System for detecting flavor of beef tallow hotpot at different stir-frying temperatures

Abstract

The invention discloses a system for detecting flavor of beef tallow hotpot at different stir-frying temperatures, comprising: a sampling module for inserting a multi-probe thermocouple array into a stir-frying material core area to monitor temperature distribution of a material body in real time, and automatically collecting micro-liter gaseous and liquid samples at stir-frying key nodes through an automatic online ultrafast sampler; according to the real-time analysis module, PTR-TOF-MS is adopted for volatile substance analysis to conduct full-spectrum analysis on transmitted gaseous samples, Online LC-MS is adopted for non-volatile substance analysis to conduct near-real-time analysis on liquid samples, the generation rate and cumulant of noxious substances are calculated in real time based on a kinetic equation in noxious substance early warning, and a risk temperature interval is early warned in advance; the intelligent evaluation module is used for predicting different sensory scores according to the real-time chemical data; and constructing a virtual stir-frying process model comprising a heat and mass transfer equation, a flavor generation kinetic equation and a harmful substance accumulation model.

Description

Technical Field

[0001] This invention relates to the field of food testing technology, and in particular to a system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot. Background Technology

[0002] This study investigates the effects of different frying temperatures on the flavor of beef tallow hot pot, aiming to systematically explore the key mechanisms by which temperature parameters influence the formation and evolution of the flavor of beef tallow hot pot base. As a specialty of Sichuan and Chongqing, beef tallow hot pot enjoys widespread popularity globally, and consumers are increasingly focused on flavor quality. While the industry has established traditional frying techniques, these rely heavily on experience and lack scientific quantitative evidence, resulting in insufficient flavor stability across different batches. This research has significant theoretical and practical implications. Theoretically, it can improve our understanding of the formation mechanism of hot pot flavor chemistry and provide data support for establishing flavor fingerprint profiles. In practice, it can provide scientific process parameters for the standardized production of beef tallow hot pot base, guiding companies to optimize and stabilize flavor quality through precise temperature control, meeting consumers' demand for consistent "taste memory." Furthermore, it provides methodological references for the process improvement of other oil-based condiments, promoting the transformation of traditional food processing from experience-driven to data-driven approaches, and ultimately contributing to the cultural inheritance and industrial development of distinctive local cuisines.

[0003] In existing technologies, GC-MS and GC-O miss polar or high-boiling-point substances and only detect the endpoint sample, ignoring the dynamic reaction rate. Therefore, a detection system for the effect of different frying temperatures on the flavor of beef tallow hot pot is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a detection system for the effect of different frying temperatures on the flavor of beef tallow hot pot.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot, comprising: Sampling module: A multi-probe thermocouple array is inserted into the core area of ​​the roasted material to monitor the temperature distribution of the material in real time. At the same time, an automatic online ultra-fast sampler automatically collects micro-level gaseous and liquid samples at key roasting nodes and transmits them instantly to downstream analysis equipment through insulated pipelines, while recording mechanical parameters simultaneously. Real-time analysis module: used to analyze substances generated during the frying process, including volatile substance analysis, non-volatile substance analysis, and hazardous substance early warning. The volatile substance analysis uses PTR-TOF-MS to perform full-spectrum analysis on the transmitted gaseous sample. The non-volatile substance analysis uses Online LC-MS to perform near real-time analysis on the liquid sample. The hazardous substance early warning is based on kinetic equations to calculate the generation rate and accumulation of hazardous substances in real time and to provide early warning of risk temperature ranges. Intelligent evaluation module: A sensory database is constructed by controlling variables through artificial sensory evaluation. Then, the analytical data of PTR-TOF-MS and LC-MS are input into the pre-constructed XGBoost model to obtain artificial sensory scores and establish a mapping relationship between "chemical data and sensory attributes". Examples of artificial sensory scores include "caramelized", "umami" and "rancid". Digital control module: Integrates temperature and mechanical parameters from the sampling module, flavor and harmful substance dynamic data from the real-time analysis module, and predicted sensory scores from the intelligent evaluation module. It constructs a virtual stir-frying process model that includes heat and mass transfer equations, flavor generation kinetic equations, and harmful substance accumulation models. It improves prediction reliability through a hybrid model that integrates mechanism and data. Based on the set optimization objectives, it uses model predictive control algorithms to simulate the results of different temperature adjustment strategies in real time, outputs the best control commands, and directly regulates the stir-frying equipment.

[0006] The above technical solution further includes: Furthermore, the multi-probe thermocouple array uses K-type thermocouples, with 12 probes evenly deployed on the inner wall of the frying pot. Among them, 6 probes are vertically inserted into the core area of ​​the material, 3 probes are close to the bottom of the pot to monitor local hot spots, and 3 probes are placed on the surface of the material to monitor the temperature of the volatilization and oxidation layer. The automatic online ultra-fast sampler uses a microfluidic sampling valve. After sampling, the sample is transported through a double-layer stainless steel pipeline. The inner layer is filled with circulating nitrogen gas at 159℃-161℃, and the outer layer is wrapped with an electric heating belt. The mechanical parameters include stirring rate and power input. The stirring rate is monitored by installing a photoelectric encoder on the stirring paddle shaft to collect the rotation speed in real time. The power input is monitored by connecting a power meter in series in the power cord of the induction cooker to record the real-time power input.

[0007] Furthermore, the multi-probe thermocouple array collects data every 0.1 seconds and sends it to the intelligent evaluation module via a wireless transmission module (ZigBee). The core area temperature threshold is set, and if a probe exceeds the range for 2 seconds, an alarm is automatically triggered and the power of the induction cooker is adjusted. The specific steps of the automatic online ultrafast sampler are as follows: Gaseous sample collection: The sampling needle is inserted into the volatile layer on the surface of the material, and the gas is extracted by a negative pressure pump and injected into the insulated pipeline; Liquid sample collection: The sampling needle is inserted into the core area of ​​the material, and the liquid is pushed to the insulated pipeline by a positive pressure pump; Cleaning and prevention of cross-contamination: After each sampling, rinse the sampling needle and tubing with nitrogen gas at 159℃-161℃; The sampling frequency of stirring rate and power input data is consistent with that of the multi-probe thermocouple array and aligned with the time stamp of temperature data. The temperature distribution changes when the stirring rate changes abruptly, and the influence of mechanical action on the formation of flavor substances are analyzed.

