufficiently accurate flavor state predictions within a 5-15 minute prediction window, providing a reliable foundation for rolling optimization.

[0040] The MPC controller of this invention employs a performance index aimed at minimizing the cumulative deviation between the predicted flavor state and the preset reference trajectory within the prediction window. , where the reference trajectory y ref The reference trajectory is designed as a first-order exponential decay curve from the current state to the target value, with a time constant τ = 180 s (3 minutes). This reference trajectory design takes into account the dynamic response characteristics of the brine process—the exponential decay trajectory makes the change of control quantity smooth and avoids the violent oscillations common in systems with large time delays.

[0041] Before solving the quadratic programming problem, the MPC controller pre-sets upper and lower limits for heating power (8kW≤P≤15kW), a rate of change constraint for heating power (|ΔP|≤1.5kW), and a constraint for the opening degree of the feeding valve (0%≤0≤100%) to ensure that the generated control quantity is always within the physical capability range of the actuator. Research shows that for controlled systems with large time lags, the MPC can achieve rapid real-time tracking of the setpoint through online cyclic correction. The MPC of this invention, in each control cycle (Δt... c =30s) Based on the current state estimate and the future control sequence, the optimal control sequence in the future control time domain is generated by recursion calculation. The control increment at the current moment is extracted from it and applied to the actuator, thus realizing the rolling time domain optimal control of the large time delay system.

[0042] This invention sets the prediction window to 5-15 minutes and the control period to 30-120 seconds, both satisfying the condition that prediction window = prediction time domain length × control period (the prediction time domain length is taken as 10-30, i.e., 10 × 30 seconds = 5 minutes to 30 × 30 seconds = 15 minutes). The configuration where the control period is an order of magnitude shorter than the prediction window allows MPC to re-optimize the future control sequence based on the latest system state within each control period. This achieves a closed-loop control strategy that proactively plans within the prediction window and makes rolling corrections at each control moment, ensuring both foresight and rapid response capabilities, effectively addressing the control challenges of systems with large time delays.

[0043] In another technical solution, the olfactory information is collected by an electronic nose device containing multiple metal oxide semiconductor gas sensors. The olfactory information is the rate of change of resistance output by each metal oxide semiconductor gas sensor after it comes into contact with the headspace gas of the brine during the sampling period, which constitutes the olfactory response vector. The taste information is collected by an electronic tongue device containing multiple ion-selective electrodes. The taste information is a taste response vector composed of the potential difference between each ion-selective electrode and the reference electrode after the electrode comes into contact with the brine during the sampling period. The visual information is collected by an industrial camera. The visual information includes images of the surface of the braising liquid and cross-sectional images of the meat during the braising process, taken within the sampling period. The cross-sectional images are obtained by non-contact imaging of the cross-section of the meat along the thickness direction using an online visual acquisition device installed in the braising equipment.

[0044] In this technical solution, olfactory information can be collected using an electronic nose device. This device can contain n metal oxide semiconductor gas sensors, where n can be an integer between 6 and 18. The sensor types can include gas-sensitive elements made of materials such as tin oxide, zinc oxide, or tungsten oxide. Each sensor has different sensitivities to aromatic substances such as alcohols, aldehydes, ketones, and esters volatilized from the brine. During the sampling period Δt... s Within, Δt s The time can be set to 30 seconds. The sampling probe of the electronic nose device extends into the headspace above the braising equipment, making full contact with the gas components volatilized from the braising liquid. Each metal oxide semiconductor gas sensor experiences a change in resistance due to the adsorption of gas molecules on its surface, and outputs the rate of change of resistance ΔR. i / R 0i , where R 0i Let S be the baseline resistance value of the i-th sensor, i = 1, 2, ..., n. The resistance change rates of all sensors together constitute the olfactory response vector S. olf =[ΔR1 / R 01 , ΔR2 / R 02 , ..., ΔR n / R 0n Taste information can be collected using an electronic tongue device, which may contain m ion-selective electrodes, where m can be an integer between 5 and 10. The electrode types can include ion-selective electrodes sensitive to sodium, potassium, chloride, and hydrogen ions, as well as one reference electrode. During the sampling period Δt... s Inside, the brine is led to the detection cell of the electronic tongue device through a bypass sampling pipeline. After the brine comes into contact with the surface of each ion-selective electrode, each electrode outputs a potential difference ΔV relative to the reference electrode. j Where j = 1, 2, ..., m, these potential differences together constitute the taste response vector S. tas =[ΔV1,ΔV2,...,ΔV m Visual information can be acquired using industrial cameras and online visual acquisition devices. The industrial camera can be a color area array camera with a resolution of H×W, where H can be 1080 to 2160 pixels and W can be 1920 to 3840 pixels. It should be mounted above the braising equipment, with the lens pointing vertically downwards at the surface of the braising liquid. During the sampling period Δt... s Image of the liquid surface of the braising liquid taken from inside the container. surf I surf It can reflect the color depth, oil distribution, and foam state of the braising liquid; the online visual acquisition device can be installed on one side of the meat conveying path inside the braising equipment, during the sampling period Δt. s Non-contact imaging was used to capture cross-sectional images of the meat along its thickness direction during the braising process. cross Icross It can reflect the color changes inside the meat and the degree of flavor absorption.

