Quality monitoring method and system used in production process of lemon-flavored thick milk

Abstract

The invention relates to the technical field of production monitoring control, and discloses a quality monitoring method and system used in the production process of lemon-flavored thick milk. The method comprises the following steps: detecting the concentration of limonene molecules at a production node in real time through a lemon essential oil volatility detection sensor to obtain gas phase release data, performing dynamic modeling analysis based on the data to obtain a limonene binding degree parameter, predicting a flavor intensity attenuation value by utilizing a CLFK-BP neural network algorithm, and determining the flavor intensity of the lemon essential oil. And establishing a lemon flavor fingerprint spectrum, calculating a flavor compound characteristic ratio, finally carrying out abnormity identification processing through a CF-LS-SVM diagnosis algorithm, and outputting a quality abnormity type and a regulation and control instruction. The technical problems that in the production process of the lemon-flavored thick milk, interaction of limonene-protein molecules cannot be monitored in real time, and flavor attenuation is difficult to accurately predict are solved, and the quality monitoring precision and the flavor stability control level of the lemon-flavored thick milk are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of production monitoring and control technology, and in particular to a quality monitoring method and system for the production process of lemon-flavored thick cream. Background Technology

[0002] In the current production process of lemon-flavored thick cream, quality control mainly relies on traditional physicochemical testing methods and sensory evaluation, including routine indicators such as pH measurement, viscosity testing, and microbial counting, as well as manual sensory evaluation to determine the flavor characteristics and overall quality of the product. These methods have been widely used in the food industry and can effectively control the basic physicochemical properties and safety of the product, providing a fundamental quality assurance means for the standardized production of lemon-flavored thick cream.

[0003] However, existing technologies have significant shortcomings: First, the detection timeliness is poor, with most physicochemical and microbiological tests requiring a long time to produce results, making true real-time monitoring impossible and leading to delays in the detection of quality problems; second, the detection accuracy and objectivity are limited, with key quality indicators such as lemon aroma intensity and product texture characteristics still mainly relying on human sensory evaluation, resulting in strong subjectivity and poor reproducibility; third, there is a lack of effective monitoring of microscopic changes during the production process, especially for key factors affecting flavor stability such as the interaction between lemon essential oil molecules and milk proteins, and the degree of fat oxidation.

[0004] More importantly, existing technologies cannot solve the unique technical challenges of lemon-flavored thick cream: the limonene molecules in lemon essential oil undergo a complex dynamic binding process with milk proteins. This molecular-level interaction changes continuously in different process stages (homogenization, heat treatment, cooling), directly affecting the flavor release intensity and persistence of the final product. Traditional monitoring methods cannot capture the dynamic changes of this microscopic molecular interaction in real time, making it impossible to accurately predict the decline trend of lemon flavor. Consequently, it is impossible to adjust process parameters in a timely manner during production to maintain the stability of flavor quality, ultimately affecting the sensory quality and shelf life of the product. Summary of the Invention

[0005] This application provides a quality monitoring method and system for the production process of lemon-flavored thick cream, which solves the technical problems of the inability to monitor the interaction between limonene and protein molecules in real time and the difficulty in accurately predicting flavor decay during the production process of lemon-flavored thick cream, and significantly improves the accuracy of quality monitoring and the level of flavor stability control of lemon-flavored thick cream.

[0006] In a first aspect, this application provides a quality monitoring method for the production process of lemon-flavored thick cream, the quality monitoring method for the production process of lemon-flavored thick cream comprising: The concentration of limonene molecules at the production stage is detected in real time using a lemon essential oil volatility detection sensor to obtain limonene gas phase release data. Based on the gas-phase release data of limonene, a kinetic modeling analysis of the limonene-protein binding state was performed to obtain the limonene binding degree parameter. The flavor intensity prediction process for the limonene binding degree parameter was performed using the CLFK-BP neural network algorithm to obtain the predicted value of lemon flavor decay. A lemon flavor fingerprint spectrum is established based on the predicted lemon flavor decay values, and the characteristic ratios of flavor compounds are calculated using a concentration ratio matrix. The characteristic ratios of the flavor compounds are input into the CF-LS-SVM diagnostic algorithm for anomaly identification to obtain the quality anomaly type and control instructions.

[0007] Secondly, this application provides a quality monitoring system for the production process of lemon-flavored thick milk, the quality monitoring system for the production process of lemon-flavored thick milk comprising: The detection module is used to detect the concentration of limonene molecules at the production node in real time using a lemon essential oil volatility detection sensor, and obtain limonene gas phase release data. The analysis module is used to perform kinetic modeling and analysis of the limonene-protein binding state based on the limonene gas phase release data, and obtain the limonene binding degree parameter. The prediction module is used to perform flavor intensity prediction processing on the limonene binding degree parameter using the CLFK-BP neural network algorithm to obtain the predicted value of lemon flavor decay. The calculation module is used to establish a lemon flavor fingerprint spectrum based on the predicted lemon flavor decay value, and to calculate the characteristic ratio of flavor compounds through the concentration ratio matrix. The identification module is used to input the characteristic ratio of the flavor compounds into the CF-LS-SVM diagnostic algorithm for anomaly identification processing to obtain the quality anomaly type and control instructions.

[0008] Thirdly, a quality monitoring device for the production process of lemon-flavored thick milk is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the quality monitoring device for the production process of lemon-flavored thick milk to execute the aforementioned quality monitoring method for the production process of lemon-flavored thick milk.

[0009] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned quality monitoring method for the production process of lemon-flavored thick milk.

[0010] The technical solution provided in this application utilizes a lemon essential oil volatility detection sensor to monitor the concentration of limonene molecules at production nodes in real time. This overcomes the technical bottleneck of traditional quality monitoring methods, which cannot capture dynamic changes at the molecular level, and achieves precise real-time tracking of the core components of lemon flavor. Based on kinetic modeling analysis using limonene gas-phase release data, a quantitative description system for the limonene-protein binding state was established for the first time, resolving the key issue of unclear interaction mechanisms between flavor molecules and the matrix. The acquisition of limonene binding degree parameters lays a scientific foundation for subsequent accurate prediction. The CLFK-BP neural network algorithm was optimized for the specific characteristics of lemon-flavored thick milk systems. By introducing limonene-specific parameters and temperature-sensitive weighting factors, the accuracy of predicting lemon flavor decay patterns was significantly enhanced, overcoming the limitation of general algorithms' insufficient applicability in specific food systems.

[0011] The establishment of lemon flavor fingerprint profiles and the calculation of concentration ratio matrices innovatively transform the complex composition of flavor compounds into quantifiable digital features, providing a scientific basis for the objective evaluation of flavor quality and completely changing the traditional model that relies on subjective sensory evaluation. The CF-LS-SVM diagnostic algorithm, through a multi-level classification architecture and optimization of lemon flavor fingerprint features, achieves intelligent diagnosis across the entire chain, from anomaly detection to type identification and degree quantification. It can not only accurately identify eight typical flavor defects such as limonene oxidation and citral loss, but more importantly, it can output specific process adjustment parameters, forming a closed-loop management system from monitoring to diagnosis to control. The organic integration of the entire technical solution solves the technical challenges of lagging quality monitoring, subjectivity, and blind control in the production of lemon-flavored thick cream, providing a complete technical solution for the standardized and intelligent production of lemon-flavored thick cream, and significantly improving the consistency and stability of product quality. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an embodiment of a quality monitoring method used in the production process of lemon-flavored thick cream in this application. Figure 2 This is a schematic diagram of an embodiment of a quality monitoring system used in the production process of lemon-flavored thick cream in this application. Figure 3 This is a schematic block diagram of the quality monitoring equipment used in the production process of lemon-flavored thick cream in an embodiment of the present invention. Detailed Implementation

[0014] This application provides a quality monitoring method and system for the production process of lemon-flavored thick cream. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the quality monitoring method for the production process of lemon-flavored thick milk in this application includes: Step S101: The concentration of limonene molecules at the production node is detected in real time using a lemon essential oil volatility detection sensor to obtain limonene gas phase release data. Step S102: Based on the gas phase release data of limonene, perform kinetic modeling and analysis on the binding state of limonene-protein to obtain the limonene binding degree parameter; Step S103: The flavor intensity prediction of the limonene binding degree parameter is processed by the CLFK-BP neural network algorithm to obtain the predicted value of lemon flavor decay. Step S104: Establish a lemon flavor fingerprint spectrum based on the predicted lemon flavor decay value, and calculate the characteristic ratio of flavor compounds through the concentration ratio matrix; Step S105: Input the flavor compound characteristic ratio into the CF-LS-SVM diagnostic algorithm for anomaly identification processing to obtain the quality anomaly type and control instructions.