[0008] Furthermore, the specific steps for the analysis of the volatile substances; The gaseous sample enters the PTR-TOF-MS ion source through an insulated pipeline to generate a full spectrum. Molecular ion peaks were matched with the NIST mass spectrometry database and a self-built flavor material spectral library, and the concentrations of each substance were calculated using the quantitative standard curve method. Dynamically generate curves, plot the generation rate curves of pyrazines and aldehydes, and identify peak temperatures; When the rate of decrease in the concentration of key flavor compounds exceeds a threshold, it is marked as "flavor loss risk". Combined with temperature data, it is determined whether the pyrolysis reaction is accelerated due to excessively high temperature.

[0009] Furthermore, the specific steps for the analysis of the non-volatile substances are as follows: Automated sample pretreatment: Liquid samples are mixed with acetonitrile-water solution through an automatic dilution module. The mixed samples are then filtered through a 0.22μm polytetrafluoroethylene filter membrane to remove particulate matter, and EDTA is added to chelate metal ions. LC-MS parameter settings: Liquid chromatography conditions: Column: C18 reversed-phase column; Mobile phase: Phase A and Phase B, gradient elution; Flow rate: 0.3 mL / min; Column temperature: 40℃; Mass spectrometry conditions: Ion source: electrospray ionization; Scan range: 50 m / z-1200 m / z, resolution 120,000; Data Acquisition and Analysis: Targeted quantification: Quantifying the concentration changes of free amino acids, capsaicin, and sugars using multiple reaction monitoring (MRM). Non-targeted screening: High-resolution mass spectrometry data is imported into the mass spectrometry data processing and analysis engine to extract molecular features, compare them with metabolite and compound reference libraries, and discover potential unknown flavor markers; Specialized monitoring of oxidation products: For secondary oxidation products, derivatization method combined with LC-MS detection is used to quantify the indicators, and temperature data is combined to establish a "temperature-oxidation product formation rate" curve.

[0010] Furthermore, the hazardous substance warning includes the following steps: Data input: Real-time temperature data is acquired from a multi-probe thermocouple array and substituted into the kinetic equation to calculate the generation rate and accumulation of harmful substances; Warning: Set threshold: When the predicted cumulative amount reaches 80% of the threshold, trigger "Hazardous Material Risk Warning" and output suggested adjustment strategies; Model optimization and validation: The parameters of the kinetic equations are dynamically updated using real-time collected cooking data, and destructive experiments are conducted periodically to compare the model predictions with the actual detection values.

[0011] Furthermore, the specific steps for the intelligent evaluation module to construct the XGBoost model; Construction of the artificial sensory database: Consumers from major target markets in different regions are recruited using stratified sampling techniques. Combined with random test sequence design and simulation of real dining scenarios, an electronic rating system is used to record sensory ratings on a 0-10 scale in real time. Facial expression recognition is used to assist in verifying the authenticity of the feedback. After removing outliers, a structured database is formed. Data preprocessing and feature engineering: PTR-TOF-MS and LC-MS chemical data are aligned using timestamp synchronization technology, Z-score normalization and Min-Max normalization are used to eliminate dimensional differences, and Pearson correlation coefficient is used to screen chemical features that are strongly correlated with sensory scores. XGBoost model training: Based on the extreme gradient boosting tree algorithm, the model complexity is controlled by setting hyperparameters, and an early stopping mechanism is used to prevent overfitting. The feature contribution is quantified by the SHAP value, and a mapping from chemical data to sensory scores is established. Real-time prediction and deployment: The trained model is exported to PMML format and integrated into the intelligent evaluation module. The output includes JSON results containing scores for "caramelized", "umami" and "rancid" flavors and confidence intervals. At the same time, the model prediction accuracy is maintained and a closed-loop optimization is formed by combining regular human sensory review and online learning mechanisms.

[0012] Furthermore, the specific steps for the digital control module to construct the virtual stir-frying process model are as follows: Constructing a mechanistic model: including heat and mass transfer equations, flavor generation kinetics equations, and a harmful substance accumulation model; Data-driven model: The score output by XGBoost is used as input to correct the prediction bias of the mechanistic model and captures dynamic relationships not covered by the mechanistic model through the long short-term memory network; Weighted fusion: For quantifiable mechanisms, a mechanistic model is used as the main approach, while for complex reactions, a data-driven model correction is used. The weights are dynamically adjusted through Bayesian optimization, and the model prediction error is evaluated through Monte Carlo simulation.

[0013] Furthermore, the specific steps for constructing the mechanism model are as follows: The heat and mass transfer equations are based on the Navier-Stokes equations and the law of conservation of energy, establishing a heat transfer model between the pot body, materials, and the environment: ,in, For the convective heat transfer term caused by stirring, For exothermic reactions, the heat source term is... For the density of the medium, For isobaric specific heat capacity, The rate of change of temperature over time. Thermal conductivity, For temperature gradient; Flavor formation kinetic equation: First-order reaction kinetic model is adopted. ,in, The generation rate constant is temperature-dependent. The concentration of the precursor substance. This refers to the concentration of volatile flavor compounds. This is a temperature-dependent degradation rate constant; Harmful substance accumulation model: Acrylamide formation adopts the Arrhenius model: ,in, The activation energy and pre-exponential factor were fitted through preliminary experiments. This represents the rate of change in amino acid concentration. This refers to the concentration of amino acids. This represents the concentration of carbohydrates.