[0045] In another technical solution, the step of extracting a fused feature vector through a multimodal information fusion model includes: The time-synchronized olfactory information, gustatory information, and visual information are respectively input into their respective modality coding networks to extract initial olfactory feature vectors, initial gustatory feature vectors, and initial visual feature vectors; The initial olfactory feature vector, the initial gustatory feature vector, and the initial visual feature vector are input into a multimodal interaction module based on a cross-attention mechanism. The cross-attention mechanism is configured to learn the dependency relationship between any two modalities and output the enhanced olfactory feature vector, enhanced gustatory feature vector, and enhanced visual feature vector after interaction. The olfactory enhancement feature vector, the gustatory enhancement feature vector, and the visual enhancement feature vector are concatenated to generate the fused feature vector.

[0046] In this technical solution, the olfactory information S after time synchronization in the multimodal information fusion process... olf Taste information S tas and visual information I surf I cross Each input is then fed into its corresponding modality coding network. The modality coding network for olfactory information can be a one-dimensional convolutional neural network (CNN_1D_olf), which processes the S... olf The resistance change rate sequence of each sensor is used as input, and then passed through L olf Layer convolution and pooling operations extract the initial olfactory feature vector F. olf ∈R dolf , where d olf For olfactory features, the modality encoding network for taste information can also employ a one-dimensional convolutional neural network CNN_1D_tas, which encodes S... tas Using the potential difference sequence of each electrode as input, the initial taste feature vector F is extracted. tas ∈R dtas , where d tas For taste features, the modality encoding network for visual information can be a two-dimensional convolutional neural network (CNN_2D_vis), and the network structure can be ResNet-50 or ResNet-101. surf and I cross The individual features or concatenations are used as input, and the initial visual feature vector F is extracted through multi-layer convolution. vis ∈R dvis , where d vis This represents the visual feature dimension. The initial feature vector F extracted by the three modality coding networks mentioned above...olf F tas F vis Input a multimodal interaction module based on a cross-attention mechanism. This module can be implemented based on the Transformer architecture. By calculating the attention weights between the initial feature vectors of any two modalities, it learns the feature dependencies between olfaction and taste, olfaction and vision, and taste and vision, and outputs the olfactory enhancement feature vector F' after the interaction. olf ∈R dolf Taste enhancement feature vector F' tas ∈R dtas and visually enhanced feature vector F' vis ∈R dvis F' olf 、F' tas and F' vis The concatenation is performed along either the channel dimension or the feature dimension to generate a fused feature vector F. fuse ∈R dfuse , where d fuse =d olf +d tas +d vis The numerical values ​​of each element in the fused feature vector comprehensively reflect the state of the braising liquid and meat in terms of aroma, taste, and color at the current sampling moment.

[0047] In another technical solution, the cross-attention mechanism is configured to learn the dependency relationship between any two modalities, including: For any selected first modality initial feature vector and second modality initial feature vector, the first modality initial feature vector is mapped to a query vector through a linear transformation, and the second modality initial feature vector is mapped to a key vector and a value vector through a linear transformation, respectively. Calculate the dot product of the query vector and the key vector, and perform softmax normalization on the dot product result to obtain the attention weight matrix. The element values ​​in the attention weight matrix represent the degree of correlation between each feature in the second modality initial feature vector and each feature in the first modality initial feature vector. Multiplying the attention weight matrix with the value vector yields cross-modal features aggregated from the second modality by the first modality, which are used to generate the enhanced feature vector corresponding to the first modality.

[0048] In this technical solution, the specific implementation of the cross-attention mechanism is based on the olfactory initial feature vector F. olf As the first mode, the initial feature vector of taste F tas Let's take the second mode as an example. olf Through linear transformation matrix W Q Mapped to query vector Q=WQ ·F olf , will F tas Through two different linear transformation matrices W K and W V Mapped to key vectors K=W respectively K ·F tas Sum vector V=W V ·F tas Calculate the dot product of the query vector and the key vector, and scale the result to obtain the attention score matrix S = (Q·K). T ) / d k 1 / 2 , where d k Let S be the dimension of the key vector. The dot product S reflects the degree of matching between each element in the olfactory feature and each element in the gustatory feature. Softmax normalization is applied to S so that the sum of the values ​​in each row is 1, resulting in the attention weight matrix A = softmax(S), where the element value a in the i-th row and j-th column is... ij a represents the degree of correlation between the j-th feature in the initial taste feature vector and the i-th feature in the initial olfactory feature vector. ij ∈[0,1] and ∑ j a ij =1. Multiplying the attention weight matrix A by the value vector V yields the cross-modal feature F, which aggregates the olfactory modality from the gustatory modality. olf ←tas=A·V, this cross-modal feature F olf ←tas and the initial olfactory feature vector F olf After performing residual connections and layer normalization, it is used to generate the olfactory enhancement feature vector F'. olf =LayerNorm(F olf +F olf ←tas). Similarly, the cross-attention calculation between any other two modalities is performed in the same way.

[0049] It should be noted that this technical solution employs a cross-attention mechanism based on query vectors, key vectors, and value vectors, rather than simple feature concatenation or weighted averaging. This is because there are non-linear, cross-modal synergistic relationships among the olfactory, gustatory, and visual information of braised meat products. For example, a deepening of the braising liquid color is often accompanied by an increase in Maillard reaction products, which are positively correlated with specific aroma compounds (such as pyrazines). Similarly, the perceived intensity of saltiness may be statistically correlated with the saturation of the meat's surface color. Simple feature concatenation or weighted averaging can only independently combine features from each modality, failing to explicitly model which features in one modality are strongly correlated with which features in another modality, thus making it difficult to accurately capture the aforementioned cross-modal synergistic information.