[0016] It is understood that the executing entity of this application can be a quality monitoring system used in the production process of lemon-flavored thick cream, or it can be a terminal or a server; no specific limitation is made here. This application's embodiments use a server as an example for illustration.

[0017] Specifically, this method for monitoring the quality of lemon-flavored thick cream involves deploying gas chromatography-mass spectrometry (GC-MS) sensors at six key production nodes: the raw material mixing tank, the homogenizer inlet and outlet, the heat processor inlet and outlet, and the cooler outlet, to construct a limonene molecule concentration monitoring network. The sensors perform qualitative and quantitative analysis of limonene molecules based on the principle of gas chromatography separation. The limonene retention time is set to 8.32 minutes, the concentration detection range is 0.1-50 mg / L, and the detection accuracy reaches ±0.05 mg / L. After acquiring the raw limonene gas phase concentration signal, the sensors perform filtering and amplification to remove noise interference. Then, the concentration change rate within a 30-second response time is calculated to obtain the limonene release rate at each production node. Based on the principle of gas-liquid equilibrium, the system performs liquid phase concentration back-calculation on the limonene gas phase concentration data to determine the distribution ratio of free and bound limonene in the liquid phase. Finally, the limonene concentration distribution data is time-stamped with synchronously acquired process parameters such as temperature, pressure, flow rate, and pH to obtain limonene gas phase release data containing concentration-process parameter correlation information.

[0018] A kinetic modeling analysis was conducted using limonene gas-phase release data to establish a limonene molecular diffusion kinetic model. The concentrations of free limonene, the concentration of bindable protein sites, and the concentration of the limonene-protein complex were modeled as state variables. The system analyzed the α-helix content changes of β-casein using circular dichroism spectroscopy. A protein conformational change marker was triggered when the α-helix content decreased by more than 15%. The binding rate constant was dynamically adjusted based on the conformational change markers, with the binding rate constant ranging from 0.02 to 0.15, the dissociation rate constant from 0.01 to 0.08, and the volatilization rate constant from 0.005 to 0.025. A nonlinear adaptive control algorithm was used to solve the limonene molecular diffusion kinetic model, obtaining real-time binding and dissociation rates of limonene and protein. These rates were then normalized, and the percentage of limonene molecule encapsulation efficiency in the milk fat environment was calculated to obtain the limonene binding degree parameter, characterizing binding stability.

[0019] The CLFK-BP neural network algorithm constructs a 13-dimensional input vector by combining the limonene binding degree parameter with initial limonene concentration, β-casein concentration, milk fat content, temperature, pH value, homogenization pressure, homogenization time, heat treatment temperature, heat treatment time, cooling rate, product viscosity, and storage time. Weights are assigned to the 13-dimensional input vector based on the temperature-sensitive limonene volatility coefficient, with the pre-exponential factor of the volatility coefficient set to 2.3 × 10^8 and the activation energy set to 58.2 kJ / mol, resulting in temperature-weighted input data. This temperature-weighted input data is then used to construct a 16-hidden-layer neural network for forward propagation calculations. The hidden-layer activation function is a modified hyperbolic tangent function, and the limonene-specific correction parameter is set to 0.15. The backpropagation algorithm iteratively optimizes the network weights, with a learning rate of 0.001 and 5000 training iterations. Training stops when the prediction error converges to within 2%. The trained neural network performs calculations on the three output neurons, outputting the lemon flavor intensity decay rate, the limonene residue rate after 24 hours, and the product shelf life in days, respectively, to obtain the predicted lemon flavor decay value.

[0020] A lemon flavor fingerprint was established using gas chromatography-mass spectrometry (GC-MS) to detect lemon-flavored thick cream based on predicted lemon flavor decay values. Concentration data for 15 key flavor compounds, including limonene, citral, citric acid, linalool, citronellol, ethyl citrate, methyl citrate, nerol, geraniol, limonene oxide, citral dimethyl acetal, triethyl citrate, α-pinene, β-pinene, and limonene dioxide, were obtained. A concentration ratio matrix was constructed based on the detected flavor compound concentration data, and the concentration ratios between compounds were calculated. The retention times for limonene and citral were 8.32 min and 12.45 min, respectively, while the retention time for citric acid was 15.67 min. The concentration ratio matrix was standardized, with the standard ratio range for a normal lemon flavor fingerprint set as follows: limonene to citral ratio 3.2-4.1, citral to citric acid ratio 0.8-1.2, and linalool to citronellol ratio 1.5-2.3. Based on eight typical flavor defect patterns—limonene oxidation, citral loss, rancidity, off-flavor formation, textural abnormalities, insufficient aroma intensity, bitterness formation, and acidity imbalance—standardized ratio data were classified and labeled, and feature identification thresholds for each defect type were established. The feature identification thresholds were compared and analyzed with the measured ratio data. When the limonene to citral ratio dropped below 2.1, it was labeled as limonene oxidation; when the absolute concentration of citral was below 0.5 mg / L, it was labeled as citral loss. This yielded flavor compound characteristic ratios that included defect type identifiers.

[0021] The CF-LS-SVM diagnostic algorithm constructs a 15-dimensional feature vector from the feature ratios of flavor compounds. A first-stage radial basis function kernel is used for flavor anomaly detection, with the kernel parameter set to 0.125, yielding anomaly detection and classification results. Based on these classification results, the detected anomaly samples are categorized. Eight typical flavor defects are input into corresponding multinomial kernel classifiers for processing. The first parameter of the multinomial kernel is set to 0.8, the second parameter to 0.2, and the power parameter to 3. The categorization results quantify the degree of anomaly, outputting an anomaly score of 0-100 using a linear kernel function with a regularization parameter set to 0.01. The Lagrange multiplier method is used to solve for the optimal classification hyperplane based on the anomaly score. A green warning and control instruction is output when the anomaly score is 0-30, a yellow warning and control instruction is output when the score is 30-60, and a red warning and control instruction is output when the score is 60-100. The early warning and control instructions establish an abnormality-control mapping relationship. When an abnormality of limonene oxidation is detected, a control instruction to reduce the heat treatment temperature by 2-5℃ is output. When an abnormality of citral loss is detected, a control instruction to increase the amount of lemon essential oil added by 5-15% is output. This results in quality abnormality types and control instructions that include specific process adjustment parameters.

[0022] In one specific embodiment, the process of performing step S101 may specifically include the following steps: Gas chromatography-mass spectrometry (GC-MS) sensors were deployed at six key production nodes: raw material mixing tank, homogenizer inlet, homogenizer outlet, heat processor inlet, heat processor outlet, and cooler outlet, to construct a limonene molecule concentration monitoring network. Qualitative and quantitative analysis of limonene molecules at each node was performed based on the principle of gas chromatography separation. The retention time of limonene was set to 8.32 minutes, the concentration detection range was 0.1-50 mg / L, and the detection accuracy was ±0.05 mg / L. The original signal of limonene gas phase concentration was filtered and amplified, and the limonene release rate of each production node was calculated by the concentration change rate within a 30-second response time. Based on the principle of gas-liquid equilibrium, the liquid phase concentration was calculated by back-calculating the gas phase concentration data of limonene to determine the distribution ratio of free limonene and bound limonene in the liquid phase. The limonene concentration distribution data was matched with the synchronously collected process parameters such as temperature, pressure, flow rate, and pH value using timestamps to obtain limonene gas phase release data containing concentration-process parameter correlation information.