[0014] Furthermore, the construction steps of the data-driven model are as follows: Flavor intensity prediction: A flavor intensity prediction function is constructed based on physicochemical principles through a mechanistic model to provide basic prediction values. At the same time, XGBoost is used to train and output different sensory scores S with historical process parameters (temperature, time, raw material ratio, etc.) as input features. Bias correction strategy: The bias correction strategy is triggered by setting a threshold. When the sensory score S is less than the set threshold, the bias correction strategy is triggered. At that time, the flavor intensity predicted by the mechanistic model There is an underestimation bias, which is corrected by increasing the weight of the flavor generation term, as shown below. ,in, This is a correction factor (which can be optimized based on historical data). This is the partial derivative of the mechanistic model with respect to temperature (reflecting the sensitivity of temperature to flavor intensity). For temperature changes over time, This is to allow for reaction time, thereby compensating for prediction biases in the mechanistic model; Residual definition: Defines the predicted value of a mechanistic model. Compared with actual flavor intensity The residual ΔF(t) = - The residuals include dynamic relationships not explained by the mechanistic model (such as the nonlinear effect of temperature fluctuations on flavor compounds). LSTM model input and output design: LSTM is used with time series process parameters as input (sequence length is τ, containing information from the past τ time steps), and the output is the residual at the current time step; LSTM network architecture: using the forget gate Controlling the retention of historical information, among which, For the Sigmoid function, This is the weight matrix. This is the hidden state from the previous moment. For the current input, For the bias term of the forget gate, input gate , Controlling the input of new information, among which, The output of the input gate, Here is the weight matrix of the input gate. For the bias term of the input gate, The candidate cell states are generated using the tanh function, representing new information, and their range is between -1 and 1. To generate the weight matrix for candidate cell states, Bias terms for generating candidate cell states. Cell state update. Information integration, among which, This represents the current cell state, and ⊙ represents element-wise multiplication. This represents the cell state at the previous time step. Output gate. , Control status output, where, For the output of the output gate, Here is the weight matrix of the output gate. This is the bias term for the output gate. The hidden state at the current moment is used to capture dynamic relationships not covered by the mechanistic model (such as the nonlinear effect of temperature fluctuations on flavor compounds), and the residual prediction value is output through a fully connected layer. ;

[0015] Model training and loss function: The training objective of LSTM is to minimize the MSE between the residual prediction and the true value. The Adam optimizer is used to update the network parameters and bias terms. Integrated decision: The flavor intensity prediction value is the superposition of the mechanistic model and the LSTM residual.

[0016] Furthermore, the digital control module employs model predictive control algorithms to simulate the results of different temperature adjustment strategies in real time, specifically in the following steps: Optimize target settings: maximize target flavor; minimize harmful substances; Control strategy simulation and optimization: Based on the current state, candidate temperature adjustment strategies are generated. In the future time domain, the trajectory of each strategy is simulated by numerical integration, the objective function value is calculated, and process constraints, safety constraints, and sensory constraints are implemented. Optimal control command output: Select the strategy that satisfies the constraints and minimizes the objective function, convert the control command into an executable format for the device, and send it to the roasting equipment control system via the OPC UA protocol.

[0017] The present invention has the following beneficial effects: 8. In this invention, full-spectrum data is obtained by PTR-TOF-MS (proton transfer reaction time-of-flight mass spectrometry), which can monitor the dynamic generation and decay of polar substances and high-boiling-point compounds in real time, cover the blind spots of the original scheme, track the reaction rate of flavor substances in the heating stage in real time, reveal the influence of temperature on nonlinear reactions, fully cover volatile substances, quantify reaction kinetics, and avoid missing key flavor contributors.

[0018] 9. In this invention, an objective sensory prediction model based on chemical data is established to reduce reliance on unstable artificial senses. In the early stages of the project, artificial sensory evaluation is used. Gradient boosting trees are used with real-time chemical data from PTR-TOF-MS and LC-MS as input and artificial sensory scores as output to train the "digital sensory" prediction model. Subsequently, based on real-time chemical data, the sensory attribute scores of the current product are predicted directly, objectively, and without fatigue, thus realizing online sensory evaluation.

[0019] 10. In this invention, the multi-probe thermocouple array not only monitors the temperature of the pot, but also directly inserts into the material to monitor the core temperature distribution of the material in real time, solving the problem of infrared monitoring being blocked by the material. At multiple key time points in the frying process, it automatically collects trace amounts of gaseous and liquid samples and transmits them instantly to downstream analysis equipment through insulated pipelines, completely avoiding flavor changes during the cooling process. Attached Figure Description

[0020] Figure 1 This is a system block diagram of a system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot, as proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, this invention is a system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot, comprising: Sampling module: The data acquisition front end uses a multi-probe thermocouple array inserted into the core area of ​​the roasted material to monitor the temperature distribution of the material in real time (accuracy ±0.5℃). At the same time, an automatic online ultra-fast sampler automatically collects microliter (μL) gaseous and liquid samples at key roasting nodes (such as every 30 seconds) and transmits them instantly to downstream analysis equipment through an insulated pipeline (maintained at 160℃) to avoid flavor loss during the cooling process and record mechanical parameters simultaneously. Real-time analysis module: The data parsing center, used to analyze substances generated during the stir-frying process. These substances include volatile substance analysis, non-volatile substance analysis, and hazardous substance warning. The volatile substance analysis uses PTR-TOF-MS (proton transfer reaction time-of-flight mass spectrometry) to perform full-spectrum analysis on the transmitted gaseous sample (data acquired once per second). Soft ionization technology reduces the generation of fragment ions and accurately reflects the dynamic generation and decay process of volatile flavor substances such as aldehydes, pyrazines, and sulfides. The non-volatile substance analysis uses Online LC-MS (online liquid chromatography-mass spectrometry) to perform near real-time analysis on liquid samples. The automated sample pretreatment unit supports the monitoring of concentration changes of non-volatile substances such as free amino acids, sugars, capsaicin, and lipid oxidation products, and combines high-resolution mass spectrometry (HRMS) data to conduct non-targeted screening and discover potential unknown flavor markers. The hazardous substance warning is based on kinetic equations (such as the Arrhenius model) to calculate the generation rate and accumulation of hazardous substances such as acrylamide and polycyclic aromatic hydrocarbons in real time, and to provide early warning of risk temperature ranges. Intelligent evaluation module: a data transformation bridge. It constructs a sensory database through human sensory evaluations with controlled variables (regional stratified sampling, random test order, and simulation of real dining scenarios). Then, it inputs the PTR-TOF-MS and LC-MS analysis data into a pre-built XGBoost model to obtain human sensory scores and establish a mapping relationship between "chemical data and sensory attributes" (R²>0.95). Based on real-time chemical data, it directly predicts the sensory scores of "caramelized aroma", "umami", and "rancidity". The digital control module serves as the decision-making and control center. It integrates the temperature and mechanical parameters from the sampling module, the flavor and harmful substance dynamic data from the real-time analysis module, and the predicted sensory scores from the intelligent evaluation module. It constructs a virtual roasting process model that includes heat and mass transfer equations, flavor generation kinetic equations, and a harmful substance accumulation model. It enhances the reliability of predictions through a hybrid model that fuses mechanisms and data. Based on set optimization objectives (such as "maximizing nutty aroma and minimizing acrylamide"), it uses a model predictive control (MPC) algorithm to simulate the results of different temperature adjustment strategies in real time, outputs the optimal control command (such as "current temperature 180℃, it is recommended to cool down to 165℃ within 30 seconds"), and directly controls the roasting equipment.