[0050] The introduction of cross-attention mechanism enables the model to automatically learn an "attention weight matrix" from features in one modality to features in another modality. Each element in this matrix represents the degree of relevance of a feature in the second modality to a feature in the first modality, with values ​​ranging from 0 to 1, and the sum of elements in the same row is 1. Specifically, when an element 'a' in the attention weight matrix... ij When the value is close to 1, it indicates that the contribution of the j-th feature of the second mode to the i-th feature of the first mode is dominant, that is, there is a strong cross-modal cooperative relationship between the two; when a ij A value close to 0 indicates a weak or unrelated correlation between the two. Through training, the model can automatically identify high attention weights between, for example, the intensity of the red channel in a brine liquid surface image and the response value of a certain sensor in the electronic nose, thereby enhancing the expression of this cross-modal association in the fused feature vector. Compared to a weighted average with fixed weights, the cross-attention mechanism can dynamically adjust the weight matrix for different input samples, making the fusion process adaptive and generating a more accurate fused feature vector that represents the current overall flavor state.

[0051] In another technical solution, the step of inputting the raw material information and the fused feature vector into a prediction model, and having the prediction model output a flavor state prediction value after a prediction window, includes: The prediction model is a mechanism-data hybrid driven model, which includes a parallel mechanism prediction module and a data-driven compensation module. The mechanism prediction module uses the temperature of the brine as a boundary condition and calculates the predicted value of the flavor state mechanism after the prediction window by solving the heat conduction equation and mass diffusion equation inside the meat. The data-driven compensation module takes the fusion feature vector at the current moment, the raw material information, and the historical fusion feature vector as input, and calculates the flavor state compensation residual value after the prediction window through a trained temporal neural network. The flavor state compensation residual value represents the part of flavor state change that the mechanism prediction module failed to explain. The flavor state mechanism prediction value is obtained by adding the flavor state compensation residual value.

[0052] In this technical solution, the prediction model is composed of a mechanism prediction module and a data-driven compensation module operating in parallel. The mechanism prediction module uses the brine temperature T as the basis for its operation. brine As a boundary condition, this T brine Temperature can be obtained in real time using a type K thermocouple or a Pt100 resistance temperature sensor installed in the brine equipment, with the sampling period corresponding to Δt. sThe same. The mechanism prediction module solves the heat conduction equation and mass diffusion equation inside the meat. The heat conduction equation can be expressed in the form of the Fourier heat conduction differential equation: ∂T / ∂t=α(∂ 2 T / ∂x 2 +∂ 2 T / ∂y 2 +∂ 2 T / ∂z 2 ), where α is the thermal diffusivity, in units of m. 2 / s, the mass diffusion equation can be expressed in the form of Fick's second law: ∂C / ∂t=D(∂ 2 C / ∂x 2 +∂ 2 C / ∂y 2 +∂ 2 C / ∂z 2 ), where D is the diffusion coefficient, with units of m. 2 / s, C is the concentration of the target flavor compound, in mol / m 3 Discrete solutions are obtained in the spatial solution domain using the finite difference method or the finite element method. The spatial discrete mesh size Δx can be set to 1 mm to 5 mm, and the prediction window Δt is calculated. p The temperature distribution inside the meat is then T(x,y,z,t0+Δt). p ) and the concentration distribution of key flavor compounds C(x,y,z,t0+Δt) p This is used as a predictor of flavor state mechanism. mec The data-driven compensation module uses the fused feature vector F at the current moment. fuse (t k ), raw material information, and historical fusion feature vectors from the previous p sampling periods [F fuse (t k-1 ),F fuse (t k-2 ),...,F fuse (t k-p As input, p can be set to 5 to 20. The network can be a trained Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU) network. The number of network layers can be set to 2 or 3, and the number of neurons in each hidden layer can be set to 64 to 256. The output layer outputs the flavor state compensation residual value ŷ. res , the ŷ res This is used to characterize the flavor state changes that the mechanistic prediction module failed to explain due to model simplification and parameter uncertainties. mec With ŷ res Adding the corresponding elements together, we obtain the predicted flavor state value ŷ=ŷ mec +ŷ res This can include salinity predictions after the prediction window. saltColor prediction value color and predicted concentrations of key aroma compounds aroma .

[0053] In another technical solution, the raw material information includes the initial temperature, geometric dimensions, and initial component concentration of the raw meat; The mechanism prediction module uses the initial temperature as the initial temperature field of the heat conduction equation, the initial component concentration as the initial concentration field of the mass diffusion equation, the geometric feature size as the spatial solution domain boundary of the heat conduction equation and the mass diffusion equation, and the brine temperature as the temperature boundary condition of the spatial solution domain to calculate the predicted value of the flavor state mechanism.

[0054] In this technical solution, the raw material information used in the mechanism prediction module can include the initial temperature T0, geometric dimension L, and initial component concentration C0 of the raw meat. T0 can be obtained by measuring the raw meat using an insertion thermometer or infrared thermometer before it enters the braising equipment; the typical range of T0 is 2°C to 8°C. L can include the average thickness or characteristic length of the raw meat piece, which can be measured after the raw meat is cut using laser ranging or structured light scanning; the typical range of L is 20mm to 80mm. C0 can include the initial moisture content, initial fat content, and initial salt content; the typical range of initial moisture content is 70% to 76%, the typical range of initial fat content is 2% to 15%, and the typical range of initial salt content is 0.1% to 0.5%. C0 can be obtained by sampling the raw meat using a near-infrared spectroscopy analyzer or chemical analysis methods. In the mechanism prediction module, T0 is assigned to each discrete node of the initial temperature field of the heat conduction equation, i.e., T(x,y,z,t0)=T0, and C0 is assigned to each discrete node of the initial concentration field of the mass diffusion equation, i.e., C(x,y,z,t0)=C0. L is used as the spatial solution domain boundary x∈[0,L] for the heat conduction equation and the mass diffusion equation, and T0 is used as the spatial solution domain boundary for the mass diffusion equation. brine The temperature boundary condition for the solution domain is T(0,t)=T(L,t)=T brine This establishes a complete boundary value problem, and the solution yields the predicted value of the flavor state mechanism. mec .