[0023] Specifically, the quality monitoring method in the production process of lemon-flavored thick cream constructs a limonene molecule concentration monitoring network by deploying gas chromatography-mass spectrometry (GC-MS) sensors at six key production nodes. The raw material mixing tank monitors the dispersion state after the initial addition of lemon essential oil; the homogenizer inlet monitors the uniformity of raw material mixing; the homogenizer outlet detects the limonene encapsulation effect after homogenization; the heat processor inlet and outlet monitor the stability changes of limonene molecules before and after heat treatment; and the cooler outlet monitors the retention of limonene in the final product. The GC-MS sensor is an analytical instrument that combines gas chromatography separation technology with mass spectrometry detection technology. The gas chromatography section separates the components in the complex mixture according to their boiling points and polarity differences, while the mass spectrometry section performs qualitative and quantitative analysis of the separated components based on the mass-to-charge ratio of the molecular ion peaks.

[0024] In the qualitative and quantitative analysis based on the principle of gas chromatography, limonene molecules are separated within the chromatographic column according to their physicochemical properties. Due to limonene's molecular weight of 136 and its specific volatility, its retention time in the column is set to 8.32 minutes, a fixed value obtained through standard calibration. The concentration detection range is set to 0.1-50 mg / L, covering the normal fluctuation range of limonene concentration during the production of lemon-flavored thick cream. A detection accuracy of ±0.05 mg / L ensures accurate capture of minute concentration changes. The mass spectrometer detector ionizes limonene molecules through electron impact ionization. The resulting ions are separated according to their mass-to-charge ratio under the influence of a magnetic field. The detector records the abundance of ions with different mass-to-charge ratios. Qualitative analysis is completed by comparison with a standard spectral library, and quantitative analysis is completed through the linear relationship between ion abundance and concentration.

[0025] Filtering and amplification are crucial steps in signal processing. The raw limonene gas phase concentration signal contains both useful signals and noise interference. Filters remove high-frequency noise and low-frequency drift by setting specific frequency response characteristics, while amplifiers amplify the weak detection signal to a suitable amplitude range for subsequent processing. The concentration change rate within a 30-second response time is calculated by continuously acquiring concentration data points over 30 seconds, dividing the difference between the initial and final concentrations by the time interval. This rate reflects the dynamic process of limonene molecules transferring from the liquid phase to the gas phase. The limonene release rate at each production node is calculated by multiplying the concentration change rate by the gas phase volumetric flow rate. The release rate data directly reflects the volatilization behavior of limonene molecules under different process conditions.

[0026] The principle of gas-liquid equilibrium is based on Henry's Law, which states that at a given temperature, the partial pressure of a component in the gas phase is directly proportional to its mole fraction in the liquid phase. Limonene gas phase concentration data is converted to liquid phase concentration using Henry's constant, a specific physical constant of limonene in a cream matrix that is related to temperature and matrix composition. The liquid phase concentration is calculated by dividing the gas phase concentration by Henry's constant to obtain the total liquid phase concentration. Then, the concentration of free limonene is calculated based on the equilibrium constant of limonene binding to proteins. The concentration of bound limonene equals the total liquid phase concentration minus the free concentration. Free limonene refers to limonene molecules that are not bound to proteins, while bound limonene refers to limonene that is bound to protein molecules through hydrophobic interactions. The distribution ratio of these two types directly affects the intensity and persistence of lemon flavor release.

[0027] Timestamp matching is the core technology for data synchronization. The acquisition time of limonene concentration distribution data must precisely correspond to that of process parameters such as temperature, pressure, flow rate, and pH. Timestamp accuracy at the millisecond level ensures data consistency. Temperature data is acquired via thermocouple sensors, pressure data via pressure transmitters, flow rate data via flow meters, and pH value via online pH electrodes. The data matching process uses timestamp sorting to combine different parameter data at the same moment into multi-dimensional data vectors. Each data vector contains information in seven dimensions: limonene concentration, free state ratio, bound state ratio, temperature, pressure, flow rate, and pH value. Correlation analysis is used to determine the mathematical relationship between limonene concentration changes and various process parameters. Increased temperature leads to enhanced limonene volatilization, pressure changes affect the gas-liquid equilibrium position, flow rate affects the mass transfer rate, and pH value affects protein conformation, thus influencing binding capacity.

[0028] In one specific embodiment, the process of performing step S102 may specifically include the following steps: A molecular diffusion dynamics model of limonene was established based on limonene gas phase release data, and the concentration of free limonene, the concentration of bindable protein sites, and the concentration of limonene-protein complex were modeled as state variables. Changes in α-helix content of β-casein were analyzed by circular dichroism spectroscopy. A protein conformational change marker was triggered when the α-helix content decreased by more than 15%. The binding rate constant was dynamically adjusted based on protein conformational change markers, with the binding rate constant range set to 0.02-0.15, the dissociation rate constant range set to 0.01-0.08, and the volatilization rate constant range set to 0.005-0.025. The diffusion dynamics model of limonene molecules was solved using a nonlinear adaptive control algorithm to obtain the real-time binding rate and dissociation rate of limonene with proteins. The binding rate and dissociation rate values ​​were normalized, and the percentage of limonene molecules encapsulated in the milk fat environment was calculated to obtain the limonene binding degree parameter characterizing the binding stability.

[0029] Specifically, the limonene molecular diffusion kinetics model is established based on the aforementioned limonene gas-phase release data. The free limonene concentration is defined as the concentration of limonene molecules that are not bound to proteins and can diffuse freely in the liquid phase. The concentration of bindable protein sites refers to the number of effective sites on the surface of β-casein molecules that are hydrophobic and can bind to limonene molecules. The limonene-protein complex concentration represents the concentration of a stable complex of limonene molecules and protein sites. The kinetic model describes the dynamic relationship between the three state variables through a system of differential equations. The rate of change of the free limonene concentration is equal to the dissociation rate of the limonene-protein complex minus the binding rate of free limonene to bindable protein sites minus the volatilization rate of free limonene. The rate of change of the bindable protein site concentration is equal to the site release rate of the complex minus the site occupancy rate. The rate of change of the limonene-protein complex concentration is equal to the binding rate minus the dissociation rate.

[0030] Circular dichroism (CD) chromatography is a spectroscopic technique for analyzing the secondary structure of proteins. Based on the different absorption characteristics of chiral amino acid residues in protein molecules to left-handed and right-handed circularly polarized light, the α-helix structure of β-casein appears as a negative peak at a specific wavelength in CD chromatography. The α-helix content is calculated by dividing the molar ellipticity at the characteristic wavelength by the standard α-helix molar ellipticity. In the detection process, a diluted β-casein solution is placed in the sample cell of a CD chromatograph, and the scanning wavelength range covers the far-ultraviolet region. The CD signal intensity at different wavelengths is recorded, and the contributions of α-helices, β-sheets, and random coil structures are separated using a spectral deconvolution algorithm to calculate the relative content of each secondary structure. The criterion for a decrease in α-helix content exceeding 15% is based on the stability range of the β-casein secondary structure under normal lemon-flavored thick milk production conditions. When a decrease in α-helix content relative to the initial value exceeds 15%, a protein conformational change marker is triggered, which serves as a trigger signal for adjusting kinetic parameters.

[0031] The binding rate constant characterizes the speed at which limonene molecules form a complex with bindable protein sites. Its value ranges from 0.02 to 0.15, corresponding to the binding kinetics under different temperature and pH conditions. Increased temperature intensifies molecular thermal motion and increases the binding rate. When the pH deviates from the isoelectric point, protein conformational changes affect the accessibility of binding sites. The dissociation rate constant describes the rate at which the limonene-protein complex decomposes into free limonene and bindable sites. Its value ranges from 0.01 to 0.08, reflecting the range of complex stability. A lower dissociation rate constant indicates a stronger complex bond. The volatilization rate constant quantifies the rate at which free limonene transfers from the liquid phase to the gas phase. Its value ranges from 0.005 to 0.025 and is related to the vapor pressure of limonene, the liquid-phase diffusion coefficient, and the gas-liquid interface mass transfer coefficient. After a protein conformational change is triggered, the binding rate constant is linearly interpolated based on the degree of decrease in α-helix content. The greater the decrease in α-helix content, the smaller the binding rate constant. Similarly, the dissociation rate constant and volatilization rate constant are adjusted accordingly based on the degree of protein conformational change.