[0023] In one embodiment, the multi-probe thermocouple array uses K-type thermocouples (diameter ≤1mm to reduce interference with material flow). Twelve probes are evenly deployed on the inner wall of the frying pot. Six probes are vertically inserted into the core area of ​​the material (to a depth of 2 / 3 of the pot's height, avoiding the area where the stirring paddle rotates), three probes are placed close to the bottom of the pot to monitor local hot spots, and three probes are placed on the surface of the material to monitor the temperature of the volatilization and oxidation layers. A standard temperature source (such as a blackbody radiation source) is used to perform three-point calibration (100℃, 150℃, 200℃) on all probes to ensure an accuracy of ±0.5℃. An infrared thermal imager is used to assist in verifying the probe positions to ensure that the probes in the core area are not blocked by materials such as chili peppers and spices, thereby achieving real-time monitoring of the material temperature distribution. The automated online ultra-fast sampler uses a microfluidic sampling valve (such as Valco Instruments MP2™) to support simultaneous sampling of gaseous and liquid samples. The minimum sampling volume is 5 μL. The sampling needle is made of high-temperature resistant stainless steel (withstanding 250℃) with an outer diameter of 0.5 mm to reduce mechanical disturbance to the material. After sampling, the sample is transferred through a double-layer stainless steel pipeline. The inner layer is filled with circulating nitrogen gas at 159℃-161℃ (an inert gas to prevent oxidation), and the outer layer is wrapped with an electric heating belt (temperature sensor provides real-time feedback, maintaining ±1℃ fluctuation). The total length of the pipeline is ≤2m to reduce transmission delay (target delay <5 seconds). The mechanical parameters include stirring rate and power input. The stirring rate is monitored by installing a photoelectric encoder on the stirring paddle shaft to collect the rotation speed in real time. The power input is monitored by connecting a power meter (such as Fluke 438-II) in series in the power cord of the induction cooker, with a sampling frequency of 1kHz, to record the real-time power input.

[0024] In this embodiment, the key nodes refer to the critical time points or stages during the stir-frying process where flavor substances are released, chemical reaction rates change abruptly, or process parameters change significantly, including: Initial stage (0-5 minutes): This is the stage of butter melting and initial heating, with a temperature of approximately 120-130℃. At this time, the butter changes from a solid to a liquid state, and moisture begins to evaporate. It is necessary to monitor the temperature uniformity to avoid localized overheating and burning.

[0025] Mid-stage (5-15 minutes): Peak period of flavor compound release and Maillard reaction, temperature approximately 140-150℃. A large amount of volatile components in spices (such as chili peppers and Sichuan peppercorns) are released. Sampling needs to be triggered when the temperature reaches 145℃ (e.g., the peak solubility point of capsaicin) to capture the instantaneous changes in flavor compounds.

[0026] Late stage (15-20 minutes): Caramelization and flavor stabilization stage, temperature approximately 160℃. At this time, non-volatile components in the flavoring (such as caramel coloring) begin to caramelize. Sampling should be performed when the temperature reaches 160℃ to avoid loss of flavor substances due to cooling.

[0027] Special nodes: such as abrupt changes in stirring rate (e.g., from low-speed mixing to high-speed dispersion) and power input adjustment points (e.g., from constant power to variable frequency control), require synchronous recording of changes in mechanical parameters to analyze their impact on flavor.

[0028] The sampling frequency needs to be matched with the key nodes, such as sampling once every 30 seconds in the mid-term stage, to ensure that instantaneous changes in flavor substances are captured.

[0029] The core area of ​​the material refers to the central region where the temperature distribution is most stable and best represents the overall state of the material during the roasting process, including: Spatial positioning: Located in the center of the frying pan, a cylindrical area with a diameter of about 5-10cm, avoiding temperature gradients caused by the edge contacting the pan wall; Temperature characteristics: Minimal temperature fluctuation range (within ±0.5℃), reflecting the overall average temperature state of the material, rather than local hot or cold spots; Physical state: Located in the center of the material pile, ensuring that the thermocouple probe can directly contact the material body, rather than the air or the pan wall.

[0030] In one embodiment, the multi-probe thermocouple array collects data every 0.1 seconds and sends it to the intelligent evaluation module via a wireless transmission module (ZigBee). The core area temperature threshold is set (e.g., 160℃±5℃). If a probe exceeds the range for 2 seconds, an alarm is automatically triggered and the power of the induction cooker is adjusted. The specific steps of the automatic online ultrafast sampler are as follows: Gaseous sample collection: The sampling needle is inserted into the volatile layer on the surface of the material, and 5μL of gas is extracted by a negative pressure pump (flow rate 0.1L / min) and injected into the insulated pipeline instantly; Liquid sample collection: The sampling needle is inserted into the core area of ​​the material, and 5μL of liquid is pushed into the insulated pipeline by a positive pressure pump (pressure 0.1MPa); Cleaning and prevention of cross-contamination: After each sampling, rinse the sampling needle and tubing with nitrogen gas at 159℃-161℃ (flow rate 0.5L / min, duration 5 seconds). The sampling frequency of stirring rate and power input data is consistent with the sampling frequency of the multi-probe thermocouple array and aligned with the temperature data timestamp. The temperature distribution change is recorded when the stirring rate changes abruptly (e.g., from 200 rpm to 400 rpm) to analyze the effect of mechanical action on the formation of flavor substances.