[0055] In another technical solution, the step of inputting the deviation into a control model and solving the control model to obtain the control quantity acting on the actuator of the brining equipment includes: The control model is a model predictive controller, which includes a predictive model and an online rolling optimization solver. The prediction model calculates the predicted flavor state at each discrete time point within the prediction window based on the control amount applied by the actuator at a historical time and the deviation amount at the current time. The online rolling optimization solver receives the deviation amount and uses minimizing the cumulative deviation between the predicted flavor state and the preset reference trajectory within the prediction window as the performance index, while satisfying the physical amplitude constraint and rate constraint of the actuator. By solving the online quadratic programming problem, it generates the control sequence in the future control time domain. The control increment at the current moment is extracted from the control sequence, and the control increment is superimposed with the control quantity at the previous moment, and then used as the control quantity to act on the actuator.

[0056] In another technical solution, the prediction model calculates the current state estimate of the braising process based on the control quantity applied by the actuator at a historical time and the deviation quantity at the current time. Using the current state estimate as the initial state, a set of candidate control sequences in the future control time domain are used as input, and the predicted flavor state at each discrete time point within the prediction window is obtained by recursively calculating through the prediction model.

[0057] In this technical solution, the model predictive controller is specifically implemented using a model predictive controller, which can be deployed in the programmable logic controller (PLC) of the brine processing equipment or in an industrial computer. The model predictive controller comprises two core components: a predictive model and an online rolling optimization solver. The predictive model is the one described above, which is used in each control cycle Δt. c Initially, Δt c It can be set to 30s, and the prediction model is based on the control quantity u(t) applied by the actuator at historical time points. k-1 )=[P(t k-1 ), O(t k-1 )]、u(t k-2 )=[P(t k-2 ), O(t k-2 ], ..., where P is the heating power, O is the opening degree of the feeding valve, and the deviation e(t) at the current moment is considered. k The current state of the braising process is estimated to obtain the estimated value of the current state x̂(t). k ). with x̂(t) k As the initial state, the future control time domain N is used. c A set of candidate control sequences As input, the predicted flavor state at each discrete time point within the prediction window is obtained through recursive calculation by the prediction model step by step. , where N p To predict the time domain length, N cTo control the length of the time domain, N p With N c Satisfying N c ≤N p Prediction time domain length N p It can be set to 10 to 30, corresponding to a prediction window of N. p ×Δt c That is, from 5 minutes to 15 minutes, the control time domain length N c It can be set to N p One-third to one-half. The online rolling optimization solver receives the deviation e(t) at the current time. k To minimize the predicted flavor state at each discrete time point within the prediction window from the preset reference trajectory y. ref The cumulative deviation between the two is the performance index, and the performance index J can be defined as follows: Where Q is the state weight matrix, R is the control increment weight matrix, and y ref It can be set to reduce the current flavor profile to a preset target value y. target A first-order smooth curve Where τ is the time constant. Simultaneously, the optimization solver must satisfy the physical limiting constraints and rate constraints of the actuator. The physical limiting constraints may include the upper and lower limits P of the heating power. min ≤P≤P max , where P min It can be set to 0, P max It depends on the rated power of the electric heating element and the upper and lower limits of the feeding valve opening. min ≤O≤O max O min It can be set to 0%, O max It can be set to 100%; the rate constraint can include the maximum rate of change of heating power |ΔP|≤ΔP max , where ΔP max It can be set to P max 10% to 20%. The online rolling optimization solver solves online quadratic programming problems. Generate control sequences in the future control time domain. This control sequence contains future N c The heating power increment and the feed valve opening increment for each control cycle are calculated. The control increment Δu(t) at the current moment is extracted from the control sequence. k ), and compare it with the control quantity u(t) from the previous time step. k-1 After superposition, the heating power value P(t) that should be applied at the current moment is obtained. k )=P(t k-1 )+ΔP(t k) and the opening value of the feed valve O(t) k )=O(tk-1 )+ΔO(t k The signal is transmitted to the actuator via an analog output module or a digital communication interface. The actuator includes a power regulator for the electric heating element and a solenoid valve on the feeding pipeline. The power regulator can adjust the received P(t) signal based on the input signal. k The supply voltage or duty cycle of the electric heating element can be adjusted according to the received O(t) value. k Adjust the feed flow rate using the value.

[0058] In another technical solution, the step of obtaining the flavor state of the finished product after braising and comparing the flavor state of the finished product with the standard flavor state, and triggering an update of the parameters of the prediction model based on the comparison result, includes: The flavor state of the finished product is compared with the preset standard flavor state, and the flavor state deviation is calculated. When the absolute value of the flavor state deviation exceeds the preset deviation threshold range, the raw material information, multi-source sensing information, fusion feature vector and finished product flavor state corresponding to the batch are combined into a training sample and added to the incremental training sample set. The prediction model is trained online using the incremental training sample set, and the internal parameters of the prediction model are updated.

[0059] In another technical solution, the preset deviation threshold range is a preset upper limit threshold. When the absolute value of the flavor state deviation is greater than the preset upper limit threshold, the parameter update is triggered. The preset upper limit threshold is a value between 0.05 and 0.5.