[0032] The nonlinear adaptive control algorithm is a numerical solution method based on the feedback principle. It adjusts the solution parameters by real-time monitoring of the state variable deviations in the limonene molecular diffusion dynamics model. The algorithm first sets initial binding and dissociation rates, then calculates the error between the model's predicted and actual measured values ​​based on the current state variables. This error information is fed back through an adaptive law to adjust the model parameters. The adaptive law determines the direction and magnitude of parameter adjustment based on the magnitude and trend of the error. The algorithm's nonlinearity is reflected in the consideration of the nonlinear interaction between limonene concentration and protein site concentration during parameter adjustment. When the limonene concentration is high, intense competition for binding sites leads to a decrease in binding efficiency; when protein sites are saturated, further increasing the limonene concentration does not significantly increase the complex formation rate. The solution calculation uses numerical integration to progressively calculate the evolution trajectory of each state variable over time. The integration step size is adaptively adjusted according to the system's dynamic characteristics to ensure the stability and accuracy of the numerical solution.

[0033] Normalization converts the binding and dissociation rates into ratios relative to the maximum theoretical values. The binding rate normalization is obtained by dividing the current binding rate by the theoretical maximum binding rate, which corresponds to the ideal state where all bindable protein sites are completely occupied by limonene molecules. The dissociation rate normalization is calculated by dividing the current dissociation rate by the theoretical maximum dissociation rate, which corresponds to the limiting state where all limonene-protein complexes are completely dissociated. The encapsulation efficiency of limonene molecules in the milk fat environment is calculated by subtracting the normalized dissociation rate from the normalized binding rate and then multiplying by a correction factor. The correction factor considers the influence of the milk fat matrix on the diffusion and binding process of limonene molecules; a higher milk fat content results in a larger correction factor value, indicating that the lipid environment is conducive to the stable encapsulation of limonene. The encapsulation efficiency percentage directly reflects the degree to which limonene molecules are effectively encapsulated by proteins and prevented from volatilization and loss. This value is used as a limonene binding degree parameter input into the subsequent flavor intensity prediction algorithm.

[0034] In one specific embodiment, the process of executing step S103 may specifically include the following steps: A 13-dimensional input vector was constructed by combining the limonene binding degree parameter with the initial limonene concentration, β-casein concentration, milk fat content, temperature, pH value, homogenization pressure, homogenization time, heat treatment temperature, heat treatment time, cooling rate, product viscosity, and storage time. The 13-dimensional input vector was weighted based on the temperature-sensitive limonene volatility coefficient. The pre-exponential factor of the volatility coefficient was set to 2.3 × 10^8, and the activation energy was set to 58.2 kJ / mol to obtain temperature-weighted input data. Temperature-weighted input data was used to construct a neural network with 16 hidden layers for forward propagation calculation. The activation function of the hidden layers was the modified hyperbolic tangent function, and the limonene-specific correction parameter was set to 0.15. The network weights are iteratively optimized using the backpropagation algorithm, with a learning rate of 0.001 and 5000 training iterations. Training stops when the prediction error converges to within 2%. The trained neural network is used to calculate and process the three output neurons, which output the lemon flavor intensity decay rate, the limonene residue rate after 24 hours, and the product shelf life days, respectively, to obtain the predicted value of lemon flavor decay.

[0035] Specifically, the 13-dimensional input vector construction of the CLFK-BP neural network algorithm uses the aforementioned limonene binding degree parameter as the first dimension. The initial limonene concentration is derived from the essential oil addition amount determination during the ingredient preparation stage. The β-casein concentration is calculated by determining the protein content in the raw milk using the Kjeldahl method. The milk fat content is obtained by the Babcock method or infrared spectroscopy. Temperature data is obtained from real-time monitoring by thermocouple sensors during the production process. The pH value is continuously measured by an online pH electrode. The homogenization pressure is recorded by a pressure sensor, which records the working pressure of the homogenizer. The homogenization time is calculated based on the residence time of the material in the homogenizer. The heat treatment temperature is obtained from the temperature controller of the sterilization equipment. The heat treatment time is determined based on the residence time of the material in the heat processor. The cooling rate is calculated by the rate of change of temperature over time. The product viscosity is measured online using a rotational viscometer. The storage time is counted from the completion of product packaging. Data preprocessing of the 13-dimensional input vector includes outlier detection and data standardization. Outliers are identified and removed using a box plot method. Data standardization converts the data of each dimension into a standard normal distribution with zero mean and unit variance.

[0036] The temperature-sensitive limonene volatility coefficient was established based on the Arrhenius equation. The relationship between the volatility coefficient and temperature is an exponential function, with a pre-exponential factor of 2.3 × 10^8 representing the limit value of the volatility coefficient as the temperature approaches infinity. The activation energy of 58.2 kJ / mol represents the minimum energy required for limonene molecules to overcome the potential barrier and transfer from the liquid phase to the gas phase. Weighting was performed by calculating the volatility coefficient at the temperature corresponding to each component of the 13-dimensional input vector and multiplying each component by its corresponding volatility coefficient weight. Temperature-related components, such as heat treatment temperature and cooling rate, directly affect the volatility coefficient calculation, while other components have their corresponding volatility coefficient weights calculated using the average temperature. The temperature-weighted input data reflects the differentiated influence of various process parameters on the limonene volatility behavior under different temperature conditions. Under high-temperature conditions, the weight of temperature-related parameters increases significantly, while under low-temperature conditions, the relative importance of non-temperature-related parameters increases.

[0037] The forward propagation computation of the 16 hidden-layer neuron network maps temperature-weighted input data to hidden-layer outputs through weighted summation and activation function transformation. Each hidden-layer neuron receives 13 input signals, which are linearly combined through corresponding connection weights, and then a bias term is added before being input to the activation function for nonlinear transformation. The modified hyperbolic tangent function is an improved activation function based on the standard hyperbolic tangent function, incorporating a limonene-specific correction parameter. The correction parameter of 0.15 is determined based on the diffusion characteristics of limonene molecules in the milk fat environment. The modified activation function expands and shifts the output range, better adapting to the nonlinear characteristics of the lemon-flavored thick milk system. During forward propagation, the outputs of each hidden-layer neuron are calculated by weighted summation of the responses of the three neurons in the output layer. The output layer directly outputs the linear combination result without using an activation function.

[0038] The backpropagation algorithm iteratively optimizes network weights using gradient descent. The algorithm first calculates the error between the output layer's predicted value and the target value, then calculates the partial derivative of the error with respect to the weights of each layer according to the chain rule. A learning rate of 0.001 controls the step size of weight updates; a smaller learning rate ensures the stability of the training process but requires more iterations. The 5000 training epochs are a convergence requirement determined based on the characteristics of the lemon-flavored thick milk data. Each training epoch includes four steps: forward propagation to calculate the predicted value, error calculation, backpropagation to calculate the gradient, and weight updates. The prediction error is calculated using the root mean square error. Convergence is considered achieved when the error change over 100 consecutive training epochs is less than 2%, and the training process automatically stops to prevent overfitting. The momentum method is used for weight updates to accelerate convergence; the momentum term considers the direction and magnitude of the previous weight update, which helps to escape local optima.

[0039] The trained neural network processes three output neurons. The first output neuron predicts the lemon flavor intensity decay rate, which represents the rate at which the lemon flavor deteriorates during storage. This is calculated by the rate of change of the ratio of the current flavor intensity to the initial flavor intensity over time. The second output neuron predicts the limonene residue rate after 24 hours, reflecting the retention degree of limonene molecules in the product. This is expressed as the ratio of the limonene concentration after 24 hours to the initial concentration. The third output neuron predicts the product's shelf life in days, calculated based on the time required for the flavor intensity to decay to an acceptable threshold, which is set at 70% of the initial flavor intensity. The predicted lemon flavor decay values ​​are combined from the results of the three output neurons, and a flavor stability evaluation index is derived using a weighted average or a comprehensive scoring method.