[0031] Edge computing nodes are deployed next to the cooking equipment to perform preliminary processing on the raw data.

[0032] The pre-processed data is uploaded to the cloud database in real time via a 5G network.

[0033] Each sample is automatically assigned a unique ID, which is associated with parameters such as cooking time, temperature, and stirring rate.

[0034] Gaseous samples are directly injected into PTR-TOF-MS (proton transfer reaction time-of-flight mass spectrometry) via an insulated pipeline for analysis; liquid samples are injected into online LC-MS (liquid chromatography-mass spectrometry) and the concentration is adjusted to the detection range by an automatic dilution module (1:10 ratio).

[0035] In one embodiment, the specific steps of the volatile substance analysis are as follows: The gaseous sample enters the PTR-TOF-MS ion source through an insulated pipeline (maintained at 160℃ with nitrogen), generating a full spectrum every second (containing molecular ion peaks of VOCs such as aldehydes, pyrazines, and sulfides). Molecular ion peaks were matched between the NIST mass spectrometry database and a self-built flavor material spectral library (containing 300+ hot pot flavor compounds), and the concentrations of each substance (in μg / kg) were calculated using the quantitative standard curve method (external standard method). Dynamically generate curves to plot the generation rate curves of pyrazines (such as 2,5-dimethylpyrazine) and aldehydes (such as hexanal), and identify peak temperatures (such as the highest concentration of pyrazines at 160℃). When the rate of decrease in concentration of key flavor compounds (such as 2-acetylpyrrolidine) exceeds a threshold (such as 10% / min), it is marked as "flavor loss risk". Combined with temperature data, it is determined whether the pyrolysis reaction is accelerated due to excessively high temperature (>180℃).

[0036] In one embodiment, the specific steps for analyzing the non-volatile substances are as follows: Automated sample pretreatment: Liquid sample (5 μL) is mixed with acetonitrile-water solution (containing 0.1% formic acid) through an automatic dilution module (1:10 ratio) to reduce viscosity for LC-MS analysis. The mixed sample is then filtered through a 0.22 μm polytetrafluoroethylene filter membrane to remove particulate matter. EDTA (0.1 mM) is added to chelate metal ions and reduce the inhibition of MS signal. LC-MS parameter settings: Liquid chromatography conditions: Column: C18 reversed-phase column (2.1×100mm, 1.8μm); Mobile phase: Phase A (0.1% formic acid aqueous solution), Phase B (acetonitrile), gradient elution (0-5min: 5% B→95% B); Flow rate: 0.3 mL / min, Column temperature: 40℃; Mass spectrometry conditions: Ion source: electrospray ionization (ESI), positive and negative ion mode switching (switching every 30 seconds); Scan range: 50 m / z-1200 m / z, resolution 120,000 (FWHM). Data Acquisition and Analysis: Targeted quantification: Quantifying the concentration changes (ppm level) of free amino acids (such as glutamic acid), capsaicin, and sugars (such as glucose) using multiple reaction monitoring (MRM) mode. Non-targeted screening: High-resolution mass spectrometry (HRMS) data is imported into a mass spectrometry data processing and analysis engine (such as Compound Discoverer software, which automatically processes the raw data files generated by HRMS (such as .raw format), extracts the retention time, precise mass number (error <5 ppm), isotope distribution, fragment ion information and other features of each compound, corrects the retention time drift between different batches of samples, merges overlapping peaks, ensures the signal consistency of the same compound in different samples, and uses statistical methods such as analysis of variance (ANOVA) and cluster analysis (PCA) to screen out compounds with significant concentration changes during the stir-frying process), performs molecular feature extraction (such as retention time, precise mass number, isotope distribution), compares with metabolite and compound reference libraries (such as the Metlin database), and explores potential unknown flavor markers (such as Maillard reaction intermediates). Example: Data acquisition: HRMS generates mass spectrometry data of the stir-fried sample (e.g., a peak at m / z 123.056 is detected at a certain temperature point); Feature extraction: Compound Discoverer was used to analyze the data and record the peak's retention time (3.2 min), exact mass number (123.056), and isotopic distribution (consistent with...). (isotope patterns) Database comparison: A search of the Metallin database for m / z 123.056 yielded a match for the candidate compound "2-acetylfuran" (molecular formula). Flavor description: "caramel aroma" Verification and correlation: Combining the cooking temperature (160°C) with sensory evaluation data ("caramel aroma" score increased), the compound was confirmed as a key flavor marker.

[0037] Specialized monitoring of oxidation products: For secondary oxidation products such as TBARS (malondialdehyde), a derivatization method (TBA reagent) combined with LC-MS detection is used to quantify the indicators, and a "temperature-oxidation product formation rate" curve is established by combining temperature data.

[0038] In one embodiment, the hazardous substance warning includes the following steps: Data input: Real-time temperature data (T(t)) is obtained from a multi-probe thermocouple array and substituted into the kinetic equation to calculate the rate of formation of harmful substances (k(T)) and the amount of accumulation (C(t)). Warning: Set thresholds: Acrylamide > 5 μg / kg (EU limit), Benzo[a]pyrene > 1 μg / kg (WHO limit). When the predicted cumulative amount reaches 80% of the threshold, trigger "Hazardous Substance Risk Warning" and output suggested adjustment strategies (such as "Current temperature 185℃, it is recommended to cool down to 165℃ within 30 seconds"). Model optimization and validation: By collecting real-time frying data (temperature, time, concentration of harmful substances), the parameters of the kinetic equation (such as Ea, A) are dynamically updated to improve prediction accuracy (target error <10%). Destructive experiments (such as forced heating to 220℃) are conducted periodically to compare the model predictions with the actual detection values ​​and ensure model reliability.