[0060] In this technical solution, regarding the model parameter update mechanism, after the braising process is completed, the electronic nose device, electronic tongue device, and industrial camera re-collect the flavor information of the braising liquid and the finished product. After processing through the same multimodal fusion process as described above, the flavor state y of the finished product is obtained. prod . y prod Compared with the preset standard flavor state y std Compare and calculate the flavor state deviation Δy = |y prod -y std |Δy can be defined as the weighted sum of the absolute differences between each element in the finished flavor state vector and the corresponding element in the standard flavor state vector, i.e. Where K is the dimension of the flavor state vector, w k Let w be the weight coefficient of the k-th dimension, satisfying ∑w k =1. Preset upper limit threshold for deviation δ th It can be set to 0.15, δ thThis is a dimensionless normalized deviation value, a scalar obtained after comprehensively normalizing deviations in various dimensions such as salinity and color. When Δy > δ th At that time, the raw material information corresponding to this batch and the multi-source sensing information S recorded during the braising process will be used. olf (t k ), S tas (t k ), I surf (t k ), I cross (t k The fused feature vector F at each sampling time point fuse (t k and the flavor state of the finished product y prod To form a complete training sample Add to incremental training sample set D incre ={D1, D2, ..., D N The incremental training sample set is stored in a host computer database or cloud storage system. When the number of samples N in the incremental training sample set reaches the preset minimum batch size B, B can be set to 16 or 32. This batch of samples is used to perform online incremental training on the temporal neural network parameters θ in the prediction model. Training can use the mini-batch stochastic gradient descent algorithm, the learning rate η can be set between 0.0001 and 0.001, and the loss function L can be defined as the mean squared error. The parameter update formula is: The weight parameters W and bias parameters b of the network are updated through backpropagation. After online incremental training is completed, the updated model parameters θ new Replace the corresponding original parameter θ in the prediction model old The prediction for subsequent batches will use θ. new The above methods enable the predictive model to gradually adapt to batch-to-batch variations in raw materials and the slow evolution of the stock during long-term production, maintaining its ability to predict flavor profiles.

[0061] The preset deviation upper limit threshold δ th This is a dimensionless, normalized comprehensive deviation value, calculated as follows: the deviation between the finished product's flavor state and the standard flavor state in each dimension is divided by the maximum allowable deviation range for the corresponding dimension, normalized, multiplied by a preset dimension weighting coefficient, and summed to obtain a scalar value between 0 and 1. Specifically, if the finished product's flavor state vector is y... prod = [y1, y2, …, y K The standard flavor state vector is y. std = [y1 std ,y2 std , …, y Kstd The maximum permissible deviation range for each dimension is Δy. max =[Δy1 max ,Δy2 max ,…,Δy K max The dimension weight coefficients are w=[w1,w2,…,w] K ], satisfying ∑w k = 1, then the normalized comprehensive deviation value The preset upper limit threshold δ th The value range is 0.05 to 0.5, corresponding to different control precision requirements: when δ th When the value is close to 0.05, the system has a low tolerance for flavor deviations in the finished product and a high sensitivity to triggering model updates, making it suitable for high-end product lines with stringent requirements for flavor consistency; when δ th A value close to 0.5 indicates a high tolerance and a low frequency of triggering updates, making it suitable for routine production where some batch-to-batch fluctuations are permissible. Those skilled in the art can select an appropriate threshold within this range, such as 0.10, 0.15, 0.20, or 0.30, based on actual product standards and process stability.

[0062] Example 1 A batch of braised beef was selected as the subject of this example. The raw meat was chunks of beef shank. Before entering the braising equipment, the raw material information for this batch was obtained through a terminal connected to the factory's production management system. This information included the type of raw meat being beef shank, the cut being the hind leg shank, the initial temperature T0 being 4°C, the geometric dimension L being 45mm (based on the average thickness of the meat chunks), the initial moisture content being 73%, the initial fat content being 5%, and the initial salt content being 0.2%. The initial temperature was measured at the center of the raw meat using an insertion thermometer, the geometric dimension was obtained by scanning after cutting using a laser rangefinder, and the initial component concentration was measured by a near-infrared spectroscopy analyzer on a sample.

[0063] The braising equipment is a jacketed steam-heated braising pot containing a recirculated stock with an initial salinity target of 2.8% and an initial temperature of 92℃. An electronic nose device comprising eight metal oxide semiconductor gas sensors is installed above the braising pot. These sensors include tin oxide, zinc oxide, and tungsten oxide gas-sensitive elements, each with different sensitivities to volatile aroma compounds in the braising liquid, such as alcohols, aldehydes, ketones, and esters. The sampling probe of the electronic nose device extends into the headspace region 150mm above the surface of the braising liquid. A bypass sampling pipeline is installed on the side wall of the braising pot, connecting to an electronic tongue device containing six ion-selective electrodes: sodium ion selective electrodes, potassium ion selective electrodes, chloride ion selective electrodes, hydrogen ion selective electrodes, and one reference electrode. A 1624×2448 resolution area array color industrial camera is installed directly above the braising pot, with its lens pointing vertically downwards at the surface of the braising liquid. An online vision acquisition device is installed on one side of the meat conveying path inside the braising pot. It has a built-in industrial camera of the same model and is used to take non-contact pictures of the cross-section of the meat along the thickness direction.