[0040] In one specific embodiment, the process of executing step S104 may specifically include the following steps: Based on the predicted value of lemon flavor decay, gas chromatography-mass spectrometry was used to detect 15 key flavor compounds in lemon-flavored thick cream, including limonene, citral, citric acid, linalool, citronellol, ethyl citrate, methyl citrate, nerol, geraniol, limonene oxide, citral dimethyl acetal, triethyl citrate, α-pinene, β-pinene, and limonene dioxide. A concentration ratio matrix was established based on the detected flavor compound concentration data, and the concentration ratio relationship between compounds was calculated. The retention times of limonene and citral were set to 8.32 minutes and 12.45 minutes, respectively, and the retention time of citric acid was set to 15.67 minutes. The concentration ratio matrix was standardized, and the standard ratio range of normal lemon flavor fingerprint was set as follows: limonene to citral ratio 3.2-4.1, citral to citric acid ratio 0.8-1.2, and linalool to citronellol ratio 1.5-2.3. Based on eight typical flavor defect patterns, namely limonene oxidation, citral loss, rancidity, off-flavor formation, texture abnormality, insufficient aroma intensity, bitterness formation, and acidity imbalance, standardized ratio data are classified and labeled, and feature identification thresholds for each defect type are established. By comparing and analyzing the feature recognition threshold with the measured ratio data, when the ratio of limonene to citral drops below 2.1, it is marked as limonene oxidation, and when the absolute concentration of citral is below 0.5 mg / L, it is marked as citral loss, thus obtaining the feature ratio of flavor compounds including defect type identifiers.

[0041] Specifically, the lemon flavor fingerprint was established based on the aforementioned predicted lemon flavor decay values. A comprehensive compound analysis of the lemon-flavored thick cream sample was performed using gas chromatography-mass spectrometry (GC-MS). The detection process included four steps: sample pretreatment, chromatographic separation, mass spectrometry identification, and quantitative analysis. Sample pretreatment involved enriching volatile flavor compounds using solid-phase microextraction (SPE). The extraction head adsorbed volatile components in the sample headspace, and then the adsorbed compounds were released through high-temperature desorption at the GC inlet. The detection of 15 key flavor compounds was qualitatively confirmed based on the characteristic ions and retention times of each compound. Limonene, as the main terpene compound, eluted at 8.32 min; citral, as an aldehyde flavor compound, was detected at 12.45 min; and citric acid, as an organic acid, showed a chromatographic peak at 15.67 min. Other compounds, such as linalool, citronellol, ethyl citrate, methyl citrate, nerol, geraniol, limonene oxide, citral dimethyl acetal, triethyl citrate, α-pinene, β-pinene, and limonene dioxide, appeared within specific retention time periods. Concentration data were quantitatively calculated using the standard curve method. The standard curve was established using standards of known concentrations, and the concentration of each compound in the sample was calculated based on the correspondence between the chromatographic peak area and the standard curve.

[0042] The concentration ratio matrix was constructed by organizing the concentration data of 15 flavor compounds into a 15×15 symmetrical matrix. Each element in the matrix represents the ratio of the concentration of a compound in a row to the concentration of a compound in a column, with diagonal elements always set to 1, indicating the ratio of a compound to itself. The concentration ratios between compounds reflect the relative content changes of different flavor components. The ratio of limonene to citral reflects the balance between terpenes and aldehydes, the ratio of citral to citric acid reflects the coordination between aldehydes and acids, and the ratio of linalool to citronellal characterizes the internal structural relationships of alcohols. Retention time was set as the time standard for chromatographic separation. The retention times of limonene (8.32 min), citral (12.45 min), and citric acid (15.67 min) were determined through repeated experiments. These time points serve as the basis for qualitative identification of compounds, ensuring the accuracy and reproducibility of the detection results.

[0043] Standardization converts the raw ratio data in the concentration ratio matrix into standardized values. The standardization process uses the Z-score method, which involves subtracting the mean of each ratio from the mean of normal samples and then dividing by the standard deviation. The standardized ratio range for the normal lemon flavor fingerprint was determined based on statistical analysis of a large number of normal samples. The limonene to citral ratio of 3.2-4.1 reflects the typical ratio of terpenes to aldehydes in normal lemon-flavored thick milk; a ratio that is too high indicates a relative deficiency of citral, while a ratio that is too low indicates a low limonene content or a high citral content. The citral to citric acid ratio of 0.8-1.2 reflects the normal balance between aldehydes and acids; ratios outside this range indicate an imbalance in the flavor structure. The linalool to citronellol ratio of 1.5-2.3 characterizes the normal distribution of alcohols; deviations from this range indicate abnormal changes in alcohol flavor components. Standardized ratio data eliminates systematic errors between different batches of samples, enhancing the comparability and consistency of the flavor fingerprint.

[0044] The classification markers for eight typical flavor defect patterns were established based on flavor science theory and practical production experience. Limonene oxidation defects are mainly caused by the reaction of limonene molecules with oxygen during storage, resulting in oxidation products and manifested as a decrease in limonene concentration and an increase in oxide concentration. Citral loss defects originate from the volatilization or chemical degradation of citral molecules, leading to a significant decrease in the absolute concentration of citral. Rancidity defects are caused by aldehydes and ketones produced by lipid oxidation, manifested as the appearance of abnormal aldehyde compounds. Off-flavor defects involve the formation of non-target compounds, such as microbial metabolites or packaging material migrations. Texture abnormalities affect the physical properties of the product, indirectly reflected in changes in the ratios of certain compounds. Insufficient aroma intensity defects are characterized by an overall low concentration of major flavor compounds. Bitterness defects are related to the excessive oxidation of certain terpenes. Acidity imbalance defects are manifested as abnormal changes in the proportion of organic acid compounds. Feature identification thresholds were determined through comparative analysis of defective samples and normal samples; each defect type corresponds to a specific abnormal pattern of compound concentration or ratio.

[0045] The comparative analysis of feature recognition thresholds and measured ratios was conducted by comparing the measured data with the threshold range one by one. When the ratio of limonene to citral dropped below 2.1, it indicated that the relative content of citral was too high or the content of limonene was severely insufficient. Combined with the increase in the concentration of limonene oxide, this was confirmed as a limonene oxidation defect. When the absolute concentration of citral was below 0.5 mg / L, it was directly marked as a citral loss defect. This threshold was determined based on the lowest effective concentration of citral contributing to the overall flavor. The comparative analysis also considered the comprehensive changes of multiple indicators. An abnormality in a single indicator was insufficient to determine the defect type; a comprehensive judgment was needed by combining the synergistic changes of related compounds. The defect type was identified using a numerical coding method. The normal state was coded as 0, and the eight defect types were coded as 1-8 respectively. Multiple defects were represented by combined codes.

[0046] In one specific embodiment, the process of executing step S105 may specifically include the following steps: Flavor compound feature ratios were constructed into a 15-dimensional feature vector. Flavor anomaly detection was performed using a first-stage radial basis function kernel function with a kernel function parameter of 0.125 to obtain anomaly detection and classification results. Based on the anomaly detection classification results, the detected abnormal samples are identified by type. The eight typical flavor defects are input into the corresponding polynomial kernel function classifier for processing. The first parameter of the polynomial kernel function is set to 0.8, the second parameter is 0.2, and the power parameter is 3. The degree of anomaly is quantitatively scored based on the type identification results. An anomaly score of 0-100 is calculated and processed using a linear kernel function, with the regularization parameter set to 0.01. The optimal classification hyperplane is solved based on the Lagrange multiplier method to determine the anomaly score. When the anomaly score is 0-30, a green warning control command is output; when the score is 30-60, a yellow warning control command is output; and when the score is 60-100, a red warning control command is output. Based on the early warning and control instructions, an abnormality-control mapping relationship is established. When an abnormality of limonene oxidation is detected, a control instruction to reduce the heat treatment temperature by 2-5℃ is output. When an abnormality of citral loss is detected, a control instruction to increase the amount of lemon essential oil added by 5-15% is output. This yields the quality abnormality type and control instruction, which include specific process adjustment parameters.