[0039] The kinetic equations include the acrylamide formation model and the polycyclic aromatic hydrocarbon formation model; Acrylamide formation model: based on the Arrhenius equation ( The acrylamide formation rate constant (k) at different temperatures (140℃, 160℃, 180℃) was determined through preliminary experiments, and the activation energy (Ea) and pre-exponential factor (A) were obtained by fitting. Polycyclic aromatic hydrocarbon formation model: using first-order reaction kinetics ( A concentration accumulation model was established by combining high-temperature coking experimental data (such as the formation rate of benzo[a]pyrene at 200℃).

[0040] In one embodiment, the specific steps of the intelligent evaluation module in constructing the XGBoost model are as follows: Construction of the artificial sensory database: Stratified sampling technique was used to recruit consumers from major target markets covering different regions such as Sichuan-Chongqing, Northern China, and Southern China (≥30 people in each region, with screening criteria including no taste or smell disorders and no recent history of spicy food consumption). This was combined with a randomized test sequence design (randomly arranging samples at cooking temperatures to avoid memory bias due to sample order, and setting a water rinsing step between each sample group) and a simulation of real dining scenarios (pairing with typical ingredients such as tripe and duck intestines, and controlling the temperature and humidity of the dining environment and the consistency of tableware). An electronic scoring system (such as Compusense®) was used to record sensory scores on a 0-10 scale in real time, and facial expression recognition (Affectiva) was used to assist in verifying the authenticity of the feedback. After removing outliers, a structured database was formed. Data preprocessing and feature engineering: PTR-TOF-MS (once per second) and LC-MS (once every 2 minutes) chemical data were aligned using timestamp synchronization technology. Z-score normalization and Min-Max normalization were used to eliminate dimensional differences. Pearson correlation coefficient was used for screening. Chemical features strongly correlated with sensory scores (e.g., 2,5-dimethylpyrazine and "caramel" score) are identified, and the model's generalization ability is enhanced by constructing interaction terms (e.g., "pyrazine concentration × temperature" to reflect the effect of temperature on flavor compounds) and nonlinear transformations (taking the logarithm of nonlinear correlated features such as TBARS value (log(TBARS+1))). XGBoost model training: Based on the extreme gradient boosting tree algorithm, the model complexity is controlled by setting hyperparameters such as learning rate (0.1), tree depth (6), and subsampling rate (0.8). The early stopping mechanism (patience=5) is combined to prevent overfitting. 5-fold cross-validation is used to ensure stability. The contribution of features is quantified by SHAP value (such as the high influence of pyrazines on "caramel aroma"). The mapping from chemical data to sensory scores is established (R²>0.95, RMSE<0.5). XGBoost Model Training Process and Validation: Data partitioning: The training set, validation set, and test set are divided in a 7:2:1 ratio (stratified sampling by region ensures consistent distribution of each set).

[0041] Training process: The tree structure is iteratively optimized using the training set data. Overfitting is monitored through the validation set (training stops when the validation set error does not decrease for 3 consecutive rounds). An early stopping mechanism (patience=5) is introduced to avoid overtraining.

[0042] Model validation: Test set performance: R²>0.95, RMSE<0.5 (sensory rating error range), MAE<0.3.

[0043] Cross-validation: Five-fold cross-validation (K=5) is used to ensure the stability of the model on different subsets of data (standard deviation <0.05).

[0044] Real-time prediction and deployment: The trained model is exported to PMML format and integrated into the intelligent evaluation module. Low-latency (<100ms) data transmission is achieved through Kafka message queue. The output is a JSON result containing scores for "caramelized", "umami", and "bad taste" and confidence intervals. At the same time, regular human sensory review (3 batches are sampled every week) and online learning mechanism (100 new data are updated incrementally every month) are combined to maintain the model prediction accuracy and form a closed-loop optimization.

[0045] In one embodiment, the specific steps of the digital control module in constructing the virtual stir-frying process model are as follows: Constructing mechanistic models (physicochemical equations): including heat and mass transfer equations, flavor generation kinetics equations, and harmful substance accumulation models; The heat and mass transfer equations are based on the Navier-Stokes equations and the law of conservation of energy, establishing a heat transfer model between the pot body, materials, and the environment: ,in, For the convective heat transfer term caused by stirring, This serves as the heat source term for exothermic reactions such as the Maillard reaction. For the density of the medium, For isobaric specific heat capacity, The rate of change of temperature over time. Thermal conductivity, For temperature gradient; Flavor formation kinetic equation: First-order reaction kinetic model is adopted. ,in, The generation rate constant is temperature-dependent. The concentration of the precursor substance. This refers to the concentration of volatile flavor compounds. This is a temperature-dependent degradation rate constant; Harmful substance accumulation model: Acrylamide formation adopts the Arrhenius model: ,in, The activation energy and pre-exponential factor were fitted through preliminary experiments. This represents the rate of change in amino acid concentration. This refers to the concentration of amino acids. Concentration of carbohydrates; Data-driven model: The XGBoost output score is used as input to correct the prediction bias of the mechanistic model (e.g., when the sensory score is below the threshold, the weight of the flavor generation item is increased), and the dynamic relationship not covered by the mechanistic model is captured by the Long Short-Term Memory Network (LSTM) (e.g., the nonlinear effect of temperature fluctuation on flavor substances). Flavor intensity prediction: A flavor intensity prediction function is constructed based on physicochemical principles through a mechanistic model to provide basic prediction values. At the same time, XGBoost is used to train and output different sensory scores S with historical process parameters (temperature, time, raw material ratio, etc.) as input features. Bias correction strategy: The bias correction strategy is triggered by setting a threshold. When the sensory score S is less than the set threshold, the bias correction strategy is triggered. At that time, the flavor intensity predicted by the mechanistic model There is an underestimation bias, which is corrected by increasing the weight of the flavor generation term, as shown below. ,in, This is a correction factor (which can be optimized based on historical data). This is the partial derivative of the mechanistic model with respect to temperature (reflecting the sensitivity of temperature to flavor intensity). For temperature changes over time, This is to allow for reaction time, thereby compensating for prediction biases in the mechanistic model; Residual definition: Defines the predicted value of a mechanistic model. Compared with actual flavor intensity The residual ΔF(t) = - The residuals include dynamic relationships not explained by the mechanistic model (such as the nonlinear effect of temperature fluctuations on flavor compounds). LSTM model input and output design: LSTM is used with time series process parameters as input (sequence length is τ, containing information from the past τ time steps), and the output is the residual at the current time step; LSTM network architecture: using the forget gate Controlling the retention of historical information, among which, For the Sigmoid function, This is the weight matrix. This is the hidden state from the previous moment. For the current input, For the bias term of the forget gate, input gate , Controlling the input of new information, among which, The output of the input gate, Here is the weight matrix of the input gate. For the bias term of the input gate, The candidate cell states are generated using the tanh function, representing new information, and their range is between -1 and 1. To generate the weight matrix for candidate cell states, Bias terms for generating candidate cell states. Cell state update. Information integration, among which, This represents the current cell state, and ⊙ represents element-wise multiplication. This represents the cell state at the previous time step. Output gate. , Control status output, where, For the output of the output gate, Here is the weight matrix of the output gate. This is the bias term for the output gate. The hidden state at the current moment is used to capture dynamic relationships not covered by the mechanistic model (such as the nonlinear effect of temperature fluctuations on flavor compounds), and the residual prediction value is output through a fully connected layer. ;