[0064] After the braising process begins, the system uses a sampling period Δt s Data from each sensor is collected synchronously every 30 seconds. The braising process lasts for 90 minutes, with a sampling period Δt. s =30s, a total of 180 sampling times were collected, and they are recorded sequentially as t1, t2, …, t 180 The following uses any one of the sampling times t. k Taking (1≤k≤180) as an example, the data acquisition and processing steps for the remaining sampling times are exactly the same. At the k-th sampling time t... k The rate of change of resistance of each sensor in the electronic nose device constitutes the olfactory response vector S. olf (t k The values ​​are: [0.217, 0.163, 0.298, 0.087, 0.435, 0.192, 0.264, 0.105], where each value represents the rate of change of the resistance of each sensor relative to the baseline resistance. The potential difference between the outputs of each electrode of the electronic tongue device and the reference electrode constitutes the taste response vector S. tas (t k = [81.2mV, 44.5mV, -118.3mV, 209.6mV, -64.2mV, 17.8mV]. Image of brine surface obtained by industrial camera. surf (t k The online visual acquisition device captures cross-sectional images of meat products. cross (t k The above data is marked with the same timestamp t by the time synchronization module. k The data set.

[0065] S after time synchronization olf (t k ), S tas (t k ), I surf (t k ) and I cross (t k The olfactory response vector is fed into a multimodal information fusion model deployed on an edge computing device. The input is a 3-layer one-dimensional convolutional neural network CNN_1D_olf, which performs one-dimensional convolution with a kernel size of 3 and max pooling operations to extract the initial olfactory feature vector F. olf ∈R 128 The taste response vector is input into a 3-layer one-dimensional convolutional neural network CNN_1D_tas to extract the initial taste feature vector F. tas ∈R 128 The liquid surface image and cross-sectional image are stitched together and then input into a ResNet-50 2D convolutional neural network to extract the initial visual feature vector F. vis ∈R 256 F olf F tas F vis Input a multimodal interaction module based on the Transformer architecture, and calculate the intermodal dependencies through a cross-attention mechanism. Taking the attention calculation of the olfactory modality on the gustatory modality as an example: F... olf Through linear transformation matrix W Q Mapping to query vector Q, and F tas Through W K and W V Map them to key vector K and value vector V respectively, and calculate the attention score matrix. , where d k =64, so S is normalized using softmax to obtain the attention weight matrix A, where a is an element of A. ij ∈[0,1] represents the correlation between the j-th taste feature and the i-th olfactory feature. Multiplying A and V together yields the cross-modal feature F. olf←tas After residual connection and layer normalization, an olfactory enhancement feature vector F' is generated. olf Similarly, perform the cross-attention calculation between the remaining modal pairs and output F'. olf ∈R 128 、F' tas ∈R 128 、F' vis ∈R 256 The fused feature vector F is obtained by concatenation. fuse (t k )∈R 512 , representing the current time t kThe overall flavor profile.

[0066] Raw material information and F fuse (t k Input the prediction model. The prediction model adopts a mechanism-data hybrid driven structure, and its core computational logic is as follows: t+T =ỹ t+T +r̂ t+T , among which ỹ t+T For the mechanism prediction value, r̂ t+T To compensate residual values ​​using data-driven methods, t+T This is the predicted value for the final flavor profile. The mechanism prediction module uses the braising liquid temperature T, which is measured in real time by a Pt100 resistance temperature sensor installed inside the braising pot. brine Using 92℃ as the boundary condition and the initial temperature of the raw meat T0=4℃ as the heat conduction equation ∂T / ∂t=α∇ 2 The initial temperature field T is assigned values ​​to each discrete node, where α = 1.4 × 10⁻⁶. -7 m 2 / s represents the thermal diffusivity of beef, with the initial component concentration C0 as the mass diffusion equation: ∂C / ∂t=D∇ 2 The initial concentration field of C is assigned values ​​to each discrete node, where the salt diffusion coefficient D = 1.2 × 10⁻⁶. -9 m 2 / s, with the geometric feature size L=45mm as the boundary of the spatial solution domain x∈[0,45mm], the above equation is solved using the finite difference method under the condition of spatial discrete grid size Δx=2mm, and the prediction window Δt is obtained. p =Predicted flavor state mechanism value after 5 minutesỹ t+5min The data-driven compensation module uses a 2-layer LSTM network, with 128 neurons in each hidden layer, based on the current time step F. fuse (t k ), raw material information, and historical fusion feature vectors from the previous 10 sampling periods [F fuse (t k-1 ),..., F fuse (t k-10 As input, the flavor state compensation residual value r̂ is output through the trained network. t+5min . ỹ t+5min With r̂ t+5min Adding the elements together, we obtain the flavor state prediction value ŷ(t). k +5min), this prediction includes the salinity prediction value ŷ salt =2.1% and color prediction value ŷ color (The L value in the L*a*b* color space is 58.3). Compare ŷ with the preset target value y. target(Salinity target value 2.8%, color L value target value 55.0) are compared to generate the deviation e(t) k )=[ŷ salt -2.8%, ŷ color -55.0] = [-0.7%, +3.3].