[0047] Specifically, the CF-LS-SVM diagnostic algorithm constructs a 15-dimensional feature vector from the aforementioned flavor compound feature ratios. This vector contains 15 independent concentration ratio elements, such as the ratio of limonene to citral, the ratio of citral to citric acid, and the ratio of linalool to citronellol, each representing the relative content relationship between specific compounds. In the first stage, a radial basis function (RBF) kernel is used for flavor anomaly detection. The RDF is a kernel function based on Euclidean distance, calculating the similarity between the input vector and the support vector. The kernel parameter 0.125 controls the width of the RDF; a smaller parameter value makes the kernel function sharper and more sensitive to small differences between samples. The anomaly detection process calculates the similarity between the 15-dimensional feature vector and normal samples in the training set. When the similarity is lower than a preset threshold, it is determined to be an anomaly sample. The similarity calculation is based on the kernel function's output value. Normal samples form clustered regions in the feature space, while anomaly samples deviate from these regions. The anomaly detection classification result is output in binary form, with normal samples marked as 0 and anomaly samples marked as 1. This result serves as input data for subsequent type identification.

[0048] Type recognition processes anomalous samples selected from anomaly detection and classification results. Eight typical flavor defects correspond to independent polynomial kernel function classifiers, each classifier specifically identifying a particular type of flavor defect. The polynomial kernel function establishes a nonlinear mapping by calculating the power of the inner product of the input vectors. The formula for the polynomial kernel function is: the inner product of the input vector and support vectors multiplied by the first parameter 0.8, plus the second parameter 0.2, and then raised to the power of 3. The first parameter controls the weight of the inner product, the second parameter provides the offset, and the power parameter 3 determines the complexity of the polynomial. The limonene oxidation defect classifier focuses on the characteristic patterns of decreasing limonene concentration and increasing oxide concentration; the citral loss defect classifier mainly identifies abnormally low absolute citral concentrations; the rancidity defect classifier detects the presence of abnormal aldehydes and ketones; and other defect classifiers correspond to the feature identification of texture abnormalities, insufficient aroma intensity, bitterness, acidity imbalance, and off-flavors, respectively. In the type recognition process, the 15-dimensional feature vector of the anomalous sample is input into each of the eight classifiers. Each classifier outputs the probability value of the sample belonging to the corresponding defect type, and the defect type corresponding to the classifier with the highest probability value is taken as the final recognition result.

[0049] The anomaly severity quantification score is calculated using a linear kernel function, the simplest form of which directly calculates the inner product of the input vector and support vectors, avoiding complex nonlinear transformations. The quantification scoring algorithm is based on the deviation between the feature vectors of the anomaly samples and the standard defect template; a larger deviation indicates a more severe anomaly. The scoring range is set from 0 to 100, with 0 representing complete normality and 100 representing severe anomaly. A regularization parameter of 0.01 is used to prevent overfitting. A penalty term for the weight vector is added to the objective function. The regularization parameter controls the penalty strength; a smaller regularization parameter allows for higher model complexity but increases the risk of overfitting. The scoring process first determines the deviation values ​​of the anomaly feature vectors across each defect dimension, then obtains a comprehensive score through weighted summation. The weight coefficients are determined based on the impact of each defect type on product quality.

[0050] The Lagrange multiplier method for finding the optimal classification hyperplane is a core optimization algorithm for Support Vector Machines (SVMs). This method transforms the constrained optimization problem into a dual problem for solution. The Lagrange function includes the original objective function, equality constraints, and inequality constraints. By introducing Lagrange multipliers, saddle point conditions for the Lagrange function are established. The optimal solution is determined by ensuring that the partial derivatives of the Lagrange function with respect to the original and dual variables are zero. The optimal classification hyperplane divides the feature space into different warning level regions. An anomaly score of 0-30 corresponds to a green warning region, indicating slight anomalies or normal fluctuations; 30-60 corresponds to a yellow warning region, indicating moderate anomalies requiring attention; and 60-100 corresponds to a red warning region, indicating severe anomalies requiring immediate action. Warning and control instructions are automatically generated based on the anomaly score interval. Green warnings output monitoring suggestions, yellow warnings output adjustment suggestions, and red warnings output emergency handling instructions.

[0051] The anomaly-control mapping relationship is established based on the characteristics of the lemon-flavored thick cream production process and the defect generation mechanism. Limonene oxidation anomalies are mainly caused by oxidation reactions under high-temperature conditions. Control measures involve reducing the heat treatment temperature by 2-5℃ to slow down the oxidation reaction rate. The degree of temperature reduction is determined according to the severity of oxidation: 2℃ for mild oxidation, 3-4℃ for moderate oxidation, and 5℃ for severe oxidation. Citral loss anomalies originate from the volatilization or degradation of citral. Control measures involve increasing the amount of lemon essential oil added by 5-15% to compensate for the lost citral content. The amount added is calculated based on the degree of loss: 5% for mild loss, 8-12% for moderate loss, and 15% for severe loss. Control measures for other defect types include pH adjustment, homogenization parameter optimization, and improved storage conditions. Each control measure has a clearly defined parameter range and implementation conditions.

[0052] In one specific embodiment, the process of performing type identification on the detected abnormal samples based on the anomaly detection classification results may specifically include the following steps: Based on the results of anomaly detection and classification, abnormal sample data were screened out, and the 15-dimensional flavor compound characteristic ratio of the abnormal samples were grouped according to eight defect types: limonene oxidation, citral loss, rancidity, off-flavor generation, abnormal texture, insufficient aroma intensity, bitterness generation, and acidity imbalance. A dedicated multinomial kernel function classifier is established for each type of flavor defect. The similarity between the input vector and the support vector is calculated by the multinomial kernel function. The kernel function calculation formula is set as the inner product of the input vector and the support vector, multiplied by the first parameter, plus the second parameter, and then exponentially operated. Based on the calculation results of the multinomial kernel function, feature space mapping is performed on each abnormal sample. The 15-dimensional flavor compound feature ratio is projected to a high-dimensional feature space to obtain the classification decision boundary for each defect type. The feature space mapping results are classified and determined by the support vector machine decision function. The distance value from the abnormal sample to each decision boundary is calculated, and the defect type with the smallest distance value is selected as the identification result. The classification results are matched and verified with the corresponding defect type labels. When the recognition accuracy of limonene oxidation type reaches more than 94% and the recognition accuracy of citral loss type reaches more than 92%, the verified flavor defect type identification data is obtained.

[0053] Specifically, the data processing for flavor defect type identification is based on the anomaly detection and classification results obtained in the first stage of the aforementioned CF-LS-SVM algorithm. The anomaly sample screening process identifies sample data that deviates from normal flavor characteristics using binary identifiers. The 15-dimensional flavor compound feature ratios of these anomaly samples include the relative content information of key flavor components such as the ratio of limonene to citral, the ratio of citral to citric acid, and the ratio of linalool to citronellol. Group processing pre-classifies anomaly samples according to eight typical flavor defect patterns: the limonene oxidation defect group mainly includes samples with decreased limonene concentration and increased limonene oxide concentration; the citral loss defect group collects samples with significantly decreased absolute citral concentration; the rancidity defect group aggregates samples with abnormally high concentrations of aldehydes and ketones; the off-flavor generation defect group includes samples with non-target flavor compounds; the texture abnormality defect group covers samples with imbalanced proportions of compounds affecting the product's physical properties; the aroma insufficiency defect group includes samples with overall low concentrations of major flavor compounds; the bitterness generation defect group collects samples with excessive oxidation of terpenes; and the acidity imbalance defect group includes samples with abnormal proportions of organic acid compounds. Grouping is performed by combining cluster pre-analysis and expert knowledge, with the sample data of each defect group serving as the training set for the corresponding classifier.