[0046] Model training and loss function: The training objective of LSTM is to minimize the MSE between the residual prediction and the true value. The Adam optimizer is used to update the network parameters and bias terms. Integrated decision: Flavor intensity prediction is the superposition of the mechanistic model and the LSTM residual. Weighted fusion: For quantifiable mechanisms (such as heat transfer), a mechanistic model is used as the main approach, while for complex reactions (such as flavor generation), a data-driven model is used for correction. The weights are dynamically adjusted through Bayesian optimization, and the model prediction error (such as flavor concentration ±5%) is evaluated through Monte Carlo simulation, providing a confidence interval for optimized control.

[0047] In one embodiment, the digital control module employs a model predictive control algorithm to simulate the results of different temperature adjustment strategies in real time, specifically in the following steps: Optimize target settings: maximize target flavor; minimize harmful substances; Control strategy simulation and optimization: Based on the current state (temperature, time, chemical concentration), generate 100+ candidate temperature adjustment strategies (e.g., "maintain 180℃", "cool down to 165℃", "heat up to 190℃ and then rapidly cool"). In the future time domain, simulate the trajectory of each strategy through numerical integration (e.g., Runge-Kutta method), calculate the objective function value, and implement process constraints (e.g., temperature range 140-200℃, stirring rate ≥100rpm), safety constraints (e.g., acrylamide <5μg / kg), and sensory constraints (e.g., "bad smell" score <2 points). Optimal control command output: Select the strategy that satisfies the constraints and minimizes the objective function (e.g., "Current temperature is 180℃, it is recommended to cool down to 165℃ within 30 seconds and maintain it until the frying is finished"), convert the control command into an executable format for the equipment (e.g., induction cooker power adjustment command, stirring paddle speed command), and send it to the frying equipment control system via the OPC UA protocol.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot, characterized in that, include: Sampling module: A multi-probe thermocouple array is inserted into the core area of ​​the roasted material to monitor the temperature distribution of the material. At the same time, an automatic online ultra-fast sampler automatically collects micro-level gaseous and liquid samples at key roasting nodes and transmits them instantly to downstream analysis equipment through insulated pipelines, while simultaneously recording mechanical parameters. Analysis module: used to analyze substances generated during the frying process, including volatile substance analysis, non-volatile substance analysis, and hazardous substance early warning. The volatile substance analysis uses PTR-TOF-MS to perform full-spectrum analysis on the transmitted gaseous sample, the non-volatile substance analysis uses Online LC-MS to analyze the liquid sample, and the hazardous substance early warning is based on kinetic equations to calculate the generation rate and accumulation of hazardous substances and provide early warning of risk temperature ranges. Intelligent evaluation module: A sensory database is constructed by controlling variables through artificial sensory evaluation. Then, the analytical data of PTR-TOF-MS and LC-MS are input into the pre-constructed XGBoost model to obtain artificial sensory scores and establish a mapping relationship between "chemical data and sensory attributes". Digital control module: Integrates temperature and mechanical parameters from the sampling module, flavor and harmful substance dynamic data from the analysis module, and predicted sensory scores from the intelligent evaluation module. It constructs a virtual stir-frying process model that includes heat and mass transfer equations, flavor generation kinetic equations, and harmful substance accumulation models. It improves prediction reliability through a hybrid model that integrates mechanism and data. Based on the set optimization objectives, it uses model predictive control algorithms to simulate the results of different temperature adjustment strategies, outputs the best control commands, and directly regulates the stir-frying equipment.

2. The system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot according to claim 1, characterized in that, The multi-probe thermocouple array uses K-type thermocouples, with 12 probes evenly deployed on the inner wall of the frying pot. Among them, 6 probes are vertically inserted into the core area of ​​the material, 3 probes are close to the bottom of the pot to monitor local hot spots, and 3 probes are placed on the surface of the material to monitor the temperature of the volatilization and oxidation layer. The automatic online ultra-fast sampler uses a microfluidic sampling valve. After sampling, the sample is transported through a double-layer stainless steel pipeline. The inner layer is filled with circulating nitrogen gas at 159℃-161℃, and the outer layer is wrapped with an electric heating belt. The mechanical parameters include stirring rate and power input. The stirring rate is monitored by installing a photoelectric encoder on the stirring paddle shaft to collect the rotational speed. The power input is monitored by connecting a power meter in series in the power cord of the induction cooker to record the power input.