[0067] The deviation e(t) k The input is a model predictive controller deployed within a programmable logic controller (PLC). The model predictive controller comprises a predictive model and an online rolling optimization solver, where the predictive model is the aforementioned predictive model. In each control cycle Δt... c At the start of 30s, the prediction model is based on the historical value of the heating power P(t) applied by the actuator at historical times. k-1 =8.5kW and historical value of feed valve opening O(t) k-1 =45%, combined with the current deviation e(t) k The current state of the braising process is estimated to obtain the estimated value of the current state x̂(t). k ). with x̂(t) k ) is the initial state, and the future control time domain N is... c The candidate control sequences within the range of 5 are used as input, and the prediction time-domain N is obtained by recursively calculating the prediction model step by step. p =Predicted flavor state at each discrete time point within 10 =. The online rolling optimization solver receives e(t) k To minimize performance metrics For the goal, among which Assuming a first-order exponential decay reference trajectory from the current state to the target value, with a time constant τ = 180s, the state weight matrix Q is a diagonal matrix with diagonal elements [1.0, 0.5], and the control increment weight matrix R is a diagonal matrix with diagonal elements [0.1, 0.05]. Simultaneously, the actuator constraints must be satisfied: heating power constraint 8kW ≤ P ≤ 15kW, power change rate constraint |ΔP| ≤ 1.5kW, and feed valve opening constraint 0% ≤ 0 ≤ 100%. This is achieved by solving a quadratic programming problem. Generate control sequences for the next 5 control cycles [u(t)] k ), u(t k+1 ), ..., u(t k+4 Extract the current control increment Δu(t) from the control sequence. k The control quantity from the previous moment is added together to obtain P(t). k )=9.2kW、O(t kThe voltage is set to 52%, and the signal is sent to the electric heating element power regulator and the feeding pipeline solenoid valve via the analog output module. The power regulator adjusts the duty cycle of the electric heating element's power supply voltage to the corresponding 9.2kW output, and the feeding solenoid valve adjusts its opening to 52%, adding the pre-prepared seasoning liquid to the braising liquid.

[0068] The braising process lasted 90 minutes. Flavor information from the braising liquid and finished beef was then collected again by an electronic nose device, an electronic tongue device, and an industrial camera. This information was processed using the same multimodal fusion process as described above to obtain the flavor profile y of the finished product. prod . y prod Compared with the preset standard flavor state y std Compare and calculate the overall flavor profile deviation. Where K=3 corresponds to the three dimensions of salinity, color L value, and aroma comprehensive index, with weighting coefficients w1=0.4, w2=0.3, and w3=0.3. This batch calculated Δy=0.18. The preset upper limit threshold for deviation δ... th =0.15, since Δy=0.18>δ th The raw material information corresponding to this batch, and the multi-source sensing information at all sampling times [S] olf , S tas , I surf ,I cross ] 1:T , fused feature vectors at each time step [F fuse ] 1:T and the flavor state of the finished product y prod To form a complete training sample D i ={raw material information, [S olf , S tas , I surf , I cross ] 1:180 , [F fuse ] 1:180 , y prod Add the incremental training sample set D stored in the host computer database. incre The system adds this sample to the incremental training sample set D. incre .

[0069] This embodiment employs an online incremental training strategy: when Δy > δ thImmediately upon detection of the current batch sample, perform a one-step parameter update on the LSTM network in the prediction model, with a learning rate η = 0.0005 and a loss function of mean squared error. After the update, replace the original parameters with the new model parameters, and use the updated parameters for flavor state prediction in the next batch of braising. Upon detecting Δy = 0.18 > 0.15, immediately perform a one-step parameter update on the LSTM network in the prediction model using the same batch sample: θ←θ-0.0005·∇ θ L(θ). After the update is complete, the new model parameters θ will be... new Replace the original parameter θ in the prediction model old The updated parameters will be used to predict the flavor profile for the next batch of braising.

[0070] To verify the control effect of this invention, this batch (using the closed-loop control method of this invention) was compared with a previous historical batch (using the same raw materials and process parameters, but without closed-loop control, relying solely on manual experience to adjust heating power and replenishment at regular intervals). The braising process of the previous historical batch also lasted 90 minutes, and its finished product flavor profile, obtained through the same testing procedure, was: salinity 2.3%, color L value 62.5. The overall flavor profile deviation from the standard flavor profile (salinity 2.8%, color L value 55.0) was calculated to obtain Δy. hist =0.41. After closed-loop control, the salinity of this batch was 2.75%, the color L value was 56.2, and Δy = 0.18 was calculated. The comparison shows that the deviation of this batch was reduced by approximately 56% compared to historical batches. Furthermore, during the braising process of this batch, between 30 and 60 minutes, the predicted salinity values ​​output by the prediction model were repeatedly lower than the target value. The MPC controller automatically increased the opening of the feeding valve from the initial 45% to 52% and appropriately increased the heating power, so that the actual salinity gradually approached the target value in the later stages of braising, avoiding the lag and over-adjustment phenomenon common in manual control. The above data indicate that the method of this invention can effectively reduce the deviation between the finished product's flavor state and the standard, significantly improving batch-to-batch consistency.

[0071] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for intelligent flavor control of braised meat products based on multi-source information fusion, characterized in that, Includes the following steps: Obtain raw material information characterizing the initial state of the raw meat, including the initial temperature, geometric dimensions, and initial component concentration of the raw meat; A set of sensors integrated into the braising equipment collects multi-source sensory information reflecting the real-time flavor status of the braising broth and meat during the braising process. The set of sensors includes sensors for acquiring olfactory information, sensors for acquiring taste information, and sensors for acquiring visual information. After synchronizing the olfactory information, the gustatory information, and the visual information in time, a fusion feature vector is extracted through a multimodal information fusion model. The fusion feature vector is used to characterize the overall flavor state at the current moment. The raw material information and the fused feature vector are input into a prediction model, and the prediction model outputs the flavor state prediction value after a prediction window of 5-15 minutes. The flavor state prediction value is then compared with a preset target value to generate a deviation. The deviation is input into a control model, which performs rolling optimization at a preset control cycle to obtain the control quantity acting on the actuator of the braising equipment. The control quantity is then output at the control cycle to adjust the braising process. The actuator includes a heating power regulator and a replenishing valve for adding seasoning liquid to the braising broth. The flavor profile of the finished product after braising is obtained, and the flavor profile of the finished product is compared with the standard flavor profile. Based on the comparison results, the parameters of the prediction model are updated.

2. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 1, characterized in that, The olfactory information is collected by an electronic nose device containing multiple metal oxide semiconductor gas sensors. The olfactory information is the rate of change of resistance output by each metal oxide semiconductor gas sensor after it comes into contact with the headspace gas of the brine during the sampling period, which constitutes the olfactory response vector. The taste information is collected by an electronic tongue device containing multiple ion-selective electrodes. The taste information is a taste response vector composed of the potential difference between each ion-selective electrode and the reference electrode after the electrode comes into contact with the brine during the sampling period. The visual information is collected by an industrial camera. The visual information includes images of the surface of the braising liquid and cross-sectional images of the meat during the braising process, taken within the sampling period. The cross-sectional images are obtained by non-contact imaging of the cross-section of the meat along the thickness direction using an online visual acquisition device installed in the braising equipment.

3. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 1, characterized in that, The step of extracting a fused feature vector through a multimodal information fusion model includes: The time-synchronized olfactory information, gustatory information, and visual information are respectively input into their respective modality coding networks to extract initial olfactory feature vectors, initial gustatory feature vectors, and initial visual feature vectors; The initial olfactory feature vector, the initial gustatory feature vector, and the initial visual feature vector are input into a multimodal interaction module based on a cross-attention mechanism. The cross-attention mechanism is configured to learn the dependency relationship between any two modalities and output the enhanced olfactory feature vector, enhanced gustatory feature vector, and enhanced visual feature vector after interaction. The olfactory enhancement feature vector, the gustatory enhancement feature vector, and the visual enhancement feature vector are concatenated to generate the fused feature vector.

4. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 3, characterized in that, The cross-attention mechanism is configured as follows: mapping the initial feature vector of the first modality to a query vector, and mapping the initial feature vector of the second modality to a key vector and a value vector; calculating the dot product of the query vector and the key vector and normalizing it to obtain the attention weight matrix; multiplying the attention weight matrix with the value vector to obtain cross-modal features, so as to generate an enhanced feature vector.

5. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 1, characterized in that, The step of inputting the raw material information and the fused feature vector into a prediction model, and having the prediction model output a flavor state prediction value after a prediction window, includes: The prediction model is a mechanism-data hybrid driven model, which includes a parallel mechanism prediction module and a data-driven compensation module. The mechanism prediction module uses the temperature of the brine as a boundary condition and calculates the predicted value of the flavor state mechanism after the prediction window by solving the heat conduction equation and mass diffusion equation inside the meat. The data-driven compensation module takes the fusion feature vector at the current moment, the raw material information, and the historical fusion feature vector as input, and calculates the flavor state compensation residual value after the prediction window through a trained temporal neural network. The flavor state compensation residual value represents the part of flavor state change that the mechanism prediction module failed to explain. The flavor state mechanism prediction value is obtained by adding the flavor state compensation residual value.

6. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 5, characterized in that, The initial component concentrations include the initial moisture content, initial fat content, and initial salt content; The mechanism prediction module uses the initial temperature as the initial temperature field of the heat conduction equation, the initial component concentration as the initial concentration field of the mass diffusion equation, the geometric feature size as the spatial solution domain boundary of the heat conduction equation and the mass diffusion equation, and the brine temperature as the temperature boundary condition of the spatial solution domain to calculate the predicted value of the flavor state mechanism.

7. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 1, characterized in that, The step of inputting the deviation into a control model and solving the control model to obtain the control quantity acting on the actuator of the brining equipment includes: The control model is a model predictive controller, which includes a predictive model and an online rolling optimization solver. The prediction model calculates the predicted flavor state at each discrete time point within the prediction window based on the control amount applied by the actuator at a historical time and the deviation amount at the current time. The online rolling optimization solver receives the deviation amount and uses minimizing the cumulative deviation between the predicted flavor state and the preset reference trajectory within the prediction window as the performance index, while satisfying the physical amplitude constraint and rate constraint of the actuator. By solving the online quadratic programming problem, it generates the control sequence in the future control time domain. The control increment at the current moment is extracted from the control sequence, and the control increment is superimposed with the control quantity at the previous moment, and then used as the control quantity to act on the actuator.

8. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 7, characterized in that, The prediction model calculates the current state estimate of the braising process based on the control quantity applied by the actuator at a historical time and the deviation quantity at the current time. Using the current state estimate as the initial state, a set of candidate control sequences in the future control time domain are used as input, and the predicted flavor state at each discrete time point within the prediction window is obtained by recursively calculating through the prediction model.

9. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 1, characterized in that, The process of obtaining the flavor state of the finished product after braising, comparing the flavor state of the finished product with the standard flavor state, and triggering parameter updates to the prediction model based on the comparison results includes: The flavor state of the finished product is compared with the preset standard flavor state, and the flavor state deviation is calculated. When the absolute value of the flavor state deviation exceeds the preset deviation threshold range, the raw material information, multi-source sensing information, fusion feature vector and finished product flavor state corresponding to the batch are combined into a training sample and added to the incremental training sample set. The prediction model is trained online using the incremental training sample set, and the internal parameters of the prediction model are updated.

10. The method for intelligent flavor control of braised meat products based on multi-source information fusion as described in claim 9, characterized in that, The preset deviation threshold range is a preset upper limit threshold. When the absolute value of the flavor state deviation is greater than the preset upper limit threshold, the parameter update is triggered. The preset upper limit threshold is a value between 0.05 and 0.5.

Images