[0054] The multinomial kernel function classifier builds a specialized classification model for each flavor defect type. The multinomial kernel function calculates the similarity between the input vector and the support vectors. The kernel function calculation formula obtains a scalar value by performing an inner product operation on the input vector and the support vectors. This scalar value is then linearly scaled by the first parameter 0.8, and then the second parameter 0.2 is added as an offset. Finally, the result is raised to the power of 3 to obtain the final kernel function output value. The input vector is the 15-dimensional flavor compound feature ratio of the abnormal sample to be classified. The support vectors are key sample points selected from the training samples. These support vectors are located near the classification boundaries of different defect types and play a decisive role in the classification decision. The support vectors of the limonene oxidation defect classifier mainly come from samples with abnormal ratios of limonene and its oxidation products. The support vectors of the citral loss defect classifier are selected from samples with significant changes in citral concentration gradients. The support vectors of other defect classifiers correspond to typical feature samples of their respective defect types. The power of 3 operation of the multinomial kernel function enhances the classifier's ability to model complex nonlinear relationships, enabling the classifier to identify higher-order interaction patterns between flavor compounds.

[0055] Feature space mapping projects the 15-dimensional flavor compound feature ratios from the original input space to a high-dimensional feature space. This mapping process is achieved through a nonlinear transformation using a polynomial kernel function. The 15 dimensions in the original space are expanded into a higher-dimensional representation in the feature space, with each dimension corresponding to different combinations and power terms of the original features. The advantage of feature space mapping is that it transforms linearly inseparable data in the original space into linearly separable data in the high-dimensional space, with each defect type forming a relatively independent clustering region. The classification decision boundary is the hyperplane that separates different defect types. The decision boundary is determined through the support vector machine training process, which searches for the separating hyperplane with the maximum margin. The margin refers to the distance from the decision boundary to the nearest sample point; the maximum margin criterion ensures good generalization performance of the classifier. The decision boundary for limonene oxidation defects separates this type of defect from other defect types. The decision boundary for citral loss defects specifically identifies the feature pattern of abnormally low citral concentration. The decision boundaries for other defect types correspond to their respective feature recognition regions.

[0056] The Support Vector Machine (SVM) decision function classifies anomaly samples by calculating their distances to each decision boundary. The distance calculation is based on the anomaly sample's coordinates in the high-dimensional feature space and the equation parameters of the decision boundaries. The sign of the distance value indicates which side of the decision boundary the sample is located on, and the absolute magnitude of the distance value indicates how close the sample is to the decision boundary. The classification process involves substituting the anomaly sample into eight defect classifiers for calculation. Each classifier outputs the distance value from the sample to its corresponding decision boundary. The classifier with the smallest distance value corresponds to the defect type as the final identification result. The smallest distance value means that the sample is closest to the typical feature pattern of the corresponding defect type. The decision function also outputs a classification confidence score, calculated based on the distance value and the distribution density of the support vectors. The smaller the distance value and the higher the support vector density, the greater the confidence score.

[0057] The matching and validation process compares the classification results with manually labeled defect types. Validation data comes from an independent set of test samples, including lemon-flavored thick cream samples with known defect types. The accuracy rate for identifying limonene oxidation is calculated by dividing the number of correctly identified limonene oxidation samples by the total number of limonene oxidation samples; the accuracy rate for identifying citral loss is calculated similarly. Accuracy thresholds are set based on the actual needs of quality monitoring for lemon-flavored thick cream. Limonene oxidation, as the most common defect type, requires an accuracy rate of over 94%, while citral loss, as a key defect affecting flavor intensity, requires an accuracy rate of over 92%. The validation process also includes confusion matrix analysis, which shows the misclassification patterns among different defect types. Analysis of misclassification patterns optimizes classifier parameters and feature selection.

[0058] The above describes the quality monitoring method used in the production process of lemon-flavored thick cream in the embodiments of this application. The following describes the quality monitoring system used in the production process of lemon-flavored thick cream in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the quality monitoring system for the production process of lemon-flavored thick cream in this application includes: The detection module is used to detect the concentration of limonene molecules at the production node in real time using a lemon essential oil volatility detection sensor, and obtain limonene gas phase release data. The analysis module is used to perform kinetic modeling and analysis of the limonene-protein binding state based on the limonene gas phase release data, and obtain the limonene binding degree parameter. The prediction module is used to perform flavor intensity prediction processing on the limonene binding degree parameter using the CLFK-BP neural network algorithm to obtain the predicted value of lemon flavor decay. The calculation module is used to establish a lemon flavor fingerprint spectrum based on the predicted lemon flavor decay value, and to calculate the characteristic ratio of flavor compounds through the concentration ratio matrix. The identification module is used to input the characteristic ratio of the flavor compounds into the CF-LS-SVM diagnostic algorithm for anomaly identification processing to obtain the quality anomaly type and control instructions.

[0059] above Figure 2 The quality monitoring system for lemon-flavored thick milk production process in this embodiment of the invention is described in detail from the perspective of modular functional entities. The quality monitoring equipment for lemon-flavored thick milk production process in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0060] Reference Figure 3 This invention also provides a quality monitoring device for the production process of lemon-flavored thick cream. This quality monitoring device can be a server, and its internal structure can be as follows: Figure 3 As shown. The quality monitoring equipment for the lemon-flavored thick cream production process includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the quality monitoring equipment for the lemon-flavored thick cream production process includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the quality monitoring equipment for the lemon-flavored thick cream production process stores the data corresponding to this embodiment. The network interface of the quality monitoring equipment for the lemon-flavored thick cream production process is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0061] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the quality monitoring equipment used in the production process of lemon-flavored thick milk where the present invention is applied.

[0062] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the quality monitoring method for the production process of lemon-flavored thick milk.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0064] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a quality monitoring device (which may be a personal computer, server, or network device, etc.) used in the production process of lemon-flavored thick cream to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A quality control method for lemon-flavored thick cream production process, characterized in that, The method includes: The concentration of limonene molecules at the production stage is detected in real time using a lemon essential oil volatility detection sensor to obtain limonene gas phase release data. Based on the gas-phase release data of limonene, a kinetic modeling analysis of the limonene-protein binding state was performed to obtain the limonene binding degree parameter. The flavor intensity prediction process for the limonene binding degree parameter was performed using the CLFK-BP neural network algorithm to obtain the predicted value of lemon flavor decay. A lemon flavor fingerprint spectrum is established based on the predicted lemon flavor decay values, and the characteristic ratios of flavor compounds are calculated using a concentration ratio matrix. The characteristic ratios of the flavor compounds are input into the CF-LS-SVM diagnostic algorithm for anomaly identification to obtain the quality anomaly type and control instructions.

2. The quality monitoring method for lemon-flavored thick cream production process according to claim 1, characterized in that, The process involves real-time detection of limonene molecule concentration at the production node using a lemon essential oil volatility detection sensor to obtain limonene gas phase release data, including: Gas chromatography-mass spectrometry (GC-MS) sensors were deployed at six key production nodes: raw material mixing tank, homogenizer inlet, homogenizer outlet, heat processor inlet, heat processor outlet, and cooler outlet, to construct a limonene molecule concentration monitoring network. Qualitative and quantitative analysis of limonene molecules at each node was performed based on the principle of gas chromatography separation. The retention time of limonene was set to 8.32 minutes, the concentration detection range was 0.1-50 mg / L, and the detection accuracy was ±0.05 mg / L. The original signal of limonene gas phase concentration was filtered and amplified, and the limonene release rate of each production node was calculated by the concentration change rate within a 30-second response time. Based on the principle of gas-liquid equilibrium, the liquid phase concentration was calculated by back-calculating the gas phase concentration data of limonene to determine the distribution ratio of free limonene and bound limonene in the liquid phase. The limonene concentration distribution data is matched with the synchronously collected process parameters such as temperature, pressure, flow rate, and pH value using timestamps to obtain the limonene gas phase release data containing concentration-process parameter correlation information.

3. The quality monitoring method for lemon-flavored thick cream production process according to claim 1, characterized in that, The kinetic modeling and analysis of the limonene-protein binding state based on the limonene gas-phase release data yields limonene binding degree parameters, including: Based on the gas-phase release data of limonene, a molecular diffusion dynamics model of limonene was established, and the concentration of free limonene, the concentration of bindable protein sites, and the concentration of limonene-protein complex were used as state variables for modeling. Changes in α-helix content of β-casein were analyzed by circular dichroism spectroscopy. A protein conformational change marker was triggered when the α-helix content decreased by more than 15%. The binding rate constant is dynamically adjusted based on the protein conformational change markers, with the binding rate constant range set to 0.02-0.15, the dissociation rate constant range set to 0.01-0.08, and the volatile rate constant range set to 0.005-0.