3. The system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot according to claim 2, characterized in that, The multi-probe thermocouple array collects data every 0.1 seconds and sends it to the intelligent evaluation module via a wireless transmission module. The core area temperature threshold is set, and if a probe exceeds the range for 2 seconds, an alarm is automatically triggered and the power of the induction cooker is adjusted.

4. The system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot according to claim 1, characterized in that, The specific steps for the intelligent evaluation module to construct the XGBoost model; Construction of the artificial sensory database: Consumers from major target markets in different regions are recruited using stratified sampling techniques. A randomized test sequence design and simulation of real dining scenarios are combined. Sensory ratings on a scale of 0-10 are recorded using an electronic rating system. Facial expression recognition is used to assist in verifying the authenticity of the feedback. Outliers are removed to form a structured database. Data preprocessing and feature engineering: Chemical data from PTR-TOF-MS and LC-MS were aligned using timestamp synchronization technology. Z-score normalization and Min-Max normalization were used to eliminate dimensional differences. Pearson correlation coefficient was used to screen chemical features that were strongly correlated with sensory scores. XGBoost model training: Based on the extreme gradient boosting tree algorithm, the model complexity is controlled by setting hyperparameters, and an early stopping mechanism is used to prevent overfitting. The feature contribution is quantified by the SHAP value, and a mapping from chemical data to sensory scores is established. Prediction and Deployment: The trained model is exported to PMML format and integrated into the intelligent evaluation module. The output includes JSON results containing different sensory scores and confidence intervals. At the same time, the model prediction accuracy is maintained and a closed-loop optimization is formed by combining regular human sensory review and online learning mechanisms.

5. The system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot according to claim 1, characterized in that, The specific steps by which the digital control module constructs the virtual stir-frying process model are as follows: Constructing a mechanistic model: including heat and mass transfer equations, flavor generation kinetics equations, and a harmful substance accumulation model; Data-driven model: The score output by the XGBoost model is used as input to correct the prediction bias of the mechanistic model and captures dynamic relationships not covered by the mechanistic model through the long short-term memory network; Weighted fusion: For quantifiable mechanisms, a mechanistic model is used as the main approach, while for complex reactions, a data-driven model correction is used. The weights are dynamically adjusted through Bayesian optimization, and the model prediction error is evaluated through Monte Carlo simulation.

6. The system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot according to claim 5, characterized in that, The specific steps for constructing the mechanism model are as follows: The heat and mass transfer equations are based on the Navier-Stokes equations and the law of conservation of energy, establishing a heat transfer model between the pot body, materials, and the environment. ,in, For the convective heat transfer term caused by stirring, For exothermic reactions, the heat source term is... For the density of the medium, For isobaric specific heat capacity, The rate of change of temperature over time. Thermal conductivity, For temperature gradient; The flavor generation kinetic equation adopts a first-order reaction kinetic model. ,in, The generation rate constant is temperature-dependent. The concentration of the precursor substance. This refers to the concentration of volatile flavor compounds. This is a temperature-dependent degradation rate constant; The harmful substance accumulation model: Acrylamide formation adopts the Arrhenius model: ,in, The activation energy and pre-exponential factor were fitted through preliminary experiments. This represents the rate of change in amino acid concentration. This refers to the concentration of amino acids. This represents the concentration of carbohydrates.

7. The system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot according to claim 5, characterized in that, The steps for constructing the data-driven model are as follows: Flavor intensity prediction: A flavor intensity prediction function is constructed based on physicochemical principles through a mechanistic model to provide basic prediction values. At the same time, XGBoost is used to train and output different sensory scores S with historical process parameters as input features. Bias correction strategy: The bias correction strategy is triggered by setting a threshold. When the sensory score S is less than the set threshold, the bias correction strategy is applied. At that time, the flavor intensity predicted by the mechanistic model There is an underestimation bias, which is corrected by increasing the weight of the flavor generation term, as shown below. ,in, For correction factor, This represents the partial derivative of the mechanistic model with respect to temperature. For temperature changes over time, This is to allow for reaction time, thereby compensating for prediction biases in the mechanistic model; Residual definition: Defines the predicted value of a mechanistic model. Compared with actual flavor intensity The residual ΔF(t) = - The residuals contain dynamic relationships that are not explained by the mechanistic model; LSTM model input and output design: LSTM is used with time series process parameters as input and the output is the residual at the current time step; LSTM network structure: using the forget gate Controlling the retention of historical information, among which, For the Sigmoid function, This is the weight matrix. This is the hidden state from the previous moment. For the current input, For the bias term of the forget gate, input gate , Controlling the input of new information, among which, The output of the input gate, Here is the weight matrix of the input gate. For the bias term of the input gate, The candidate cell states are generated using the tanh function, representing new information, and their range is between -1 and 1. To generate the weight matrix for candidate cell states, Cell state update for bias terms used to generate candidate cell states Information integration, among which, This represents the current cell state, and ⊙ represents element-wise multiplication. The cell state at the previous time step; output gate , Control status output, where, For the output of the output gate, This is the weight matrix of the output gate. This is the bias term for the output gate. The hidden state at the current moment is used to capture dynamic relationships not covered by the mechanistic model, and the residual prediction value is output through a fully connected layer. ; Model training and loss function: The training objective of LSTM is to minimize the MSE between the residual prediction and the true value. The Adam optimizer is used to update the network parameters and bias terms. Integrated decision: The flavor intensity prediction value is the superposition of the mechanistic model and the LSTM residual.

8. The system for detecting the effect of different frying temperatures on the flavor of beef tallow hot pot according to claim 1, characterized in that, The specific steps by which the digital control module uses model predictive control algorithms to simulate the results of different temperature adjustment strategies are as follows: Optimize target settings: maximize target flavor; minimize harmful substances; Control strategy simulation and optimization: Based on the current state, candidate temperature adjustment strategies are generated. In the future time domain, the trajectory of each strategy is simulated by numerical integration, the objective function value is calculated, and process constraints, safety constraints, and sensory constraints are implemented. Optimal control command output: Select the strategy that satisfies the constraints and minimizes the objective function, convert the control command into an executable format for the device, and send it to the roasting equipment control system via the OPC UA protocol.

Images