025. The diffusion dynamics model of limonene molecules was solved using a nonlinear adaptive control algorithm to obtain the real-time binding rate and dissociation rate of limonene with proteins. The binding rate and dissociation rate values ​​were normalized, and the percentage of limonene molecules' encapsulation efficiency in the milk fat environment was calculated to obtain the limonene binding degree parameter characterizing the binding stability.

4. The quality monitoring method for lemon-flavored thick cream production process according to claim 1, characterized in that, The step of using the CLFK-BP neural network algorithm to predict the flavor intensity of the limonene binding degree parameter to obtain a predicted value for lemon flavor decay includes: The limonene binding degree parameter is combined with the initial concentration of limonene, β-casein concentration, milk fat content, temperature, pH value, homogenization pressure, homogenization time, heat treatment temperature, heat treatment time, cooling rate, product viscosity, and storage time to construct a 13-dimensional input vector. The 13-dimensional input vector was weighted based on the temperature-sensitive limonene volatility coefficient. The pre-exponential factor of the volatility coefficient was set to 2.3 × 10^8, and the activation energy was set to 58.2 kJ / mol to obtain temperature-weighted input data. The temperature-weighted input data was input into a 16-hidden-layer neural network for forward propagation calculation. The activation function of the hidden layer was a modified hyperbolic tangent function, and the limonene-specific correction parameter was set to 0.

15. The network weights are iteratively optimized using the backpropagation algorithm, with a learning rate of 0.001 and 5000 training iterations. Training stops when the prediction error converges to within 2%. The trained neural network is used to calculate and process the three output neurons, which output the lemon flavor intensity decay rate, the limonene residue rate after 24 hours, and the product shelf life days, respectively, to obtain the predicted lemon flavor decay value.

5. The quality monitoring method for lemon-flavored thick cream production process according to claim 1, characterized in that, The step of establishing a lemon flavor fingerprint spectrum based on the predicted lemon flavor decay value and calculating the characteristic ratios of flavor compounds through a concentration ratio matrix includes: Based on the predicted lemon flavor decay values, gas chromatography-mass spectrometry was used to detect lemon-flavored thick cream, and the concentration data of 15 key flavor compounds, including limonene, citral, citric acid, linalool, citronellol, ethyl citrate, methyl citrate, nerol, geraniol, limonene oxide, citral dimethyl acetal, triethyl citrate, α-pinene, β-pinene, and limonene dioxide, were obtained. A concentration ratio matrix was established based on the detected flavor compound concentration data, and the concentration ratio relationship between compounds was calculated. The retention times of limonene and citral were set to 8.32 minutes and 12.45 minutes, respectively, and the retention time of citric acid was set to 15.67 minutes. The concentration ratio matrix was standardized, and the standard ratio range of normal lemon flavor fingerprint was set as follows: limonene to citral ratio 3.2-4.1, citral to citric acid ratio 0.8-1.2, and linalool to citronellol ratio 1.5-2.

3. Based on eight typical flavor defect patterns, namely limonene oxidation, citral loss, rancidity, off-flavor formation, abnormal texture, insufficient aroma intensity, bitterness formation, and acidity imbalance, the standardized ratio data are classified and labeled, and feature recognition thresholds for each defect type are established. The feature recognition threshold is compared and analyzed with the measured ratio data. When the ratio of limonene to citral drops below 2.1, it is marked as limonene oxidation. When the absolute concentration of citral is below 0.5 mg / L, it is marked as citral loss. The feature ratio of the flavor compound containing the defect type identifier is obtained.

6. The quality monitoring method for lemon-flavored thick cream production process according to claim 1, characterized in that, The step of inputting the characteristic ratio of the flavor compounds into the CF-LS-SVM diagnostic algorithm for anomaly identification processing to obtain the quality anomaly type and control instructions includes: The flavor compound feature ratios were constructed into a 15-dimensional feature vector. Flavor anomaly detection was performed using a first-stage radial basis function kernel function with the kernel function parameter set to 0.125 to obtain the anomaly detection classification results. Based on the anomaly detection and classification results, the detected abnormal samples are identified by type. The eight typical flavor defects are input into the corresponding polynomial kernel function classifiers for processing. The first parameter of the polynomial kernel function is set to 0.8, the second parameter is set to 0.2, and the power parameter is set to 3. The degree of anomaly is quantified and scored based on the type identification results. An anomaly score of 0-100 is calculated and processed using a linear kernel function, with the regularization parameter set to 0.

01. The optimal classification hyperplane is solved based on the Lagrange multiplier method for the anomaly score. When the anomaly score is 0-30, a green warning control command is output; when the score is 30-60, a yellow warning control command is output; and when the score is 60-100, a red warning control command is output. An abnormality-control mapping relationship is established based on the aforementioned early warning and control instructions. When an abnormality in limonene oxidation is detected, a control instruction to reduce the heat treatment temperature by 2-5°C is output. When an abnormality in citral loss is detected, a control instruction to increase the amount of lemon essential oil added by 5-15% is output. This yields the quality abnormality type and control instruction, which include specific process adjustment parameters.

7. The quality monitoring method for lemon-flavored thick cream production process according to claim 6, characterized in that, The abnormal samples detected are identified based on the anomaly detection and classification results. Eight typical flavor defects are input into corresponding multinomial kernel function classifiers for processing. The first parameter of the multinomial kernel function is set to 0.8, the second parameter to 0.2, and the power parameter to 3, including: Based on the anomaly detection and classification results, abnormal sample data were selected, and the 15-dimensional flavor compound characteristic ratio of the abnormal samples were grouped according to eight defect types: limonene oxidation, citral loss, rancidity, off-flavor generation, texture abnormality, insufficient aroma intensity, bitterness generation, and acidity imbalance. A dedicated multinomial kernel function classifier is established for each type of flavor defect. The similarity between the input vector and the support vector is calculated by the multinomial kernel function. The kernel function calculation formula is set as the inner product of the input vector and the support vector, multiplied by the first parameter, plus the second parameter, and then exponentially operated. Based on the calculation results of the polynomial kernel function, feature space mapping is performed on each abnormal sample. The 15-dimensional flavor compound feature ratio is projected to a high-dimensional feature space to obtain the classification decision boundary for each defect type. The feature space mapping results are classified and determined by the support vector machine decision function. The distance value from the abnormal sample to each decision boundary is calculated, and the defect type with the smallest distance value is selected as the identification result. The classification results are matched and verified with the corresponding defect type labels. When the recognition accuracy of limonene oxidation type reaches 94% or more and the recognition accuracy of citral loss type reaches 92% or more, the verified flavor defect type identification data is obtained.

8. A quality monitoring system for the production process of lemon-flavored thick milk, characterized in that, For implementing the quality monitoring method for the production process of lemon-flavored thick cream as described in any one of claims 1-7, the quality monitoring system for the production process of lemon-flavored thick cream comprises: The detection module is used to detect the concentration of limonene molecules at the production node in real time using a lemon essential oil volatility detection sensor, and obtain limonene gas phase release data. The analysis module is used to perform kinetic modeling and analysis of the limonene-protein binding state based on the limonene gas phase release data, and obtain the limonene binding degree parameter. The prediction module is used to perform flavor intensity prediction processing on the limonene binding degree parameter using the CLFK-BP neural network algorithm to obtain the predicted value of lemon flavor decay. The calculation module is used to establish a lemon flavor fingerprint spectrum based on the predicted lemon flavor decay value, and to calculate the characteristic ratio of flavor compounds through the concentration ratio matrix. The identification module is used to input the characteristic ratio of the flavor compounds into the CF-LS-SVM diagnostic algorithm for anomaly identification processing to obtain the quality anomaly type and control instructions.

9. A quality monitoring device for use in the production process of lemon-flavored thick milk, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, the processor executing the computer program to implement the quality monitoring method for the production process of lemon-flavored thick milk as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by a processor, the processor performs a quality monitoring method for the production process of lemon-flavored thick cream as described in any one of claims 1 to 7.

Images